Object Recognition System

Abstract

To achieve the high object recognition accuracy in an object recognition system using a 3D point group acquired by a 3D measurement sensor mounted on a vehicle.SOLUTION: An object recognition system 1e includes: a 3D measurement sensor 4; a 3D point cloud processing part 10c; an object recognition processing part 14e which performs the object recognition processing on a bird's-eye view image of the own vehicle sent from the 3D point cloud processing part 10c; and a 3D point cloud generation processing part 17e with object information which generates a 3D point group with the object information by associating a 3D point cloud of the own vehicle with the bird's-eye view image of the own vehicle which has been subjected to the object recognition processing transmitted from the object recognition processing part 14e. The highly accurate object recognition is possible by obtaining a 3D point cloud with the object information in real time while making a vehicle travel independently.SELECTED DRAWING: Figure 8

Description

[Technical field]

[0001] The present invention relates to a high-precision object recognition system that acquires a 3D point cloud of the surrounding area using an on-board 3D measurement means while moving a means of transportation such as a vehicle, and uses this to perform recognition processing of objects in the vicinity. In particular, the present invention relates to an object recognition system in a single means of transportation that accumulates the 3D point cloud it acquires itself, performs imaging processing to generate a bird's-eye image, and then performs object recognition processing on this. [Background technology]

[0002] To recognize obstacles such as people and vehicles, which is essential for autonomous driving, 3D measurement sensors that measure the surrounding 3D shape, such as LiDAR (Light Detection and Ranging) mounted on the vehicle, are used. However, if a 3D point cloud is acquired only by the sensor mounted on the vehicle, measurement data can only be obtained in a narrow range near the vehicle due to the limitations of the sensor's measurement distance and range. In addition, only a part of the surface of an object can be measured, and the entire shape of the object cannot be grasped. Furthermore, depending on the surrounding environment, there may be areas that are blind spots of the sensor.

[0003] Non-Patent Document 1 discloses an invention of an object recognition system that shares and integrates 3D point clouds acquired by LiDAR among multiple vehicles, and uses the integrated 3D point cloud to estimate the position and orientation of objects such as surrounding vehicles.

[0004] Non-Patent Document 2 discloses an invention that performs object recognition processing using current measurement data obtained from a LiDAR mounted on a vehicle. [Prior art documents] [Non-patent literature]

[0005] [Non-Patent Document 1] Qi Chen et al., “Cooper: Cooperative Perception for Connected Autonomous Vehicles based on 3D Point Clouds,” 2019 IEEE 39th International Conference on Distributed Computing Systems(ICDCS), 2019, pp. 514-524 [Non-Patent Document 2] Alejandro Barrera et al., “BirdNet+: End-to-End 3D Object Detection in LiDAR Bird's Eye View,” 2020 IEEE 23rd International Conference on Intelligent Transportation System(ITSC), 2020, pp. 1-6 Summary of the Invention [Problem to be solved by the invention]

[0006] In order to integrate the 3D point clouds acquired by each of the multiple vehicles, data on the relative positions between the vehicles is required. In the invention disclosed in Non-Patent Document 1, a given condition is that highly accurate data on the relative positions of the multiple vehicles is provided. That is, in this invention, the positions of the multiple vehicles are predetermined, and an ideal situation far from a realistic situation in which each vehicle transmits and receives data in real time while moving is assumed. Therefore, when integrating the 3D point clouds acquired by each of the multiple vehicles while moving, if the data on the relative positions between the vehicles contains an error, it is inevitable that the object recognition accuracy will be greatly reduced. In addition, there is a problem that a large capacity communication line is required to transmit and receive the 3D point cloud, which is a huge amount of data, between multiple vehicles and share it, and this is only possible when the communication conditions are good.

[0007] In addition, the invention disclosed in Non-Patent Document 2 uses only the 3D point cloud at the current time, and depending on the sensor installed, only a sparse 3D point cloud can be obtained, resulting in insufficient object recognition accuracy.

