JP6340227B2 - Human detection device - Google Patents

Description

  The present invention relates to a person detection apparatus, and more particularly to a person detection apparatus that detects a person from a captured image taken by a camera.

  There is an object detection device that determines whether or not a detection target exists in an image taken by a camera. For example, the object detection device is mounted on a vehicle together with a camera. The object detection device can notify the driver of the presence of a pedestrian by determining whether or not a person is present in an image captured by the camera. By using the object detection device, the driver can easily grasp the situation outside the vehicle.

  The object detection apparatus uses pattern matching to determine whether or not a detection target exists in an image. Examples of pattern matching algorithms include neural networks and support vector machines. The object detection device sets an area (window area) for detecting an object for the input image. The object detection device determines whether or not the detection target is included in the window region using a program in which the above algorithm is implemented.

  Patent Documents 1 and 2 disclose object detection apparatuses using pattern matching.

  The pedestrian recognition device according to Patent Document 1 specifies a region of a captured image similar to a pedestrian image as a pedestrian candidate. The pedestrian recognition device detects a pedestrian from the photographed image by detecting an edge indicating the characteristics of the lower body of the pedestrian in the identified pedestrian candidate. The edges indicating the characteristics of the lower body are an obliquely right edge and an obliquely left edge corresponding to both legs of the pedestrian. In addition, an edge connecting the diagonally right edge and the diagonally left edge is detected from the pedestrian candidate. The pedestrian recognition device determines whether a pedestrian exists in the pedestrian candidate based on the ratio of the lengths of these three edges.

  The object identification device according to Patent Document 2 calculates the similarity of six parts of a person (head, right shoulder, left shoulder, inseam, right foot, and left foot) for each block of a captured image. The target object identification device calculates an integrated similarity obtained by integrating similarities of parts based on a preset positional relationship between the parts. Based on the result of comparing the integrated similarity with a predetermined threshold value, it is determined whether or not a pedestrian exists in the captured image.

JP 2013-97459 A JP 2013-23080 A

  As described above, the pedestrian recognition device according to Patent Document 1 specifies a pedestrian using a pedestrian image. Since the pedestrian image set in advance is an image including the whole body, even if a pedestrian whose upper or lower body is hidden is present in the photographed image, an area including these pedestrians is specified as a pedestrian candidate. I can't.

  The object identification device according to Patent Document 2 integrates similarities of six parts of a person (head, right shoulder, left shoulder, inseam, right foot, and left foot), and the integrated similarity exceeds a threshold value. In this case, it is determined that a person exists. However, if any part of the person is hidden, the integrated similarity decreases. As a result, when a pedestrian whose part is hidden is present in the captured image, the integrated similarity corresponding to this pedestrian may be lower than the threshold value. In this case, there is a possibility that this pedestrian cannot be detected.

  As described above, the conventional person detection device has a problem that the accuracy of detecting a person from a captured image is lowered.

  In view of the above problems, an object of the present invention is to provide a person detection device that can improve the accuracy of detecting a person from a captured image.

  In order to solve the above-mentioned problem, the invention according to claim 1 is a human detection device, wherein a first window region is set for a frame image, and a second window region located below the first window region. A detection window setting unit that sets the first upper body likelihood that the first window region will include the upper body of the person, and the second upper body likelihood that the second window region will include the upper body of the person, An upper body likelihood calculating unit that calculates the first lower body likelihood that the first window region will include the lower body of the person, and the second lower body likelihood that the second window region will include the lower body of the person. The frame image is calculated based on the lower body likelihood calculating unit for calculating the first upper body likelihood, the second upper body likelihood, the first lower body likelihood, and the second lower body likelihood. A determination unit that determines whether or not a person exists.

  The invention according to claim 2 is the person detection device according to claim 1, wherein the determination unit is configured such that an integrated value of the first upper body likelihood and the second lower body likelihood is a first threshold value. If the integrated value of the second upper body likelihood and the first lower body likelihood is equal to or smaller than a second threshold value, it is determined that a person exists in the frame image.

  The invention according to claim 3 is the person detection device according to claim 1, wherein the determination unit determines that the difference absolute value between the first upper body likelihood and the second upper body likelihood is a predetermined upper body reference value. If it is above, or if the difference absolute value between the first lower body likelihood and the second lower body likelihood is greater than or equal to a predetermined lower body reference value, it is determined that a person exists in the frame image.

  The invention according to claim 4 is the person detection device according to any one of claims 1 to 3, wherein the frame image is determined when the determination unit determines that a person is present in the frame image. A hidden detection unit that determines whether or not the upper or lower body of the person is hidden.

  A fifth aspect of the present invention is the person detection device according to any one of the first to fourth aspects, wherein a first feature extraction process is further performed on the first window region to perform a first upper portion. A first feature extraction unit configured to generate feature data and execute the first feature extraction process on the second window region to generate first lower feature data; and a second feature for the first window region A second feature extraction unit that executes extraction processing to generate second upper feature data, and executes the second feature extraction processing on the second window region to generate second lower feature data. The upper body likelihood calculation unit calculates the first upper body likelihood from the first upper feature data, calculates the second upper body likelihood from the first lower feature data, and the lower body likelihood calculation unit The first lower body likelihood is calculated from the second upper feature data, and the second Calculating the second lower body likelihood from part feature data.

  According to a seventh aspect of the present invention, a step of setting a first window area and a second window area positioned below the first window area for a frame image in a computer mounted on the person detection apparatus. Calculating a first upper body likelihood that the first window area will include the upper body of a person and a second upper body likelihood that the second window area will include the upper body of the person; Calculating a first lower body likelihood that the first window region will include the lower body of the person and a second lower body likelihood that the second window region will include the lower body of the person; and the first upper body Determining whether a person is present in the frame image based on likelihood, the second upper body likelihood, the first lower body likelihood, and the second lower body likelihood; It is a person detection program for making it happen.

  The person detection apparatus according to the present invention calculates a first upper body likelihood and a first lower body likelihood from the first window region, and calculates a second upper body likelihood and a second lower body likelihood from the second window region. The person detection apparatus according to the present invention can determine whether or not a person is present in the captured image with high accuracy by using these four likelihoods.

