JP7810883B2 - Object Recognition System - Google Patents

Description

The present invention relates to an object recognition system that recognizes objects using point cloud data acquired by a three-dimensional sensor.

Conventionally, systems for monitoring monitored areas use point cloud data acquired by three-dimensional sensors placed within the monitored area to recognize the presence or absence of objects to be wary of, such as intruders. One example of technology related to three-dimensional sensors that recognize the three-dimensional shapes of objects within a monitored area is the laser radar device disclosed in Patent Document 1 below. This laser radar device is configured so that when a laser beam emitted by a light-emitting unit is deflected and irradiated externally by an oscillating mirror, the deflection direction of the oscillating mirror scans in both vertical and horizontal directions. The externally reflected light corresponding to the laser beam irradiation is received by a light-receiving unit via the oscillating mirror, thereby acquiring point cloud data as a measurement result of the three-dimensional shape of objects within a predetermined measurement range.

Patent No. 6772667

However, there is a limit to the range that can be monitored with a single 3D sensor, and the presence of obstacles such as pillars will result in blind spots that cannot be monitored. For this reason, using multiple 3D sensors is one way to expand the monitoring area and cover blind spots. However, the recognition unit, which uses the output from each 3D sensor to perform recognition processing to recognize objects within the monitoring range, needs to synchronize the output of each 3D sensor so that it can be handled in the same coordinate system. To do this, it is necessary to accurately grasp the relative positional relationship between each 3D sensor, including the direction of rotation. For this reason, the user can input information about the relative positional relationship between each 3D sensor into the recognition unit when installing each sensor, but this is not only cumbersome for the user, but also has the problem of impeding recognition processing if incorrect information is input.

The present invention was made to solve the above-mentioned problems, and its purpose is to provide a configuration that can identify the relative positional relationship between each three-dimensional sensor without receiving information from an external source.

In order to achieve the above object, the invention described in claim 1 of the claims is as follows:
a plurality of three-dimensional sensors (10a, 10b, 10c) each acquiring point cloud data as a measurement result of the three-dimensional shape of an object within a predetermined measurement range (S);
a recognition unit (20) capable of recognizing the object using the point cloud data acquired by each of the plurality of three-dimensional sensors;
An object recognition system (1) comprising:
The recognition unit
an overhead image generating unit (21) that generates overhead images (Pa, Pb) by projecting at least a portion of the point cloud data acquired by the three-dimensional sensor onto a plane perpendicular to the vertical direction based on the position of the three-dimensional sensor;
an extraction unit (21) that extracts a common shape (Ra, Rb) consisting of point cloud data of an object commonly included in the plurality of overhead images generated by the overhead image generation unit for each of the plurality of three-dimensional sensors;
a relative position specifying unit (21) that specifies a relative positional relationship between the plurality of three-dimensional sensors based on a result of aligning the bird's-eye images by rotating the bird's-eye images around the corners after moving the plurality of bird's-eye images so that the positions of the corners of the common shape coincide with each other;
The present invention is characterized by comprising:
The symbols in parentheses above indicate the correspondence with the specific means described in the embodiments to be described later.

In the invention of claim 1, in the recognition unit, an overhead image is generated by an overhead image generation unit so that at least a portion of the point cloud data acquired by the three-dimensional sensor is projected onto a plane (hereinafter also referred to as a reference plane) perpendicular to the vertical direction based on the position of the three-dimensional sensor, an extraction unit extracts a common shape made up of point cloud data of objects that are commonly included in the multiple overhead images generated for each of the multiple three-dimensional sensors, and after moving the multiple overhead images so that the positions of the corners of the common shape coincide, the overhead images are rotated around the corners to align them, and based on the result, the relative position identification unit identifies the relative positional relationship of the multiple three-dimensional sensors.

When multiple overhead images generated based on the positions of the three-dimensional sensors are moved so that the positions of the corners of the common shape coincide, and then the overhead images are rotated around the corners to align them, the three-dimensional sensors are scattered on the reference plane that includes the common shape. For example, when an overhead image based on the position of a second three-dimensional sensor is moved and rotated to align the positions of the corners of the common shape with respect to an overhead image based on the position of a first three-dimensional sensor, the first and second three-dimensional sensors are scattered on the reference plane to reflect their actual relative positional relationship. Therefore, the relative positional relationship between the three-dimensional sensors can be determined based on the results of the above-described alignment, without receiving external information such as relative position information input by a user or GPS information.