[0008] The present invention has been made in consideration of the conventional technology described above, and has a first object recognition system that acquires a 3D point cloud of the surrounding area while moving using a 3D measurement means mounted on a vehicle, and performs object recognition processing on an image generated from the 3D point cloud, when performing object recognition processing on an image generated by a single means of transportation, to achieve high object recognition accuracy. [Means for solving the problem]

[0009] The object recognition system according to claim 1 comprises: Single Means of transportation; a 3D measurement means mounted on the moving means and measuring the surroundings to obtain a 3D point cloud; A control means mounted on the moving means for performing surrounding object recognition based on an object recognition process using the three-dimensional point cloud; An object recognition system comprising: The control means A self-position estimation process for estimating the position of the moving means Department and , The three-dimensional point cloud obtained is stored by converting the obtained three-dimensional point cloud into a coordinate system of the moving means at the current time based on the position of the moving means at the time when the three-dimensional point cloud was obtained and the position of the moving means at the current time. thing This allows us to obtain a cumulative 3D point cloud that is denser than the 3D point cloud obtained at the current time. A point cloud accumulation processing unit that , Imaging process to generate a bird's-eye view image from the accumulated 3D point cloud Department and , The object recognition process is performed on the bird's-eye view image to perform object recognition for estimating the object type for each pixel constituting the bird's-eye view image. An object recognition processing unit; a processing unit for generating a three-dimensional point cloud with object information by linking a result of performing the object recognition processing on the bird's-eye image with the three-dimensional point cloud, thereby generating a three-dimensional point cloud with object information; having The vehicle travels independently while obtaining a 3D point cloud with object information in real time. It is characterized by the following. Effect of the Invention

[0011] According to the object recognition system described in claim 1, the point cloud accumulation process is continued while performing the self-position estimation process, thereby obtaining an accumulated three-dimensional point cloud that is denser than only the three-dimensional point cloud obtained at the current time, and object recognition can be performed by performing object recognition processing on the bird's-eye image generated from this accumulated three-dimensional point cloud, thereby achieving high object recognition accuracy with an object recognition system that has a single means of transportation. In addition, by linking the results of object recognition processing performed on the integrated bird's-eye view image or the bird's-eye view image with a 3D point cloud, a 3D point cloud with object information can be generated. [Brief description of the drawings]

[0013] [Figure 1] 1 is an overall configuration diagram of an object recognition system according to a first embodiment. [Diagram 2] 1 is a functional block diagram of an object recognition system according to a first embodiment. [Diagram 3] 2 is an explanatory diagram showing a procedure for generating an integrated bird's-eye image from a 3D point cloud measured by a vehicle itself and a 3D point cloud measured by another vehicle in the object recognition system according to the first embodiment. FIG. [Figure 4] FIG. 11 is an overall configuration diagram of an object recognition system according to a second embodiment. [Diagram 5] FIG. 11 is a functional block diagram of an object recognition system according to a second embodiment. [Figure 6] FIG. 11 is a functional block diagram of an object recognition system according to a third embodiment. [Figure 7] FIG. 13 is a functional block diagram of an object recognition system according to a fourth embodiment. [Figure 8] FIG. 13 is a functional block diagram of an object recognition system according to a fifth embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0014] A first embodiment of the present invention will be described with reference to FIGS. As shown in FIG. 1, the object recognition system 1a of the first embodiment measures objects around each vehicle 2 while driving a plurality of vehicles 2, 2 as a means of transportation, acquires a three-dimensional point cloud, wirelessly transmits and receives data between the vehicles as shown by the arrows, and performs data integration processing in each vehicle 2, and performs object recognition processing on the integrated data to categorize objects around each vehicle 2. Specifically, each part of the integrated bird's-eye image, which is the integrated data, can be divided into categories that are considered to be the same in advance, such as cars, people, buildings, etc., and displayed by color coding or the like. Note that, although two vehicles 2, 2 are shown in FIG. 1, three or more vehicles may be used. In the following description, one of the two vehicles 2, 2 shown in the figure is referred to as the subject vehicle 2a, and the other vehicle is referred to as the other vehicle 2b.

[0015] As shown in FIG. 1, each vehicle 2 is equipped with a position and attitude sensor 3 that acquires data regarding position and attitude (hereinafter also simply referred to as "position and attitude"), and a 3D measurement sensor 4 as a 3D measurement means that acquires 3D point cloud data (hereinafter also simply referred to as "3D point cloud") by measuring the surroundings.

[0016] GNSS (short for Global Navigation Satellite System), which is a general term for systems that calculate the position of a receiver from the distance between the receiver and multiple satellites, can be used as the position and attitude sensor 3. Specific examples of GNSS include GPS, GLONASS, and Galileo.

[0017] As an example of the 3D measurement sensor 4, LiDAR (Light Detection And Ranging) can be used. LiDAR can measure the distance to an object by emitting a laser from the sensor and measuring the time it takes for the laser to reflect off the object and return. With 3D LiDAR, a 3D point cloud (a collection of [x, y, z] coordinates) can be obtained by stacking multiple laser emitting parts vertically that can scan 360 degrees by rotating.