It is a functional block diagram which shows the structure of the person detection apparatus which concerns on the 1st Embodiment of this invention. It is a figure which shows the upper pedestrian image and lower pedestrian image which are learned by the person detection apparatus shown in FIG. It is a flowchart which shows operation | movement of the person detection apparatus shown in FIG. It is a figure which shows the gray scale image produced | generated by the image conversion part shown in FIG. It is a figure which shows the upper window and lower window produced | generated from the detection window shown in FIG. It is a figure which shows the specific example of the upper body likelihood and lower body likelihood calculated by the likelihood calculation part shown in FIG. It is a flowchart of the judgment process performed by the judgment part shown in FIG. It is a flowchart of the hidden detection process shown in FIG. It is a flowchart of the judgment process performed by the person detection apparatus which concerns on the 2nd Embodiment of this invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. In the drawings, the same or corresponding parts are denoted by the same reference numerals and description thereof will not be repeated.

[First Embodiment]
{1. overall structure}
FIG. 1 is a functional block diagram showing the configuration of the person detection apparatus 1 according to the first embodiment of the present invention. The person detection device 1 detects a person (pedestrian) from an image captured by the camera 100. The person detection device 1 is mounted on a vehicle such as an automobile together with the camera 100.

  The camera 100 is installed on the dashboard of the automobile, and generates image data 20 by photographing the scenery in front. The image data 20 is a moving image composed of a plurality of frames, and is input to the person detection device 1 in units of frames. The person detection device 1 determines whether or not there is a pedestrian in one frame (hereinafter referred to as “captured image 21”).

  As shown in FIG. 1, the person detection device 1 includes an image conversion unit 11, a detection window setting unit 12, feature extraction units 13 </ b> A and 13 </ b> B, a likelihood calculation unit 14, a determination unit 15, and a learning unit 16. And a storage unit 17. The image conversion unit 11 converts the color photographed image 21 into a grayscale image 22.

  The detection window setting unit 12 sets a detection window for the grayscale image 22. The detection window is a unit area for detecting a pedestrian from the captured image 21. Further, the detection window setting unit 12 sets the area above the detection window as the upper window 31 and cuts out the upper window 31 from the grayscale image 22. The detection window setting unit 12 sets an area under the detection window as the lower window 41 and cuts out the lower window 41 from the grayscale image 22. Details of the detection window, the upper window 31, and the lower window 41 will be described later.

  The feature extraction unit 13A performs image processing for extracting the feature amount of the upper body of the pedestrian on the upper window 31 and the lower window 41. Specifically, the feature extraction unit 13A extracts the HOG (Histograms Of Oriented Gradients) feature amount of the upper window 31 to generate the upper feature data 32A. Further, by extracting the HOG feature amount of the lower window 41, lower feature data 42A is generated.

  The feature extraction unit 13 </ b> B performs image processing for extracting the feature amount of the lower body of the pedestrian on the upper window 31 and the lower window 41. Specifically, the feature extraction unit 13B generates the upper feature data 32B by executing processing for emphasizing the edge extending in the diagonally right direction and the edge extending in the diagonally leftward direction on the upper window 31. Similarly, the lower feature data 42 </ b> B is generated by executing the processing for enhancing the edge on the lower window 41.

  The likelihood calculating unit 14 includes an upper body likelihood calculating unit 14A and a lower body likelihood calculating unit 14B.

  The upper body likelihood calculating unit 14A calculates the upper body likelihood 33A from which the upper window 31 will include the upper body of the pedestrian from the upper feature data 32A. The upper body likelihood calculation unit 14A calculates the upper body likelihood 43A from which the lower window 41 will include the upper body of the pedestrian from the lower feature data 42A. For the calculation of the upper body likelihoods 33A and 43A, pattern data 50A in which a pattern of the upper body of a pedestrian is recorded is used.

  The lower body likelihood calculating unit 14B calculates a lower body likelihood 33B from which the upper window 31 will include the lower body of the pedestrian from the upper feature data 32B. The lower body likelihood calculating unit 14B calculates a lower body likelihood 43B from which the lower window 41 will include the lower body of the pedestrian from the lower feature data 42B. For calculation of the lower body likelihoods 33B and 43B, pattern data 50B in which a pattern of the lower body of the pedestrian is recorded is used.

  Based on the upper body likelihoods 33A and 43A calculated by the upper body likelihood calculation unit 14A and the lower body likelihoods 33B and 43B calculated by the lower body likelihood calculation unit 14B, the determination unit 15 determines whether the pedestrian is in the detection window. Determine if it exists. Details of the operation of the determination unit 15 will be described later.

  Before starting the process of detecting the pedestrian from the captured image 21, the learning unit 16 learns the pattern of the upper body of the pedestrian to generate the pattern data 50A, and learns the pattern of the lower body of the pedestrian to obtain the pattern data. 50A is generated. Details of the learning unit 16 will be described later.

  The storage unit 17 is, for example, a hard disk device or a flash memory. The storage unit 17 stores pattern data 50A and 50B.

{2. Operation of Person Detection Device 1}
{2.1. Pedestrian learning}
The person detecting device 1 learns an image including a pedestrian and creates pattern data 50A and 50B before starting detection of the pedestrian from the captured image 21.

  FIG. 2 is a diagram illustrating an example of an image learned by the person detection device 1. The pedestrian image 51 shown in FIG. 2 is grayscale and has the same width and height as the detection window. The pedestrian image 51 is directly input to the detection window setting unit 12 without going through the image conversion unit 11. The detection window setting unit 12 generates an upper pedestrian image 51A and a lower pedestrian image 51B from the pedestrian image 51. When the detection window setting unit 12 generates the upper pedestrian image 51 </ b> A and the lower pedestrian image 51 </ b> B, the detection window setting unit 12 does not perform a process of setting a detection window on the pedestrian image 51.

  The upper pedestrian image 51A is cut out from the pedestrian image 51 so as to include the upper half of the pedestrian 60 in the pedestrian image 51. When the height of the pedestrian image 51 is H, the upper pedestrian image 51A is an image above the horizontal straight line L1 at a height of 0.45 × H with reference to the lower end of the pedestrian image 51. It is.

  The lower pedestrian image 51B is cut out from the pedestrian image 51 so as to include the lower half of the pedestrian 60. The lower pedestrian image 51B is an image below the horizontal straight line L2 at a height of 0.55 × H with the lower end of the pedestrian image 51 as a reference.

  In FIG. 2, a part of the upper pedestrian image 51A overlaps with a part of the lower pedestrian image 51B, but the upper pedestrian image 51A and the lower pedestrian image 51B do not overlap. May be. In FIG. 2, the height of the upper pedestrian image 51A is the same as the height of the lower pedestrian image 51B, but the height of the upper pedestrian image 51A is different from the height of the lower pedestrian image 51B. May be.