In the invention of claim 2, the overhead image generation unit generates an overhead image by excluding point cloud data of objects whose movement is estimated using previous measurement results from the point cloud data acquired by the three-dimensional sensor. As a result, the common shape is extracted from point cloud data of stationary objects rather than point cloud data of moving objects, making it possible to extract point cloud data suitable for alignment and more accurately identify the relative positional relationship between each three-dimensional sensor. In addition to reducing the impact of data acquisition noise, limiting the amount of information reduces the processing load.

In the invention of claim 3, the overhead image generation unit generates an overhead image by excluding point cloud data acquired by the three-dimensional sensor whose height from the floor in a predetermined measurement range is equal to or less than a predetermined value. This allows the common shape to be extracted from point cloud data of relatively tall objects (e.g., walls or pillars) rather than point cloud data of relatively short objects (e.g., people or transported goods), making it possible to extract point cloud data suitable for alignment and more accurately identify the relative positional relationship between each three-dimensional sensor. Furthermore, not only can the effects of data acquisition noise be reduced, but the processing load can also be reduced by limiting the amount of information.

In the invention of claim 4, in the recognition unit, a layout diagram of each object placed within a predetermined measurement range is acquired in advance by a layout diagram acquisition unit, and the absolute position identification unit identifies the position within the predetermined measurement range of the three-dimensional sensor using the layout diagram acquired by the layout diagram acquisition unit and the point cloud data acquired by the three-dimensional sensor. In this way, by using not only the point cloud data acquired by the three-dimensional sensor but also the layout diagram of each object placed within the predetermined measurement range, the position within the predetermined measurement range of the three-dimensional sensor can be identified more accurately compared to when simply using the point cloud data acquired by the three-dimensional sensor.

1 is a block diagram showing a schematic configuration of an object recognition system according to a first embodiment. FIG. 2 is a schematic perspective view illustrating the positional relationship between two three-dimensional sensors arranged within a monitoring area. FIG. 3 is an explanatory diagram illustrating an image of point cloud data acquired in the arrangement state of FIG. 2 . 10 is a flowchart illustrating the flow of a relative position specifying process performed by a control unit in the first embodiment. 4 is an explanatory diagram illustrating an overhead image generated from point cloud data of a three-dimensional sensor 10a in the state of FIG. 3; FIG. 4 is an explanatory diagram illustrating an overhead image generated from point cloud data of a three-dimensional sensor 10b in the state of FIG. 3; FIG. 7 is an explanatory diagram illustrating a state in which the bird's-eye view image shown in FIG. 6 is moved relative to the bird's-eye view image shown in FIG. 5 to align the positions of the corners of the common shape. FIG. 8 is an explanatory diagram illustrating a state in which the overhead image of FIG. 6 is rotated and aligned from the state of FIG. 7; FIG. FIG. 2 is an explanatory diagram illustrating a layout of objects placed in a part of a monitoring target area. FIG. 10 is a schematic perspective view illustrating the positional relationship between a three-dimensional sensor and each object that is placed within a monitored area that includes an area corresponding to the layout diagram of FIG. 9 . 10 is a flowchart illustrating the flow of an absolute position specifying process performed by a control unit in the second embodiment. 11 is an explanatory diagram illustrating an overhead image generated from point cloud data of a three-dimensional sensor in the state of FIG. 10 . FIG. 10 is an explanatory diagram illustrating a case where a plurality of position candidates are calculated that have a high degree of coincidence between a first feature shape and a second feature shape; FIG. FIG. 10 is an explanatory diagram illustrating a state in which the layout diagram of FIG. 9 has been corrected to remove line segments.