[0018] The position and orientation sensor 3 and the three-dimensional measuring sensor 4 are disposed at predetermined positions on the roof of the vehicle 2 so that they can perform their functions most effectively. Inside the vehicle 2, a data processing unit 5a is provided as a control means for processing data acquired by the position and orientation sensor 3 and the three-dimensional measuring sensor 4 and performing object recognition. As the data processing unit 5a, a personal computer or the like having an application for executing data processing, which will be described later, implemented can be used. Inside the vehicle 2, a wireless communication device 6 is provided as a communication means. The wireless communication device 6 has a transmitter 7 for transmitting data processed by the data processing unit 5a to other vehicles 2, a receiver 8 for receiving data transmitted from other vehicles 2 and sending it to the data processing unit 5a, and an antenna 9 provided on the roof of the vehicle 2 and connected to the transmitter 7 and the receiver 8.

[0019] 2, the data processing unit 5a has a three-dimensional point cloud processing unit 10a. The position and orientation sensor 3 and the three-dimensional measurement sensor 4 are connected to the three-dimensional point cloud processing unit 10a, and the position and orientation of the host vehicle acquired by the position and orientation sensor 3 and the three-dimensional point cloud of the host vehicle acquired by the three-dimensional measurement sensor 4 are input. The three-dimensional point cloud processing unit 10a also has an imaging processing unit 11. The imaging processing unit 11 converts the three-dimensional point cloud acquired by the three-dimensional measurement sensor 4 into a bird's-eye image viewed from above the vehicle 2 directly below.

[0020] As shown in FIG. 3(a), a 3D LiDAR, which is a 3D measurement sensor 4 mounted on a vehicle 2 (own vehicle 2a), scans the entire surroundings (360°) of the own vehicle by rotating while emitting a laser, acquires a 3D point cloud (a set of [x, y, z] coordinates), and sends it to an imaging processing unit 11 (see FIG. 2). In the example of FIG. 3(a), a set of [x, y] coordinates at a predetermined position in the height direction [z] is displayed, but in reality, a 3D point cloud including values ​​in the height direction [z] is acquired, and the three-dimensional shapes of various objects S present around the own vehicle 2a can be grasped by this 3D point cloud. In addition, as shown in FIG. 3(b), a similar measurement is also performed on another vehicle 2b in the vicinity of the own vehicle 2a, and a 3D point cloud is acquired. Note that Figure 3 shows a case where the measurement range of the host vehicle 2a partially overlaps with the measurement range of another vehicle 2b located in the vicinity of the host vehicle 2a. In this case, by integrating the data obtained by the host vehicle 2a and the data obtained by the other vehicle 2b, as will be described later, an integrated bird's-eye view image showing a single continuous area can be generated as shown in Figure 3(e).

[0021] In the 3D point cloud processing unit 10a of the vehicle 2a, the imaging processing unit 11 converts the 3D point cloud shown in FIG. 3(a) input from the 3D measurement sensor 4 into a bird's-eye image shown in FIG. 3(c) which is 2D data. At that time, data processing is performed so that each pixel constituting the bird's-eye image includes not only position data but also standardized data indicating height, reflection intensity, and density, so that a specific set of pixels grasped in the planar bird's-eye image can be identified as a specific type of object S (e.g., a car, a person, a building, etc.). Similarly to the vehicle 2a, the other vehicle 2b in the vicinity of the vehicle 2a also performs imaging processing of the 3D point cloud acquired by the other vehicle shown in FIG. 3(b), and a bird's-eye image as shown in FIG. 3(d) is acquired.

[0022] As shown in FIG. 2, the 3D point cloud processing unit 10a of the vehicle 2a sends the bird's-eye image of the vehicle 2a generated by the imaging processing unit 11 and the position and attitude of the vehicle 2a input from the position and attitude sensor 3 to the wireless communication device 6, and transmits them to the other vehicle 2b by the transmitter 7 and antenna 9. The 3D point cloud has a large data capacity, and depending on the communication capacity of the communication line, it may be difficult to transmit and receive in real time. However, in this embodiment, what is sent to the other vehicle 2b is a bird's-eye image with a small data volume obtained by imaging the 3D point cloud, so the communication volume can be reduced and it is easy to share it by vehicle-to-vehicle communication. In addition, the 3D point cloud processing unit 10a of the vehicle 2a sends the position and attitude of the vehicle 2a input from the position and attitude sensor 3 and the 3D point cloud of the vehicle 2a to the relative position estimation processing unit 12. In addition, the 3D point cloud processing unit 10a of the vehicle 2a sends the bird's-eye image of the vehicle 2a generated by the imaging processing unit 11 to the data integration processing unit 13 described later.