  The feature extraction unit 13A executes a process of extracting the HOG feature amount from the upper pedestrian image 51A to generate feature data 62A for learning. The learning unit 16 learns the pattern of the upper body of the pedestrian using the learning feature data 62A generated from the upper pedestrian image 51A. The learning unit 16 generates pattern data 50A reflecting the learning result of the upper body pattern.

  The feature extraction unit 13B executes processing for emphasizing edges on the lower pedestrian image 51B to generate feature data 62B for learning. The learning unit 16 learns the pattern of the lower body of the pedestrian using the learning feature data 62B generated from the lower pedestrian image 51B. The learning unit 16 generates pattern data 50B reflecting the learning result of the lower body pattern.

  The person detection device 1 creates pattern data 50A and 50B using a plurality of pedestrian images 51. Thereby, various patterns of the upper body and lower body of the pedestrian can be reflected in the pattern data 50A and 50B. In a plurality of pedestrian images 51 used for learning, it is desirable that the pedestrians included in each image are normalized. Here, that the pedestrian is normalized means that the size of the pedestrian, the position of the head, the position of the foot, etc. included in each of the pedestrian images 51 are aligned.

{2.2. Detection of pedestrians}
FIG. 3 is a flowchart showing the operation of the person detection apparatus 1 that detects a pedestrian from the captured image 21. The person detection device 1 executes the process shown in FIG. 3 each time a captured image 21 is input from the camera 100 and detects a pedestrian from the input captured image 21.

  When the person detection device 1 receives the captured image 21 from the camera 100, the image conversion unit 11 converts the captured image 21 into a grayscale image 22 (step S1). The person detection device 1 executes steps S2 to S6, sets a detection window for the grayscale image 22, and determines whether a pedestrian exists in the detection window. Details of steps S2 to S6 will be described later.

  The person detection device 1 checks whether or not the setting of the detection window for the grayscale image 22 is completed after the determination process (step S5) is completed (step S6). If the setting of the detection window has not ended (No in step S6), the person detection device 1 returns to step S2 in order to set a new detection window. On the other hand, when all the detection windows have been set (Yes in step S6), the person detection device 1 ends the process shown in FIG. The person detection device 1 executes the process shown in FIG. 3 again when a new captured image 21 is input.

  Hereinafter, steps S2 to S5 will be described in detail.

(Setting of detection window (step S2))
The detection window setting unit 12 inputs the grayscale image 22 generated by the image conversion unit 11. The detection window setting unit 12 sets the detection window, the upper window 31, and the lower window 41 for the input grayscale image 22 (step S2).

  FIG. 4 is a diagram illustrating an example of the grayscale image 22. The detection window setting unit 12 sets a detection window 23 for detecting a pedestrian. That is, each time step S2 is executed, one detection window 23 is set.

  As shown in FIG. 3, by repeating steps S <b> 2 to S <b> 6, a plurality of detection windows 23 are set for the grayscale image 22. The detection window 23 located at the upper left vertex of the gray scale image 22 (see FIG. 4) is set first. The detection window setting unit 12 sets a plurality of detection windows 23 so as to scan the grayscale image 22 from the upper left to the lower right. When the detection window 23 is newly set, the detection window setting unit 12 desirably sets a new detection window 23 so as to overlap a part of the detection window 23 that has already been set.

  The detection window setting unit 12 generates an upper window 31 and a lower window 41 from the set detection window 23. FIG. 5 is a diagram showing a procedure for generating the upper window 31 and the lower window 41 from the detection window 23 set for the grayscale image 22 shown in FIG. The procedure for generating the upper window 31 and the lower window 41 is the same as the procedure for generating the upper pedestrian image 51A and the lower pedestrian image 51B from the pedestrian image 51 (see FIG. 2).

  That is, as shown in FIG. 5, when the height of the detection window 23 is H, the detection window setting unit 12 has a horizontal direction at a height of 0.45 × H with respect to the lower end of the detection window 23. The region above the straight line L3 is set in the upper window 31. The detection window setting unit 12 generates the upper window 31 by cutting out the upper region from the grayscale image 22. The detection window setting unit 12 sets an area below the horizontal straight line L4 at a height of 0.55 × H in the lower window 41 with the lower end of the detection window 23 as a reference. The detection window setting unit 12 generates the lower window 41 by cutting out the lower region from the grayscale image 22.

  In this way, the detection window setting unit 12 sets the upper window 31 and the lower window 41 located below the upper window 31 for the grayscale image 22 (step S2). Note that, as in the upper pedestrian image 51A and the lower pedestrian image 51B, the upper window 31 and the lower window 41 do not have to partially overlap each other. The height of the upper window 31 may be different from that of the lower window 41.

(Feature extraction (step S3))
In step S <b> 3, the feature extraction unit 13 </ b> A extracts the feature amount of the upper body of the pedestrian from each of the upper window 31 and the lower window 41. The feature extraction unit 13B extracts the feature amount of the lower body of the pedestrian from each of the upper window 31 and the lower window 41. The feature extraction unit 13A extracts the HOG feature value of the upper window 31 to generate the upper feature data 32A, and extracts the HOG feature value of the lower window 41 to generate the lower feature data 42A.

  The feature extraction unit 13B performs processing for emphasizing the edge extending in the right diagonal direction and the edge extending in the left diagonal direction on the upper window 31 to generate the upper feature data 32B. The feature extraction unit 13B executes processing for emphasizing an edge extending in the right oblique direction and an edge extending in the left oblique direction, and generates the lower feature data 42B.

(Likelihood calculation (step S4))
In step S4, the upper body likelihood calculating unit 14A calculates the upper body likelihood 33A from the upper feature data 32A and the upper body likelihood 43A from the lower feature data 42A using the pattern data 50A. The upper body likelihood 33A is a numerical value indicating the degree that the upper window 31 includes the upper body of the pedestrian. The upper body likelihood 43A is a numerical value indicating the degree that the lower window 41 includes the upper body of the pedestrian.

  In step S4, the lower body likelihood calculating unit 14B calculates the lower body likelihood 33B from the upper feature data 32B and the lower body likelihood 43B from the lower feature data 42B using the pattern data 50B. The lower body likelihood 33B is a numerical value indicating the degree to which the upper window 31 includes the lower body of the pedestrian. The lower body likelihood 43B is a numerical value indicating the degree that the lower window 41 includes the lower body of the pedestrian.