[First embodiment]
A first embodiment of an object recognition system according to the present invention will be described below with reference to the drawings.
The object recognition system 1 according to this embodiment is a system that recognizes objects within a predetermined measurement range by using point cloud data acquired by a plurality of three-dimensional sensors. In this embodiment, as shown in Fig. 1, the object recognition system 1 includes three-dimensional sensors 10a and 10b as a plurality of three-dimensional sensors arranged within a monitored area S corresponding to the predetermined measurement range, and a recognition device 20 that functions as a recognition unit that recognizes objects within the monitored area S by using the point cloud data acquired by each of the three-dimensional sensors 10a and 10b, and is configured to be able to detect intruders and the like entering the monitored area S.

Three-dimensional sensor 10a is configured so that when the infrared light emitted by the light-emitting unit is deflected and emitted externally by the oscillating mirror, the deflection direction by the oscillating mirror scans vertically and horizontally. The externally reflected light corresponding to this infrared light is received by the light-receiving unit via the oscillating mirror, thereby acquiring point cloud data as a measurement result of the three-dimensional shape of an object within a specified measurement range. Three-dimensional sensor 10b is configured similarly to three-dimensional sensor 10a, and its specific detailed configuration can be equivalent to that of the laser radar device disclosed in Patent No. 6772667, for example.

As shown in Figure 2, the three-dimensional sensors 10a and 10b configured in this manner are installed on the ceiling or other surface of the monitoring area S at a predetermined distance apart so that the entire monitoring area S can be monitored. The three-dimensional sensors 10a and 10b are configured to output the acquired point cloud data to the recognition device 20 via wired or wireless connection at predetermined times.

The recognition device 20 is, for example, a personal computer, and as shown in FIG. 1, in addition to a control unit 21 consisting of a CPU or the like and a storage unit 22 consisting of semiconductor memory or the like, it is equipped with a display unit 23 whose display content is controlled by the control unit 21, an operation unit 24 that outputs operation signals to the control unit 21 in response to input operations, and a communication unit 25 for communicating with external devices such as a monitoring server.

The recognition device 20 performs a recognition process to recognize objects within the monitored area S using the point cloud data output from the three-dimensional sensors 10a and 10b, respectively, and if an intruder or the like is detected during this recognition process, the recognition device 20 transmits the detection results or the like to the monitoring server via the communication unit 25. Note that Figure 3 shows an example of the point cloud data acquired from the three-dimensional sensors 10a and 10b in the arrangement shown in Figure 2.

As described above, in this embodiment, multiple three-dimensional sensors (three-dimensional sensor 10a and three-dimensional sensor 10b) are used instead of a single three-dimensional sensor, thereby expanding the monitored area S and covering blind spots. When multiple three-dimensional sensors are used in this way, the outputs of each three-dimensional sensor must be synchronized so that they can be handled in the same coordinate system, and to do this, it is necessary to accurately grasp the relative positional relationship between each three-dimensional sensor, including the rotational direction.

In this embodiment, after installing three-dimensional sensors 10a and 10b on the ceiling or the like of the monitored area S, the relative positional relationship between three-dimensional sensors 10a and 10b, including the rotational direction, is determined by a relative position determination process performed by the control unit 21 of the recognition device 20. Information relating to the determined relative positional relationship (hereinafter also referred to as relative positional relationship information) is stored in the memory unit 22. Then, in the recognition process performed thereafter by the control unit 21, the relative positional relationship information previously stored in the memory unit 22 as described above is used to synchronize the point cloud data acquired by both three-dimensional sensors 10a and 10b so that they can be handled in the same coordinate system, and the synchronized point cloud data is used to perform processing to recognize objects within the monitored area S.

Hereinafter, the relative position specifying process performed by the control unit 21 of the recognition device 20 in this embodiment will be specifically described with reference to the flowchart shown in FIG.
When the control unit 21 starts the relative position identification process in response to a predetermined operation on the operation unit 24, the control unit 21 first performs the three-dimensional coordinate information conversion process shown in step S101 of Fig. 4. In this process, the point cloud data acquired from the three-dimensional sensor 10a, which is the first three-dimensional sensor, is converted into three-dimensional coordinate information (x, y, z) with the position of the three-dimensional sensor 10a as the origin (reference).