[0023] As shown in FIG. 2, the data processing unit 5a includes a relative position estimation processing unit 12. The relative position estimation processing unit 12 receives the three-dimensional point cloud of the vehicle 2a and the position and attitude of the vehicle 2a from the three-dimensional point cloud processing unit 10a, and also receives the bird's-eye image of the other vehicle 2b and the position and attitude of the other vehicle 2b received by the receiving unit 8 of the wireless communication device 6, calculates the inter-vehicle relative position and attitude, which is data indicating the relative positional relationship and attitude between the vehicle 2a and the other vehicle 2b, and sends it to the data integration processing unit 13. The relative position estimation processing in this embodiment is accurate and fast because it is calculated from coordinate data obtained by GNSS. The bird's-eye image of the other vehicle 2b received by the receiving unit 8 of the wireless communication device 6 is sent directly to the data integration processing unit 13 via the relative position estimation processing unit 12.

[0024] As shown in Fig. 2, the data processing unit 5a includes a data integration processing unit 13. The data integration processing unit 13 converts the bird's-eye image of the other vehicle 2b shown in Fig. 3(d) sent via the relative position estimation processing unit 12 based on the inter-vehicle relative position and attitude sent from the relative position estimation processing unit 12 into data in the coordinate system of the bird's-eye image of the host vehicle 2a shown in Fig. 3(c) sent from the three-dimensional point cloud processing unit 10a, and integrates it with the bird's-eye image of the host vehicle 2a to generate an integrated bird's-eye image as shown in Fig. 3(e), which is a single continuous image data.

[0025] 2, the data processing unit 5a includes an object recognition processing unit 14. The object recognition processing unit 14 performs semantic segmentation processing using, for example, CNN on the integrated bird's-eye image generated by the data integration processing unit 13, and estimates the object type for each pixel constituting the bird's-eye image.

[0026] Semantic segmentation involves classifying the smallest unit of recognition data (each pixel in an image, or each point when dealing with 3D point clouds) into predefined categories that are considered to be the same, such as cars, people, buildings, etc., and in recent years, methods using CNN have become mainstream.

[0027] CNN is an abbreviation for Convolutional Neural Network, and is an algorithm that extracts image features by repeatedly convolving and downsampling input image data using a filter of approximately 3 x 3 size.

[0028] In this embodiment, a different color is determined for each category to be classified. Then, for each pixel of the integrated bird's-eye image shown in FIG. 3(e), the color of the category to which the pixel belongs is assigned, so that the object S recognized in the integrated bird's-eye image is colored, and a planar image in which the distinction between cars, people, buildings, etc. can be easily understood can be obtained. Although it is possible to perform semantic segmentation processing on a three-dimensional point cloud, in this embodiment, the semantic segmentation processing is performed on a bird's-eye image with a small amount of data obtained by converting the three-dimensional point cloud, so that the processing is faster than when object recognition processing is performed on a three-dimensional point cloud.

[0029] The second embodiment will be described with reference to FIGS. 4 and 5, focusing on the differences from the first embodiment. In the configuration of the second embodiment, the same parts as those in the first embodiment are denoted by the same reference numerals in the drawings, and the description of the first embodiment is to be incorporated herein by reference.

[0030] As shown in Fig. 4, the object recognition system 1b of the second embodiment generates an integrated bird's-eye image in each vehicle 2 by transmitting and receiving data between the vehicles while a plurality of vehicles 2, 2 are traveling, and performs object recognition processing on the image to categorize surrounding objects S. As shown in Fig. 4 and the functional block diagram Fig. 5, each vehicle 2 is not equipped with a position / attitude sensor 3 for acquiring the position / attitude of the vehicle 2a, but as shown in Fig. 5, a 3D point cloud processing unit 10b of a data processing unit 5b includes a self-position estimation processing unit 15 not included in the first embodiment.

[0031] That is, in the first embodiment, each vehicle 2, for example, the vehicle itself 2a, acquires the position and attitude of the vehicle itself 2a by means of the mounted position and attitude sensor 3, but in the first embodiment, the self-position estimation processing unit 15 estimates the position of the vehicle itself 2a, and the result is used in data processing as the position and attitude of the vehicle itself 2a.