  FIG. 6 is a diagram illustrating specific examples of upper body likelihoods 33A and 43A and lower body likelihoods 33B and 43B5. Detection windows 123, 223, and 323 shown in FIG. 6 are images set for a grayscale image 22 different from the grayscale image 22 shown in FIG.

  Upper body likelihoods 33A and 43A and lower body likelihoods 33B and 43B are numerical values of 0 or more and 1 or less, respectively. When the upper window 31 and the lower window 41 include the upper body of the pedestrian, the upper body likelihoods 33A and 43A approach one. When the upper window 31 and the lower window 41 do not include the upper body of the pedestrian, the upper body likelihoods 33A and 43A approach zero. Similarly, when the upper window 31 and the lower window 41 include the lower body of the pedestrian, the lower body likelihoods 33B and 43B approach 1 and when the upper window 31 and the lower window 41 do not include the lower body of the pedestrian, the lower body likelihood. The degrees 33B and 43B approach zero.

  The detection window 123 includes the whole body of a pedestrian that proceeds in the left direction. The upper window 123 a and the lower window 123 b are generated from the detection window 123.

  The upper window 123a includes the upper body of a pedestrian that proceeds in the left direction. The upper body likelihood 33A calculated from the upper window 123a is 0.94, and has a value close to 1 (for example, 0.7 or more). On the other hand, the lower window 123b does not include the upper body of the pedestrian. The upper body likelihood 43A calculated from the lower window 123b is 0.43 and has a value close to 0 (for example, 0.5 or less).

  The lower window 123b includes the lower body of a pedestrian that proceeds in the left direction. The lower body likelihood 43B calculated from the lower window 123b is 0.89 and has a value close to 1. On the other hand, the upper window 123a does not include the lower body of the pedestrian. The lower body likelihood 33B calculated from the upper window 123a is 0.32, and has a value close to 0.

  When the detection window 23 includes the whole body of the pedestrian as in the example of the detection window 123, the upper window 31 includes the upper body of the pedestrian and the lower window 41 includes the lower body of the pedestrian. In this case, the upper body likelihood 33A and the lower body likelihood 43B have values close to 1, whereas the upper body likelihood 43A and the lower body likelihood 33B have values close to 0.

  The detection window 223 includes a pedestrian facing backward, and the lower half of the pedestrian's whole body is hidden. The upper window 223a and the lower window 223b are generated from the detection window 223.

  The upper window 223a includes the upper body of a pedestrian facing backward. The upper body likelihood 33A calculated from the upper window 223a is 0.91, and has a value close to 1. On the other hand, the lower window 223b does not include the upper body of the pedestrian. The upper body likelihood 43A calculated from the lower window 223b is 0.26 and has a value close to 0.

  In the detection window 223, the lower body of the pedestrian facing backward is hidden. For this reason, the lower window 223b does not include the lower body of the pedestrian. The lower body likelihood 43B calculated from the lower window 223b is 0.28 and has a value close to zero. Similarly to the lower window 223b, the upper window 223a does not include the lower body of the pedestrian. The lower body likelihood 33B calculated from the upper window 223a is 0.39 and has a value close to 0.

  Like the detection window 223, the detection window 23 includes a pedestrian, but when the lower body of the pedestrian is hidden, the upper window 31 includes the upper body of the pedestrian and the lower window 41 does not include the lower body of the pedestrian. . In this case, the upper body likelihood 33A has a value close to 1, whereas the upper body likelihood 43A and the lower body likelihoods 33B and 43B have values close to 0.

  Contrary to the detection window 223, the detection window 23 includes a pedestrian, but when the upper body of the pedestrian is hidden, the upper window 31 does not include the upper body of the pedestrian and the lower window 41 covers the lower body of the pedestrian. Including. In this case, the lower body likelihood 43B has a value close to 1, whereas the upper body likelihoods 33A and 43A and the lower body likelihood 33B have values close to 0.

  The detection window 323 does not include a pedestrian and includes a telephone pole in the right region. For this reason, the upper window 323a and the lower window 323b do not include the upper body and the lower body of the pedestrian. The upper body likelihood 33A and the lower body likelihood 33B calculated from the upper window 323a are 0.25 and 0.42, respectively, and have values close to 0. Similarly, the upper body likelihood 43A and the lower body likelihood 43B calculated from the lower window 323b are 0.42 and 0.34, respectively, and have values close to 0. When the detection window 23 does not include a pedestrian as in the detection window 323, the upper body likelihoods 33A and 43A and the lower body likelihoods 33B and 43B all have values close to zero.

  Thus, four kinds of likelihood changes with whether the upper window 31 and the lower window 41 include a pedestrian. In the determination process (step S5) described later, it is determined whether or not there is a pedestrian in the detection window 23 in consideration of the relationship between these four types of likelihoods.

(Judgment process (step S5))
Reference is again made to FIG. After the four types of likelihood are calculated by the likelihood calculating unit 14 (step S4), the determining unit 15 determines whether or not a pedestrian exists in the detection window 23 based on these four types of likelihood. (Step S5). Hereinafter, the determination process (step S5) will be described in detail.

  FIG. 7 is a flowchart of the determination process (step S5) executed by the determination unit 15. In FIG. 7, Lu1 indicates the value of the upper body likelihood 33A, and Lu2 indicates the value of the upper body likelihood 43A. Ld1 indicates the value of the lower body likelihood 33B, and Ld2 indicates the value of the lower body likelihood 43B.

  The determination unit 15 calculates two types of integrated values S1 and S2 (step S51). The integrated value S1 is calculated by integrating the upper body likelihood 33A calculated from the upper window 31 and the lower body likelihood 43B calculated from the lower window 41. The integrated value S2 is calculated by integrating the upper body likelihood 43A calculated from the lower window 41 and the lower body likelihood 33B calculated from the upper window 31.

  The determination unit 15 compares the integrated value S1 with a preset threshold value Th1 (step S52).