Next, the overhead image generation process shown in step S103 is performed, and the three-dimensional coordinate information converted as described above is projected onto a plane (reference plane) perpendicular to the vertical direction (z-axis), thereby generating an overhead image. With the three-dimensional sensor 10a arranged as shown in FIG. 2, for example, an overhead image Pa such as that shown in FIG. 5 is generated. Note that the control unit 21 that performs the overhead image generation process can be considered an example of a "bird's-eye image generation unit."

Next, in the determination process shown in step S105, it is determined whether or not overhead images have been generated for all three-dimensional sensors. Because an overhead image has not been generated for three-dimensional sensor 10b (No in S105), the point cloud data for three-dimensional sensor 10b is also converted into three-dimensional coordinate information with the position of that three-dimensional sensor 10b as the origin (S101), and an overhead image is generated by projecting the converted three-dimensional coordinate information onto a plane (reference plane) perpendicular to the vertical direction (S103). For three-dimensional sensor 10b positioned as shown in Figure 2, for example, an overhead image Pb such as that shown in Figure 6 is generated.

Once overhead images are generated for all three-dimensional sensors as described above (Yes in S105), the common shape extraction process shown in step S107 is performed. In this process, a common shape consisting of point cloud data of objects commonly included in each overhead image is extracted. When overhead images Pa and Pb are generated as described above, common shapes Ra and Rb shown in Figures 5 and 6 are extracted. Note that the control unit 21 that performs the common shape extraction process can be considered an example of an "extraction unit."

Next, the alignment process shown in step S109 is performed, in which the multiple overhead images are moved and rotated to align them so that the common shapes match. Specifically, for example, the overhead image Pb of FIG. 6 is moved and superimposed on the overhead image Pa of FIG. 5 so that the positions of corners Rao and Rbo, which are easily recognizable parts of the common shapes Ra and Rb extracted as described above, match (see FIG. 7). Furthermore, the overhead image Pb is rotated around the corners Rao and Rbo so that the surface of the common shape Ra in overhead image Pa and the surface of the common shape Rb in overhead image Pb overlap (see FIG. 8). With this alignment, as can be seen from FIG. 8, the three-dimensional sensors 10a and 10b are scattered on the reference plane to reflect their actual relative positional relationship.

For this reason, the relative positional relationship between the two three-dimensional sensors 10a, 10b, including the direction of rotation, can be determined from the scattered state described above, and in the relative positional relationship information storage process shown in step S111, relative positional relationship information regarding the relative positional relationship determined as described above is stored in the memory unit 22. Note that the control unit 21, which performs the process of determining the relative positional relationship between the two three-dimensional sensors 10a, 10b, including the direction of rotation, from the scattered state described above, can correspond to an example of a "relative position determination unit."

As described above, in the object recognition system 1 according to this embodiment, in the relative position identification process performed by the control unit 21 of the recognition device 20, overhead images (Pa, Pb) are generated by projecting the point cloud data acquired by the three-dimensional sensors (10a, 10b) onto a reference plane based on the positions of the three-dimensional sensors, a common shape (Ra, Rb) consisting of point cloud data of objects commonly included in the multiple overhead images generated by the multiple three-dimensional sensors is extracted, and the multiple overhead images are moved and rotated to align the common shape, and the relative positional relationship of the multiple three-dimensional sensors is identified based on the result.

When the overhead image Pb based on the position of the three-dimensional sensor 10b is moved and rotated to align it with the overhead image Pa based on the position of the three-dimensional sensor 10a so that the common shapes Ra and Rb coincide, the three-dimensional sensors 10a and 10b are scattered on the reference plane in a manner that reflects their actual relative positional relationship. Therefore, the relative positional relationship between each three-dimensional sensor can be automatically determined based on the results of the above-described alignment, without receiving external information such as relative position information input by the user or GPS information.