[0032] First of all, SLAM can be mentioned as a method of self-location estimation processing performed by the self-location estimation processing unit 15. SLAM is an abbreviation for Simultaneous Localization And Mapping (processing for simultaneously estimating self-location and creating a map), and is a method for estimating a self-location by comparing data measured by a sensor at a certain time with measurement data from one time before, and creating a map by arranging the measurement data based on the estimated self-location.

[0033] As another method of self-location estimation processing, a method of estimating self-location by aligning measurement data by performing scan matching such as ICP between multiple 3D point clouds obtained by multiple vehicles 2 can be adopted. ICP is an abbreviation for Iterative Closest Point, and is a method of optimally aligning point cloud data by overlapping different point cloud data, alternately matching the closest point from each point and minimizing the sum of the distance. Scan matching is a method of estimating self-location by searching for the position where the current point cloud data and the point cloud data of the map overlap most.

[0034] When generating an integrated bird's-eye view image, data regarding the relative position and attitude between the vehicle 2a and the other vehicle 2b is required. However, according to the second embodiment, such data can be acquired by self-position estimation processing without using the position and attitude sensor 3, thereby reducing the cost of the position and attitude sensor 3 that was required in the first embodiment.

[0035] The third embodiment will be described with reference to FIG. 6, focusing on the differences from the second embodiment. In the configuration of the third embodiment, the same parts as those in the first and second embodiments are denoted by the same reference numerals in the drawings as those in the first and second embodiments, and the explanations of the first and second embodiments are to be cited.

[0036] The three-dimensional point cloud processing unit 10c of the third embodiment shown in FIG. 6 differs from the three-dimensional point cloud processing unit 10b of the second embodiment shown in FIG. 5 in that it has a point cloud accumulation processing unit 16. The point cloud accumulation processing unit 16 is a data processing method for obtaining a denser three-dimensional point cloud than only the three-dimensional point cloud obtained at the current time by a point cloud accumulation process that accumulates three-dimensional point clouds input from the three-dimensional measurement sensor 4 from the past to the present, and more specifically, accumulates the three-dimensional point cloud by converting the three-dimensional point cloud into the coordinate system of the vehicle 2 at the current time based on the position of the vehicle 2a at the time when the three-dimensional point cloud was obtained and the position of the vehicle 2a at the current time. According to the third embodiment, an integrated bird's-eye view image based on the dense three-dimensional point cloud obtained by the point cloud accumulation processing unit 16 can be obtained, thereby improving the object recognition accuracy.

[0037] The fourth embodiment will be described with reference to FIG. 7, focusing on the differences from the third embodiment. In the configuration of the fourth embodiment, the same parts as those in the first to third embodiments are denoted by the same reference numerals in the drawings as those in the first to third embodiments, and the explanations of the first to third embodiments are to be cited.

[0038] The data processing unit 5d of the fourth embodiment shown in Fig. 7 differs from the data processing unit 5c of the third embodiment shown in Fig. 6 in that it has an object-information-attached 3D point cloud generation processing unit 17. The object-information-attached 3D point cloud generation processing unit 17 performs processing to link the 3D point cloud of the vehicle 2a acquired by the 3D measurement sensor 4 with the semantic segmentation result in the object recognition processing unit 14, i.e., an integrated bird's-eye image in which the object type is specified for each pixel, and assigns the color of the category into which the point is classified to each point of the 3D point cloud, thereby coloring the object S recognized in the 3D point cloud and obtaining a 3D point cloud including object information that allows easy distinction between cars, people, buildings, etc.

[0039] As described above, according to the first to fourth embodiments, the relative position is estimated based on the measurement data shared with the other vehicle 2b, and the measurement data of the multiple vehicles 2 is integrated, so that it is possible to perform highly accurate object recognition in real time in each vehicle 2 while the vehicle 2 is traveling. In particular, since the three-dimensional measurement data is converted into a bird's-eye view image to reduce the amount of data, it is possible to share the data with the other vehicle 2b via a narrowband communication line. Furthermore, according to the third and fourth embodiments, it is possible to reduce blind spots and improve object recognition accuracy by accumulating past three-dimensional point clouds and performing object recognition processing. Furthermore, according to the fourth embodiment, it is possible to obtain a three-dimensional point cloud with object information.

[0040] The fifth embodiment will be described with reference to FIG. In the configuration of the fifth embodiment, the same parts as those in the other embodiments are denoted by the same reference numerals in the drawings, and the explanation of those embodiments will be incorporated herein by reference.