  The criteria for determining the threshold value Th1 will be described with reference to the four types of likelihood calculated from the detection windows 123 and 223 in FIG. The detection window 123 includes the whole body of the pedestrian. Therefore, the upper body likelihood 33A calculated from the upper window 123a and the lower body likelihood 43B calculated from the lower window 123b both have values close to 1. The detection window 223 includes a pedestrian whose lower body is hidden. Therefore, the upper body likelihood 33A calculated from the upper window 223a has a value close to 1, and the lower body likelihood 43B calculated from the lower window 223b has a value close to 0. Therefore, in the case where the detection window 23 includes the whole body of the pedestrian, the pedestrian whose lower body is hidden, or the pedestrian whose upper body is hidden, at least the upper body likelihood 33A and the lower body likelihood 43B are included. One has a value close to 1. In these cases, the integrated value S1 is considered to be larger than 1. The threshold value Th1 is determined in consideration of the tendency of the upper body likelihood 33A and the lower body likelihood 43B when the detection window 23 includes a pedestrian's whole body, a pedestrian whose lower body is hidden, or a pedestrian whose upper body is hidden. The The threshold value Th1 is set to 1.0, for example. When integrated value S1 is smaller than threshold value Th1 (No in step S52), determination unit 15 determines that there is no pedestrian in detection window 23 (step S56), and ends the process shown in FIG.

  On the other hand, when integrated value S1 is equal to or greater than threshold value Th1 (Yes in step S52), determination unit 15 compares integrated value S2 with preset threshold value Th2 (step S53).

  The criteria for determining the threshold value Th2 will be described with reference to the four types of likelihood calculated from the detection window 223 of FIG. The detection window 223 includes a pedestrian whose upper body is hidden. For this reason, the upper body likelihood 33A calculated from the upper window 223a has a value close to 1, and the lower body likelihood 33B calculated from the upper window 223a has a value close to 0. The lower window 123b does not include the lower body of the pedestrian because the lower body of the pedestrian is hidden. For this reason, the upper body likelihood 43A and the lower body likelihood 43B calculated from the lower window 123b have values close to zero. That is, when the detection window 23 includes a pedestrian whose lower body is hidden, both the upper body likelihood 43A and the lower body likelihood 33B have values close to zero. Similarly, when the detection window 23 includes the whole body of a pedestrian and a pedestrian whose upper body is hidden, the upper body likelihood 43A and the lower body likelihood 33B both have values close to zero. Thus, the threshold Th2 is calculated when the detection window 23 includes the pedestrian's whole body, the pedestrian whose upper body is hidden, the upper body likelihood 43A calculated when the lower body includes a pedestrian whose lower body is hidden, and It is determined in consideration of the tendency of the lower body likelihood 33B. For example, the threshold value Th2 is set to 0.9.

  When integrated value S2 is larger than threshold value Th2 (No in step S53), determination unit 15 determines that there is no pedestrian in detection window 23 (step S56), and ends the process shown in FIG.

  On the other hand, when integrated value S2 is equal to or smaller than threshold value Th2 (Yes in step S53), determination unit 15 determines that there is a pedestrian in detection window 23 (step S54). The determination part 15 performs step S55, and determines whether the upper body or lower body of the pedestrian who exists in the detection window 23 is hidden. Details of step S55 will be described later.

  Hereinafter, as a specific example, the determination process (step S5) for each detection window shown in FIG. 6 will be described. The threshold value Th1 is set to 1.0, and the threshold value Th2 is set to 0.9.

  First, the detection window 123 will be described. The upper body likelihood 33A and the lower body likelihood 33B calculated from the upper window 123a are 0.94 and 0.32. The upper body likelihood 43A and the lower body likelihood 43B calculated from the lower window 123b are 0.43 and 0.89. As a result of step S51, the integrated value S1 is 1.83. The integrated value S2 is 0.75. Integrated value S1 is equal to or greater than threshold value Th1 (Yes in step S52). Integrated value S2 is equal to or smaller than threshold value Th2 (Yes in step S53). Therefore, the determination unit 15 determines that there is a pedestrian in the detection window 123 (step S54).

  In the detection window 223, the upper body likelihood 33A and the lower body likelihood 33B calculated from the upper window 223a are 0.91 and 0.39. The upper body likelihood 43A and the lower body likelihood 43B calculated from the lower window 223b are 0.26 and 0.28. As a result of step S51, the integrated value S1 is 1.19. The integrated value S2 is 0.65. Integrated value S1 is equal to or greater than threshold value Th1 (Yes in step S52). Integrated value S2 is equal to or smaller than threshold value Th2 (Yes in step S53). Therefore, the determination unit 15 determines that there is a pedestrian in the detection window 223 (step S54).

  In the detection window 323, the upper body likelihood 33A and the lower body likelihood 33B calculated from the upper window 323a are 0.25 and 0.42. The upper body likelihood 43A and the lower body likelihood 43B calculated from the lower window 323b are 0.26 and 0.0.34. As a result of step S51, the integrated value S1 is 0.59. The integrated value S2 is 0.68. Since integrated value S1 is smaller than threshold value Th1 (No in step S52), determination unit 15 determines that there is no pedestrian in detection window 23 (step S56).

  Thus, the determination unit 15 integrates the integrated value S1 obtained by integrating the upper body likelihood 33A calculated from the upper window 31 and the lower body likelihood 43B calculated from the lower window 41, and the lower body likelihood calculated from the upper window 31. Whether or not there is a pedestrian in the detection window 23 is determined using the integrated value S2 obtained by integrating the degree 33B and the upper body likelihood 43A calculated from the lower window 41. Thereby, it can be judged with high accuracy whether a pedestrian exists in the detection window 23 or not. Hereinafter, the reason will be described.

  Although the detection window 23 does not include a pedestrian, the upper body likelihood 33A calculated from the upper window 31 and the lower body likelihood 43B calculated from the lower window 41 may have values close to 1. In this case, there is a high possibility that the integrated value S1 will be equal to or greater than the threshold value Th1 (Yes in step S52). When it is determined whether or not there is a pedestrian in the detection window 23 using only step S52, there is a possibility that a non-pedestrian is erroneously detected as a pedestrian.

  However, when the detection window 23 includes a pedestrian's whole body, a pedestrian whose upper body is hidden, or a pedestrian whose lower body is hidden, the upper body likelihood 33A and the lower body likelihood 33B calculated from the upper window 31 are both close to 1. The upper body likelihood 43A and the lower body likelihood 43B calculated from the lower window 41 are not both close to 1. Therefore, when integrated value S1 is equal to or greater than threshold value Th1 (Yes in step S52), determination unit 15 compares the integrated value S2 with threshold value Th2, thereby erroneously determining a non-pedestrian as a pedestrian. Detection can be prevented.

(Hidden detection process (step S55))
FIG. 8 is a flowchart of the hidden detection process (step S55). The process shown in FIG. 8 is executed when the determination unit 15 determines that there is a pedestrian in the detection window 23 (step S54).