As a first variation of this embodiment, the overhead image generation process may generate an overhead image by excluding point cloud data of objects whose movement is estimated using previous measurement results from the point cloud data acquired by the three-dimensional sensor. Specifically, for example, point cloud data whose difference from the previous or previous measurement result is equal to or less than a predetermined value is defined as point cloud data of stationary objects, and point cloud data whose difference from the previous or previous measurement result exceeds the predetermined value is defined as point cloud data of moving objects, and an overhead image is generated using only the point cloud data of stationary objects, excluding the point cloud data of moving objects (point cloud data of objects whose movement is estimated).

As a result, the common shape is extracted from point cloud data of stationary objects rather than point cloud data of moving objects, making it possible to extract point cloud data suitable for alignment and more accurately identify the relative positional relationship between each 3D sensor. In addition to reducing the impact of data acquisition noise, limiting the amount of information also reduces the processing load.

Furthermore, as a second variation of this embodiment, in the overhead image generation process, the overhead image may be generated by excluding point cloud data acquired by the three-dimensional sensor whose height from the floor in the monitored area (predetermined measurement range) S is equal to or less than a predetermined value. Specifically, for example, the predetermined value may be set to 2 m, and the overhead image may be generated using only point cloud data whose height from the floor exceeds 2 m, excluding point cloud data whose height from the floor is equal to or less than 2 m.

As a result, the common shape is extracted from point cloud data of relatively tall objects (such as walls or pillars) rather than point cloud data of relatively short objects (such as people or transported goods), making it possible to extract point cloud data suitable for alignment and more accurately identify the relative positional relationship between each 3D sensor. In addition to reducing the impact of data acquisition noise, limiting the amount of information also reduces the processing load.

In addition, in the above-mentioned overhead image generation process, the overhead image is not limited to being generated by projecting all point cloud data acquired by the three-dimensional sensor onto the reference plane based on the position of the three-dimensional sensor. The overhead image may also be generated by projecting a portion of the point cloud data acquired by the three-dimensional sensor onto the reference plane based on the position of the three-dimensional sensor (for example, point cloud data within a specified distance range from the three-dimensional sensor).

Second Embodiment
Next, an object recognition system according to a second embodiment of the present invention will be described with reference to the drawings.
The second embodiment differs from the first embodiment mainly in that the position of the three-dimensional sensor is identified using a layout diagram acquired in advance. Therefore, components that are substantially the same as those in the first embodiment are given the same reference numerals and their description will be omitted.

When deciding where to install a 3D sensor on the ceiling or other surface in a monitored area, maps or actual objects may be used as reference. However, if there are no clear landmarks in the surrounding area (for example, if there are few distinctively shaped objects), it can be difficult to accurately determine the installation position of the 3D sensor relative to the monitored area. For this reason, if a system administrator, or other person other than the person who installed the 3D sensor, estimates and registers the position and rotation angle (detection angle) of the 3D sensor in the monitored area, there is a risk of measurement errors occurring due to discrepancies in the installation position and installation angle.

Therefore, in the object recognition system 1 according to this embodiment, a layout diagram M of at least some of the objects located within the monitored area (predetermined measurement range) S is acquired in advance and stored in the memory unit 22. The control unit 21 of the recognition device 20 performs absolute position identification processing to identify the position of the three-dimensional sensor within the monitored area S using the layout diagram M stored in advance in the memory unit 22 and the point cloud data acquired by the three-dimensional sensor. By accurately identifying the position of each three-dimensional sensor within the monitored area S in this manner, the accuracy of the point cloud data acquired in the relative position identification processing in the first embodiment can be improved. In this embodiment, the layout diagram M of each object is, for example, an architectural drawing (a floor plan created from the same perspective as the aerial image) as shown in FIG. 9, which is composed of simple line segments. The layout diagram M is acquired from the monitoring server via the communication unit 25 by the layout diagram acquisition processing performed by the control unit 21 and stored in the memory unit 22. The control unit 21 and communication unit 25 that perform the layout diagram acquisition processing in this manner may correspond to an example of a "layout diagram acquisition unit." Furthermore, the control unit 21 that performs the absolute position identification process can be an example of an "absolute position identification unit."

The following describes in detail, with reference to the flowchart shown in Figure 11, the absolute position identification process performed by the control unit 21 of the recognition device 20 when one three-dimensional sensor 10c is installed as shown in Figure 10 within a monitored area S that includes an area corresponding to the layout diagram M in Figure 9.