[0041] In the first to fourth embodiments, multiple vehicles 2 are used to share measurement data between the vehicles, and each vehicle 2 generates an integrated bird's-eye image and performs object recognition processing. In contrast, in the fifth embodiment, a single vehicle 2 equipped with a three-dimensional measurement sensor 4 is used, object recognition processing is performed on the bird's-eye image acquired by the vehicle 2a, and the recognition result is linked to the three-dimensional point cloud of the vehicle 2a to generate a three-dimensional point cloud with object information. Therefore, the wireless communication device 6, the relative position estimation processing unit 12, and the data integration processing unit 13 are not provided.

[0042] The object recognition system 1e of the fifth embodiment has the same three-dimensional measurement sensor 4 as the object recognition systems 1a, 1b, 1c, and 1d of the first to fourth embodiments (FIGS. 2, 5, 6, and 7), and has the same three-dimensional point cloud processing unit 10c as the object recognition systems 1c and 1d of the third and fourth embodiments (FIGS. 6 and 7). The object recognition system 1e of the fifth embodiment also has an object recognition processing unit 14e that performs object recognition processing on the bird's-eye image of the vehicle 2a sent from the three-dimensional point cloud processing unit 10c, and an object-information-attached three-dimensional point cloud generation processing unit 17e that links the bird's-eye image of the vehicle 2a that has been subjected to the object recognition processing and sent from the object recognition processing unit 14e to the three-dimensional point cloud of the vehicle 2a to generate a three-dimensional point cloud with object information.

[0043] According to the fifth embodiment, it is possible to perform highly accurate object recognition by obtaining a 3D point cloud with object information in real time while the vehicle 2 is traveling independently. In particular, by storing past 3D point clouds and performing object recognition processing, it is possible to reduce blind spots and improve object recognition accuracy.

[0044] In the embodiment described above, the host vehicle 2a acquires the position and attitude of the other vehicle 2b through wireless communication from the other vehicle 2b, and calculates the inter-vehicle relative position and attitude between the host vehicle 2a and the other vehicle 2b. However, if conditions such as the distance to the other vehicle 2b are met, the position and attitude of the other vehicle 2b may be directly measured using a three-dimensional measuring means 4 such as a three-dimensional LiDAR mounted on the host vehicle 2a, and used to calculate the inter-vehicle relative position and attitude of the host vehicle 2a. [Explanation of symbols]

[0045] 1a, 1b, 1c, 1d, 1e...Object recognition system 2. Vehicles as a means of transportation 2a…Vehicle 2b…Other vehicles 3. Position and orientation sensor 4. 3D measurement sensor as a 3D measurement tool 5a, 5b, 5c, 5d, 5e...Data processing unit as control means 6. Wireless communication devices as a means of communication 10a, 10b, 10c...3D point cloud processing unit 11...Image processing unit 12...Relative position estimation processing unit 13...Data integration processing section 14, 14e...Object recognition processing unit 15...Self-location estimation processing unit 16...Point cloud accumulation processing unit 17, 17e…Processing unit for generating 3D point cloud with object information S…object

Claims

[Claim 1] A single means of transportation; a three-dimensional measuring means mounted on the moving means and measuring the surroundings to obtain a three-dimensional point cloud; a control means mounted on the moving means and configured to recognize surrounding objects based on an object recognition process using the three-dimensional point cloud; An object recognition system comprising: The control means A self-location estimation processing unit that estimates a location of the moving means; a point cloud accumulation processing unit that accumulates the three-dimensional point clouds acquired by measurements from the past to the present by converting the obtained three-dimensional point cloud into a coordinate system of the moving means at the current time based on the position of the moving means at the time when the three-dimensional point cloud was obtained and the position of the moving means at the current time, thereby acquiring an accumulated three-dimensional point cloud that is denser than only the three-dimensional point cloud obtained at the current time; an imaging processor for generating a bird's-eye view image from the accumulated three-dimensional point cloud; an object recognition processing unit that performs object recognition to estimate an object type for each pixel constituting the bird's-eye view image by performing the object recognition processing on the bird's-eye view image; a processing unit for generating a three-dimensional point cloud with object information, which generates a three-dimensional point cloud with object information by linking a result of performing the object recognition processing on the bird's-eye image with the three-dimensional point cloud; having An object recognition system characterized by obtaining a three-dimensional point cloud with object information in real time while the moving means is traveling independently.

Images