  In the hidden detection process (step S55), among the four types of likelihood, the upper body likelihood 33A calculated from the upper window 31 and the lower body likelihood 43B calculated from the lower window 41 are used.

  Determination unit 15 compares upper body likelihood 33A with threshold value Th3 (step S551). The threshold value Th3 is a reference value for determining whether or not the upper body of the pedestrian exists in the upper window 31, and is, for example, 0.8.

  When upper body likelihood 33A is equal to or less than threshold value Th3 (Yes in step S551), determination unit 15 determines that the upper body of the pedestrian existing in detection window 23 is hidden (step S552). On the other hand, when upper body likelihood 33A is larger than threshold value Th3 (No in step S551), determination unit 15 determines that the upper body of the pedestrian existing in detection window 23 is not hidden (step S555). That is, the determination unit 15 determines that the upper body of the pedestrian can be visually recognized in the detection window 23.

  Next, the determination unit 15 compares the lower body likelihood 43B with the threshold value Th4 (step S553). The threshold value Th4 is a reference value for determining whether or not the lower body of the pedestrian is present in the lower window 41, and is, for example, 0.8. The threshold value Th4 may be different from the threshold value Th3.

  If lower body likelihood 43B is equal to or less than threshold value Th4 (Yes in step S553), determination unit 15 determines that the lower body of the pedestrian existing in detection window 23 is hidden (step S554). On the other hand, when lower body likelihood 43B is larger than threshold Th4 (No in step S553), determination unit 15 determines that the lower body of the pedestrian existing in detection window 23 is not hidden (step S556). That is, the determination unit 15 determines that the lower body of the pedestrian can be visually recognized in the detection window 23.

  Hereinafter, a specific example of the hidden detection process (step S55) will be described using the detection windows 123 and 223. The determination unit 15 determines that a pedestrian is present in the detection window 123 (step S54, see FIG. 7). In this case, the processing shown in FIG.

  As shown in FIG. 6, the upper body likelihood 33A calculated from the upper window 123a is 0.94, and the lower body likelihood 43B calculated from the lower window 123b is 0.89. Since the upper body likelihood 33A is larger than the threshold value Th3 (No in step S551), the determination unit 15 determines that the upper body is not hidden in the detection window 123 (step S555). The determination unit 15 determines that there is no hiding of the lower body in the detection window 123 (step S556) because the lower body likelihood 43B is greater than the threshold value Th4 (No in step S553). The determination unit 15 detects the pedestrian from the detection window 123 and generates result data 25 indicating that the detected upper and lower pedestrians are not hidden.

  Next, the detection window 223 will be described. As described above, the determination unit 15 determines that there is a pedestrian in the detection window 223 (step S54, see FIG. 7). In this case, the process shown in FIG. 8 is executed for the detection window 223.

  As shown in FIG. 6, the upper body likelihood 33A calculated from the upper window 223a is 0.91, and the lower body likelihood 43B calculated from the lower window 223b is 0.28. The determination unit 15 determines that the upper body is not hidden in the detection window 223 because the upper body likelihood 33A is larger than the threshold Th3 (No in step S551) (step S555). The determination unit 15 determines that the lower body is hidden in the detection window 223 because the lower body likelihood 43B is equal to or less than the threshold value Th4 (Yes in Step S553) (Step S554). The determination unit 15 generates result data 25 indicating that a pedestrian is detected from the detection window 223 and that the lower body of the detected pedestrian is hidden.

  As described above, the person detection apparatus 1 calculates the upper body likelihood 33A and the lower body likelihood 33B from the upper window 31, and calculates the upper body likelihood 43A and the lower body likelihood 43B from the lower window 41. The person detection device 1 determines whether there is a pedestrian in the detection window 23 using these four types of likelihoods. Thereby, a pedestrian can be detected from the captured image 21 with high accuracy.

[Second Embodiment]
The object detection device according to the second embodiment of the present invention has the same configuration as the person detection device 1 according to the first embodiment (see FIG. 1). The object detection apparatus according to the present embodiment detects the pedestrian from the captured image 21 by executing the process shown in FIG.

  In the present embodiment, the operation of the determination unit 15 (determination process (step S5)) is different from the operation of the determination unit 15 in the first embodiment. Hereinafter, the operation of the determination unit 15 in the present embodiment will be mainly described.

  FIG. 9 is a flowchart of the determination process (step S5) executed by the object detection apparatus according to the present embodiment. Hereinafter, the determination process executed by the object detection device according to the present embodiment will be described with reference to FIGS. 6 and 9.

  In FIG. 9, Lu1 shows the value which upper body likelihood 33A has, and Lu2 shows the value which upper body likelihood 43A has. Ld1 indicates the value of the lower body likelihood 33B, and Ld2 indicates the value of the lower body likelihood 43B.

  The determination unit 15 calculates two types of absolute difference values D1 and D2 (step S71). The difference absolute value D1 is calculated from the upper body likelihood 33A and the lower body likelihood 33B calculated from the upper window 31. The difference absolute value D1 is calculated by subtracting the lower body likelihood 33B from the upper body likelihood 33A and taking the absolute value of the difference value obtained by the subtraction. The difference absolute value D2 is calculated from the upper body likelihood 43A and the lower body likelihood 43B calculated from the lower window 41. The difference absolute value D2 is calculated by subtracting the upper body likelihood 43A from the lower body likelihood 43B and taking the absolute value of the difference value obtained by the subtraction.

  The determination unit 15 compares the absolute difference value D1 with a preset threshold value Th5 (step S72).

  When the detection window 23 includes the whole body of a pedestrian, the upper body likelihood 33A calculated from the upper window 31 has a value close to 1, and the lower body likelihood 33B calculated from the upper window 31 is close to 0. Has a value. When the detection window 23 includes a pedestrian whose lower body is hidden, the upper body likelihood 33A calculated from the upper window 31 has a value close to 1, and the lower body likelihood 33B calculated from the upper window 31 is 0. Has a close value. Thus, when the detection window 23 includes a pedestrian whose pedestrian's whole body or lower body is hidden, the difference absolute value D1 is considered to have a value equal to or greater than a certain value. The threshold value Th5 is determined in consideration of the tendency of the upper body likelihood 33A and the lower body likelihood 33B when the detection window 23 includes a pedestrian whose entire body or lower body of the pedestrian is hidden. For example, the threshold value Th5 is set to 0.4.

  When the difference absolute value D1 is equal to or greater than the threshold Th5 (Yes in Step S72), the determination unit 15 determines that a pedestrian exists in the detection window 23 (Step S74). And the judgment part 15 performs a hidden detection process (step S55).

  On the other hand, when difference absolute value D1 is smaller than threshold value Th5 (No in step S72), determination unit 15 compares difference absolute value D2 with threshold value Th6 (step S73).

  When the detection window 23 includes the whole body of a pedestrian, the upper body likelihood 43A calculated from the lower window 41 has a value close to 0, and the lower body likelihood 43B calculated from the lower window 41 has a value close to 1. Have. When the detection window 23 includes a pedestrian whose upper body is hidden, the upper body likelihood 43A calculated from the lower window 41 has a value close to 0, and the lower body likelihood 43B calculated from the lower window 41 is close to 1. Has a value. Thus, when the detection window 23 includes a pedestrian whose pedestrian's whole body or upper body is hidden, the difference absolute value D2 is considered to have a value equal to or greater than a certain value, like the difference absolute value D1. The threshold value Th6 is determined in consideration of the tendency of the upper body likelihood 33A and the lower body likelihood 33B when the detection window 23 includes a pedestrian whose entire body or pedestrian is hidden. For example, the threshold value Th6 is set to 0.4. The threshold value Th6 may be different from the threshold value Th5.

  When difference absolute value D2 is smaller than threshold value Th6 (No in step S73), determination unit 15 determines that there is no pedestrian in detection window 23 (step S75). On the other hand, when the difference absolute value D2 is greater than or equal to the threshold Th6 (Yes in step S73), the determination unit 15 determines that a pedestrian is present in the detection window 23 (step S74).

  Hereinafter, as a specific example, the determination process (step S5) for each detection window shown in FIG. 6 will be described. Here, the threshold values Th5 and Th6 are set to 0.4.

  First, the detection window 123 will be described. The upper body likelihood 33A and the lower body likelihood 33B calculated from the upper window 123a are 0.94 and 0.32. The upper body likelihood 43A and the lower body likelihood 43B calculated from the lower window 123b are 0.43 and 0.89. As a result of step S71, the absolute difference values D1 and D2 are 0.62 and 0.46, respectively. The determination unit 15 has the absolute difference value D1 equal to or greater than the threshold value Th5 (Yes in step S72). It is determined that there is a pedestrian in the detection window 123 (step S74).

  Next, the detection window 223 will be described. The upper body likelihood 33A and the lower body likelihood 33B calculated from the upper window 223a are 0.91 and 0.39. The upper body likelihood 43A and the lower body likelihood 43B calculated from the lower window 223b are 0.26 and 0.28. As a result of step S71, the absolute difference values D1 and D2 are 0.52 and 0.02, respectively. The determination unit 15 determines that there is a pedestrian in the detection window 223 because the absolute difference value D1 is equal to or greater than the threshold Th5 (Yes in Step S72) (Step S74).

  Next, the detection window 323 will be described. The upper body likelihood 33A and the lower body likelihood 33B calculated from the upper window 323a are 0.25 and 0.42. The upper body likelihood 43A and the lower body likelihood 43B calculated from the lower window 123b are 0.26 and 0.0.34. As a result of step S71, the absolute difference values D1 and D2 are 0.17 and 0.08, respectively. Difference absolute value D1 is smaller than threshold value Th5 (No in step S72), and difference absolute value D2 is smaller than threshold value Th6 (No in step S73). Accordingly, the determination unit 15 determines that there is no pedestrian in the detection window 23 (step S75).

  Thus, the determination unit 15 calculates the absolute difference D1 between the upper body likelihood 33A and the lower body likelihood 33B calculated from the upper window 31, and the upper body likelihood 43A and the lower body likelihood 43B calculated from the lower window 41. Whether or not there is a pedestrian in the detection window 23 is determined based on the difference absolute value D2. Thereby, similarly to the said 1st Embodiment, it can be judged with high precision whether a pedestrian exists in the detection window 23 or not.

  In the above-described embodiment, the detection window setting unit 12 uses the detection window 23 for setting the upper window 31 and the lower window 41. However, the present invention is not limited to this. For example, the detection window setting unit 12 sets the lower window 41 by first setting the upper window 31 and setting a coordinate having a predetermined positional relationship with the center of the upper window 31 at the center of the lower window 41. Also good. In this case, a coordinate located below the upper window 31 and at a predetermined distance from the upper window 31 may be set at the center of the lower window 41.

  Alternatively, the detection window setting unit 12 may set the lower window 41 first. In this case, the detection window setting unit 12 may be set at the center of the upper window 31 at a predetermined distance from the center of the lower window 41 and positioned above the center of the lower window 41.

  In the above embodiment, the example in which the feature extraction unit 13A extracts the HOG feature amount from each of the upper window 31, the lower window 41, and the upper pedestrian image 51A has been described, but the present invention is not limited thereto. The feature extraction unit 13A may perform image processing that can extract the features of the upper body of the pedestrian from the upper pedestrian image 51A. For example, the feature extraction unit 13A may execute processing for enhancing an edge. That is, the feature extraction unit 13A may perform the same image processing as that performed on the upper pedestrian image 51A on the upper window 31 and the lower window 41.

  Similarly, although the feature extraction part 13B demonstrated the example which performs the process which emphasizes an edge with respect to each of the upper window 31, the lower window 41, and the lower pedestrian image 51B, it is not restricted to this. The feature extraction unit 13B may perform image processing that can extract the features of the lower body of the pedestrian from the lower pedestrian image 51B. For example, the feature extraction unit 13B may execute a process of extracting the HOG feature amount. That is, the feature extraction unit 13B may perform the same image processing as that performed on the lower pedestrian image 51B on the upper window 31 and the lower window 41.

  In the above embodiment, an example in which the upper body likelihood calculation unit 14A and the lower body likelihood calculation unit 14B use a neural network has been described, but the present invention is not limited thereto. The upper body likelihood calculation unit 14A and the lower body likelihood calculation unit 14B may use an algorithm other than the neural network as an algorithm for calculating the likelihood. Further, the upper body likelihood calculating unit 14A and the lower body likelihood calculating unit 14B may use different algorithms as algorithms for calculating the likelihood. For example, the upper body likelihood calculation unit 14A may use a support vector machine, and the lower body likelihood calculation unit 14B may use a neural network.