When the control unit 21 starts the absolute position identification process in response to a predetermined operation on the operation unit 24, the three-dimensional coordinate information conversion process shown in step S201 of FIG. 11 is first performed. In this process, similar to the three-dimensional coordinate information conversion process shown in step S101 described above, the point cloud data acquired from the three-dimensional sensor 10c, which is the three-dimensional sensor whose absolute position is to be identified, is converted into three-dimensional coordinate information (x, y, z) with the position of the three-dimensional sensor 10c as the origin (reference).

Next, the overhead image generation process shown in step S203 is performed. Similar to the overhead image generation process shown in step S103 described above, the three-dimensional coordinate information converted as described above is projected onto a plane (reference plane) perpendicular to the vertical direction (z-axis), thereby generating an overhead image. With a three-dimensional sensor 10c positioned as shown in Figure 10, for example, an overhead image Pc like the one shown in Figure 12 is generated.

Next, the first characteristic shape extraction process shown in step S205 is performed. In this process, a characteristic shape that is easily recognizable in the overhead image generated as described above (for example, a shape with approximately 90° corners) is extracted as the first characteristic shape. When the overhead image Pc is generated as described above, for example, the characteristic shape R1 shown in Figure 12 is extracted as the first characteristic shape.

Next, the second characteristic shape extraction process shown in step S207 is performed. In this process, an easily recognizable characteristic shape is extracted as the second characteristic shape from the layout drawing M stored in advance in the storage unit 22. If a layout drawing M such as the one shown in FIG. 9 is stored, for example, the characteristic shape R2 shown in FIG. 9 is extracted as the second characteristic shape. Note that the second characteristic shape does not necessarily have to be extracted from the layout drawing M each time, but may also be extracted in advance when the layout drawing M is acquired and stored in the storage unit 22.

Next, the sensor position candidate calculation process shown in step S209 is performed. In this process, a matching process is performed based on the first characteristic shape (characteristic shape R1) extracted from the bird's-eye view image Pc of the position reference of the three-dimensional sensor 10c and the second characteristic shape (characteristic shape R2) extracted from the layout drawing M in which the position of the three-dimensional sensor 10c is not taken into account, and a position candidate on the layout drawing M of the three-dimensional sensor 10c that has a high matching rate (the degree to which the two shapes match is high) is calculated, including the zoom rate and rotation direction. Note that the matching rate can be calculated, for example, as equivalent to the value obtained by dividing the number of point data (number of sensor feature matches) considered to match for the characteristic shapes by the total number of point data (number of sensor features).

If one position candidate is calculated (Yes in S211), the absolute position information storage process shown in step S217 is performed. In this process, the three-dimensional sensor is assumed to be located at that position candidate, and information about the position candidate, including the rotation direction within the monitored area S, obtained from the relationship between the position of the position candidate in the layout drawing M and the area occupied by the layout drawing M in the monitored area S, is stored in the memory unit 22 as information about the three-dimensional sensor's position within the monitored area S (absolute position information).

On the other hand, when multiple position candidates are calculated on the layout diagram M of the three-dimensional sensor 10c with a high matching rate, for example, two of the three position candidates C1 to C3, C1 and C2, may be calculated with a matching rate of nearly 100%, as shown in Figure 13. This is because when extracting an easily recognizable feature shape on the layout diagram M, if the shape is simply determined by line segments without taking into account factors such as wall thickness, multiple position candidates will be calculated, as shown in Figure 13. Note that for ease of explanation, the position candidates C1 to C3 are shown slightly offset from each other in Figure 13.

For this reason, when multiple position candidates are calculated as described above (No in S211), the second characteristic shape correction process shown in step S213 is performed. In this process, since the approximate position of the three-dimensional sensor 10c on the layout drawing M is known, the layout drawing M is corrected to remove line segments that are behind the approximate position of the three-dimensional sensor 10c from the second characteristic shapes extracted as described above, and a characteristic shape that is easy to recognize in this corrected layout drawing Mc is newly extracted as the second characteristic shape R2c. For example, the layout drawing M shown in FIG. 9 is corrected to look like the layout drawing Mc shown in FIG. 14. This makes it possible to remove line segments that cannot be recognized by the three-dimensional sensor 10c from the layout drawing.