  In the above embodiment, an example in which the pedestrian image 51 is input to the person detection device 1 during learning has been described, but the present invention is not limited to this. For example, the upper pedestrian image 51 </ b> A and the lower pedestrian image 51 </ b> B may be input to the person detection device 1. In this case, the upper pedestrian image 51A is directly input to the feature extraction unit 13A without going through the image conversion unit 11 and the detection window setting unit 12. Similarly, the lower pedestrian image 51B is directly input to the feature extraction unit 13B.

  Moreover, although the said part demonstrated the example in which the judgment part 15 performs a hidden detection process (step S55) when a pedestrian is detected in the detection window 23 in the said embodiment, it is not restricted to this. The determination unit 15 may not execute the hidden detection process (step S55).

  In the above embodiment, the example in which the size of the pedestrian image input to the object detection device 1 is the same as the size of the detection window 23 has been described, but the present invention is not limited to this. As long as the aspect ratio of the pedestrian image 51 input to the object detection device 1 matches the aspect ratio of the detection window 23, the size of the pedestrian image 51 may not be the same as the size of the detection window 23. In this case, before the upper window 31 and the lower window 41 are input to the feature extraction units 13A and 13B, the sizes of the upper window 31 and the lower window 41 are matched with the sizes of the upper pedestrian image 51A and the lower pedestrian image 51B. It should be changed as follows. Alternatively, before the upper feature data 32A and the lower feature data 42A are input to the upper body likelihood calculation unit 14A, and before the upper feature data 32B and the lower feature data 42B are input to the lower body likelihood calculation unit 14B, these features What is necessary is just to change the size of data so that it may correspond to the size of 51 A of upper pedestrian images, and the lower pedestrian image 51B. Alternatively, the size of the detection window 23 may be changed to match the size of the pedestrian image 31 before the upper window 31 and the lower window 41 are generated from the detection window 23.

  In the human detection device 1 described in the above embodiment, each functional unit may be individually made into one chip by a semiconductor device such as an LSI, or may be made into one chip so as to include a part or the whole. . The method of circuit integration is not limited to LSI, and may be realized by a dedicated circuit or a general-purpose processor. An FPGA (Field Programmable Gate Array) that can be programmed after manufacturing the LSI, or a reconfigurable processor that can reconfigure the connection and setting of circuit cells inside the LSI may be used.

  In addition, a part or all of the processing of each functional block of the above embodiment may be realized by a program. A part or all of the processing of each functional block in the above embodiment is performed by a central processing unit (CPU) in the computer. In addition, a program for performing each processing is stored in a storage device such as a hard disk or a ROM, and is read out and executed in the ROM or the RAM.

  Each processing of the above embodiment may be realized by hardware, or may be realized by software (including a case where the processing is realized together with an OS (Operating System), middleware, or a predetermined library). Further, it may be realized by mixed processing of software and hardware.

  A computer program that causes a computer to execute the above-described method and a computer-readable recording medium that records the program are included in the scope of the present invention. Here, examples of the computer-readable recording medium include a flexible disk, a hard disk, a CD-ROM, an MO, a DVD, a DVD-ROM, a DVD-RAM, a BD (Blu-ray Disc), and a semiconductor memory. .

  The computer program is not limited to the one recorded on the recording medium, and may be transmitted via a telecommunication line, a wireless or wired communication line, a network represented by the Internet, or the like.

DESCRIPTION OF SYMBOLS 1 Person detection apparatus 11 Image conversion part 12 Detection window setting part 13A, 13B Feature extraction part 14 Likelihood calculation part 14A Upper body likelihood calculation part 14B Lower body likelihood calculation part 15 Judgment part 16 Learning part 17 Storage part 21 Captured image 23, 123, 223, 323 Detection window 31, 123a, 223a, 323a Upper window 41, 123b, 223b, 323b Lower window 33A, 43A Upper body likelihood 33B, 43B Lower body likelihood

Claims (6)

A detection window setting unit that sets a first window area for a frame image and sets a second window area located below the first window area;
An upper body likelihood calculating unit that calculates a first upper body likelihood that the first window region will include the upper body of the person and a second upper body likelihood that the second window region will include the upper body of the person; ,
A lower body likelihood calculator that calculates a first lower body likelihood that the first window region will include the lower body of the person and a second lower body likelihood that the second window region will include the lower body of the person; ,
Judgment whether or not a person is present in the frame image based on the first upper body likelihood, the second upper body likelihood, the first lower body likelihood, and the second lower body likelihood And
A person detection device comprising: The person detection device according to claim 1,
The determination unit is configured such that an integrated value of the first upper body likelihood and the second lower body likelihood is not less than a first threshold value, and the second upper body likelihood and the first lower body likelihood are A person detection device that determines that a person is present in the frame image when the integrated value is equal to or less than a second threshold value. The person detection device according to claim 1,
The determination unit, when the difference absolute value between the first upper body likelihood and the second upper body likelihood is greater than or equal to a predetermined upper body reference value, or the first lower body likelihood and the second lower body likelihood A person detection device that determines that a person is present in the frame image when the absolute difference value of the frame is equal to or greater than a predetermined lower body reference value. The person detection device according to any one of claims 1 to 3, further comprising:
When the determination unit determines that a person is present in the frame image, a hidden detection unit that determines whether the upper body or lower body of the person in the frame image is hidden;
A person detection device comprising: The person detection device according to any one of claims 1 to 4, further comprising:
First feature extraction processing is performed on the first window region to generate first upper feature data, and first feature extraction processing is performed on the second window region to generate first lower feature data. A first feature extraction unit to be generated;
Second feature extraction processing is performed on the first window region to generate second upper feature data, and second feature extraction processing is performed on the second window region to generate second lower feature data. A second feature extraction unit to be generated;
With
The upper body likelihood calculating unit calculates the first upper body likelihood from the first upper feature data, calculates the second upper body likelihood from the first lower feature data,
The lower body likelihood calculating unit calculates the first lower body likelihood from the second upper feature data and calculates the second lower body likelihood from the second lower feature data. In the computer installed in the person detection device,
Setting a first window area for the frame image and setting a second window area located below the first window area;
Calculating a first upper body likelihood that the first window region will include the upper body of a person and a second upper body likelihood that the second window region will include the upper body of the person;
Calculating a first lower body likelihood that the first window region will include the lower body of a person and a second lower body likelihood that the second window region will include the lower body of the person;
Determining whether a person exists in the frame image based on the first upper body likelihood, the second upper body likelihood, the first lower body likelihood, and the second lower body likelihood. When,
A person detection program for running.