Once the layout plan has been corrected as described above, the sensor position candidate calculation process shown in step S215 is performed, and a matching process is performed based on the first characteristic shape R1 extracted from the overhead image Pc and the second characteristic shape R2c extracted from the corrected layout plan Mc. This calculates one position candidate that includes the rotational direction of the three-dimensional sensor 10c on the layout plan Mc, which results in a high matching rate, and information about this position candidate is stored in the memory unit 22 as absolute position information (S217).

As described above, in the object recognition system 1 according to this embodiment, the recognition device 20 acquires in advance a layout diagram M of each object placed within the monitored area S, and the position of the three-dimensional sensor within the monitored area S is identified by an absolute position identification process performed by the control unit 21 using the layout diagram M thus acquired and the point cloud data acquired by the three-dimensional sensor. In this way, by using not only the point cloud data acquired by the three-dimensional sensor but also the layout diagram M of each object placed in the monitored area S, the position of the three-dimensional sensor within the monitored area S can be identified more accurately compared to when simply using the point cloud data acquired by the three-dimensional sensor.

The present invention is not limited to the above-described embodiments, and may be embodied as follows, for example.
(1) The three-dimensional sensors 10a to 10c are not limited to being installed on the ceiling or the like of the monitored area (predetermined measurement range) S, but may be installed near the ceiling by being attached to the top of a support, for example.

(2) In the first embodiment and the like, the multiple three-dimensional sensors are not limited to two three-dimensional sensors 10a and 10b. Three or more three-dimensional sensors may be used, and the relative positional relationship between the three-dimensional sensors, including the rotational direction, may be determined by aligning the respective overhead images.

1... Object recognition system 10a, 10b, 10c... Three-dimensional sensor 20... Recognition device (recognition unit)
21...Control unit (bird's-eye view image generation unit, extraction unit, relative position identification unit, layout diagram acquisition unit, absolute position identification unit)
M, Mc... Layout diagram Ra, Rb... Common shape S... Monitoring area (predetermined measurement range)
Pa, Pb, Pc... Bird's-eye view image

Claims (4)

a plurality of three-dimensional sensors each acquiring point cloud data as a measurement result of the three-dimensional shape of an object within a predetermined measurement range;
a recognition unit capable of recognizing the object by using the point cloud data acquired by each of the plurality of three-dimensional sensors;
An object recognition system comprising:
The recognition unit
an overhead image generating unit that generates an overhead image by projecting at least a portion of the point cloud data acquired by the three-dimensional sensor onto a plane perpendicular to a vertical direction based on the position of the three-dimensional sensor;
an extraction unit that extracts a common shape made up of point cloud data of an object that is commonly included in the plurality of overhead images generated by the overhead image generation unit for each of the plurality of 3D sensors;
a relative position specifying unit that specifies a relative positional relationship between the plurality of three-dimensional sensors based on a result of aligning the bird's-eye images by rotating the bird's-eye images around the corners after moving the plurality of bird's-eye images so that the positions of the corners of the common shape coincide with each other;
An object recognition system comprising: The object recognition system described in claim 1, characterized in that the overhead image generation unit generates the overhead image by excluding point cloud data of objects whose movement is estimated using previous measurement results from the point cloud data acquired by the three-dimensional sensor. The object recognition system described in claim 1 or 2, characterized in that the overhead image generation unit generates the overhead image by excluding, from the point cloud data acquired by the three-dimensional sensor, point cloud data whose height from the floor in the specified measurement range is equal to or less than a specified value. The recognition unit
a layout diagram acquisition unit that acquires in advance a layout diagram of each object that is to be placed within the predetermined measurement range;
an absolute position specifying unit that specifies a position within the predetermined measurement range of the three-dimensional sensor by using the layout plan acquired by the layout plan acquisition unit and the point cloud data acquired by the three-dimensional sensor;
The object recognition system according to any one of claims 1 to 3, comprising: