Method for identifying vehicles in service area based on three-dimensional laser radar data

Abstract

The invention discloses a method for identifying vehicles in a service area based on three-dimensional laser radar data, and the method comprises the following steps: collecting environment information of the service area, and constructing a colored base map line model comprising key ground features of the service area; constructing a local coordinate system and a world coordinate system; carrying out region division on the colored base map line model, and carrying out laser radar arrangement at a safe region position of the colored base map line model based on the priority of vehicle identification positions; based on the arranged laser radar, classification and real-time state identification are carried out on vehicles in a service area, and collected data are uploaded to a server side in real time. The beneficial effects of the invention are that according to the actual condition of the service area and the coverage range of the laser radars, the layout points of the laser radars can be intelligently optimized, the number of the laser radars can be reduced to the maximum extent, and the laser radar coverage of each key area of the service area is ensured to be free of dead angles while the cost is reduced.

Description

Technical Field

[0001] The present application relates to the field of intelligent transportation technology, and in particular to a method for identifying vehicles in a service area based on three-dimensional laser radar data. Background Art

[0002] Compared with service areas, other vehicle identification technology application scenarios such as parking lots and autonomous driving have significant differences in tasks and technical requirements. Vehicle identification in parking lots mainly focuses on access management and parking guidance, with the emphasis on vehicle identity verification and parking space allocation. It has lower requirements for vehicle location accuracy, but higher requirements for vehicle identity recognition accuracy. Vehicle identification for autonomous driving focuses on real-time monitoring of driving status, emphasizing high-precision trajectory recognition and immediate response to ensure autonomous decision-making and safe driving of vehicles.

[0003] However, vehicle management in service areas not only involves the identification of vehicle identities, but also requires real-time monitoring and analysis of vehicle behavior in complex environments. The driving status of vehicles in service areas is more complicated, including deceleration, parking, starting and other states, and the vehicle driving routes are not fixed, lacking the regularity like in autonomous driving scenarios. In addition, the environment in the service area is complex, with a large number of people and facilities, which puts higher anti-interference requirements on radar technology, especially when distinguishing vehicles from other objects. The technical difficulty is relatively high. Traditional vehicle monitoring technology relies on limited cameras or manual monitoring, which cannot fully cover various dynamic situations in the service area, and is inefficient in high-traffic, high-density environments. Summary of the invention

[0004] One of the objects of the present application is to provide a method for identifying vehicles in a service area based on three-dimensional lidar data, which can solve at least one of the defects in the above-mentioned background technology.

[0005] In order to achieve at least one of the above purposes, the technical solution adopted by the present application is: a method for identifying vehicles in a service area based on three-dimensional laser radar data, comprising the following steps: S100: Collect service area environmental information and construct a colored base map line model including key ground features of the service area; S200: constructing a local coordinate system and a world coordinate system; wherein the local coordinate system is used to aggregate and analyze vehicle driving behavior data, and the world coordinate system is used for cross-system data integration and information sharing; S300: Divide the colored base map line model into regions; arrange laser radars in safe area positions of the colored base map line model based on the priority of the vehicle identification position, and use the minimum number of laser radars to achieve full coverage of the service area according to the approximation principle; S400: Classify and identify the real-time status of vehicles in the service area based on the deployed laser radar, and upload the collected data to the server in real time.

[0006] Preferably, step S100 includes the following specific processes: S110: Multiple cross targets are evenly distributed in the service area, and the 3D coordinates of the cross targets are measured using GNSS equipment; S120: Use drones to take photos of the service area in a wide range and at multiple angles, and construct a colored 3D point cloud model of the service area based on the collected image data; S130: De-noising the obtained three-dimensional point cloud model, and using the de-noised three-dimensional point cloud model to construct a colored base map line model.

[0007] Preferably, in step S130, denoising the three-dimensional point cloud model includes the following specific processes: S131: judging the ground of the point cloud by using the RANSAC algorithm; S132: judging the height of each point from the ground in the point cloud data, if the height of the point from the ground is within a set first range, the point is considered to be a low-range point and is retained, otherwise proceeding to the next step; S133: Determine whether the height of the point from the ground is within the second range, if so, proceed to the next step, otherwise, determine that the point is a noise point and remove it; S134: Use the point as the center of the sphere to search within a range of a set radius. If the number of other points within the spherical search range is greater than the set threshold, the point is considered a low-level point and is retained; otherwise, it is considered a noise point and is removed.

[0008] Preferably, in step S130, the construction process of the colored base map line model is as follows: S135: Based on the denoised point cloud data, remove all points that are above a first range from the ground; S136: Project the remaining point cloud data onto the ground to generate a colored base map model of the ground; S137: Using computer vision technology, the spatial layout and lane information of different areas in the service area are extracted from the colored base map model generated in step S136 to obtain a colored base map line model of the service area.

[0009] Preferably, in step S300, the region division of the colored base map line model includes the following process: S310: Divide the service area into three areas according to the colored base map line model, including an area that needs to be identified, an area for immovable objects, and an area where radar can be set up; S320: Divide the entire service area into grids and number the divided grids in horizontal and vertical directions; S330: Obtain the numbered positions of the areas to be identified and the coordinates of the four corner points of each grid according to the divided grids combined with the colored base map line model.

[0010] Preferably, the vehicle identification priority at the entrance and exit of the service area is higher than that at other locations; then the laser radar arrangement at the entrance and exit of the service area in step S300 includes the following process: S340: Locate the midpoint coordinates (x 1 ,y 1 ), and the coordinates of the center point of the service area (x 2 ,y 2 ); S350: Calculate the placement vector v of the laser radar according to the obtained coordinates = (x 2 -x 1 ,y 2 -y 1 ); S360: Determine a candidate area for radar placement according to the direction of the placement vector v; S370: Traverse each grid in the candidate area to calculate the recognition range of the laser radar, select the local optimal grid and determine the precise layout position of the laser radar through the approximation principle.

[0011] Preferably, the laser radar arrangement process for other areas of the service area except the entrance and exit positions is as follows: S381: Determine the grid corresponding to the edge of the coverage range of the laser radar at the entrance and exit of the service area, and then obtain the grid area number that is not fully recognized in the service area; S382: traverse each empty grid area in the horizontal or vertical direction and simulate the placement of the laser radar; S383: Recording the empty grid area and edge position covered by the laser radar at each placement position; S384: Selecting a placement point that covers the largest number of empty grid areas and has the largest number of intersecting grids with the grid area occupied by the laser radar at the entrance and exit of the service area as the optimal placement grid area; S385: Determine the precise placement position of the laser radar in the optimal placement grid area by using the approximation principle.

[0012] Preferably, the process of determining the precise placement position of the laser radar based on the approximation principle is as follows: S301: traversing all layout areas where laser radars can be placed in the corresponding local optimal grid; S302: Divide the layout area into a plurality of placement points according to a set interval distance; S303: Perform a hypothetical placement of the laser radar for each placement point and determine the coverage range of the laser radar, and select the placement point with the best coverage range as the precise layout position of the laser radar.

[0013] Preferably, step S400 includes the following process: S410: Target annotation is performed on the 3D point cloud collected by the LiDAR, including three types of targets: cars, vans, and trucks. Other types are annotated as negative samples. S420: Building a vehicle recognition model on the server and performing model training; S430: Calibrate the coordinate system of the laser radar to ensure that the coordinates of the point cloud data collected by the laser radar are unified; S440: The point cloud data collected by the laser radar is uploaded to the server, and then detected in combination with the trained vehicle recognition model, and the vehicle is tracked and the behavior is recognized according to the detection results.

[0014] Preferably, the process of model training using the PointPillars model in step S420 is as follows: S421: Divide the three-dimensional point cloud data into multiple cylinders; if the number of points contained in a single cylinder exceeds a threshold, calculate the features of the cylinder; if the number of points in a single cylinder is less than the threshold, fill it with zero; S422: Aggregate the points in each column, calculate the mean feature of the column and compress it into a feature vector of fixed length; splice the feature vectors of all columns to form a two-dimensional feature map; S423: Processing the two-dimensional feature map through the backbone network, gradually capturing the regional shape, texture and spatial relationship between different regions, thereby extracting high-level spatial features in the two-dimensional feature map; S424: The extracted high-level spatial features are fused through the Feature Pyramid Networks network structure and passed to the head module, and then the head module generates the final detection results, including target category, location and confidence.

[0015] Compared with the prior art, the beneficial effects of this application are: (1) This application generates service area point clouds by using drone photogrammetry technology. Compared with traditional manual measurement or high-precision ground laser scanning, it not only greatly reduces equipment investment and labor costs, but also can quickly complete three-dimensional modeling of large areas, which is particularly suitable for infrastructure planning and management of large-scale service areas.

[0016] (2) This application can intelligently optimize the LiDAR deployment points based on the actual situation of the service area and the coverage of the LiDAR, minimize the number of LiDARs, and ensure that the LiDAR coverage of key areas in the service area is complete while reducing costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 The figure is a schematic diagram of the overall workflow of this application.

[0018] Figure 2 Schematic diagram of gridding the colored base map line model of the service area for this application.

[0019] Figure 3 This is a schematic diagram of the arrangement of laser radars at the entrances and exits of the service area in this application.

[0020] Figure 4 Schematic diagram of the coverage of the laser radar at the entrance and exit of the service area in this application.

[0021] Figure 5 This is a schematic diagram of the completed LiDAR layout for this application. DETAILED DESCRIPTION

[0022] Below, the present application is further described in conjunction with specific implementation methods. It should be noted that in the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms should not be understood as necessarily being directed to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification.

[0023] In the description of the present application, it should be noted that directional words, such as the terms "center", "lateral", "longitudinal", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise", etc., indicating directions and positional relationships are based on the directions or positional relationships shown in the accompanying drawings, and are only for the convenience of narrating the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and cannot be understood as limiting the specific scope of protection of the present application.

[0024] It should be noted that the terms "first", "second", etc. in the description and claims of the present application are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0025] In this application, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be connected, detachably connected, or integrated; it can be mechanically connected or electrically connected; it can be directly connected or indirectly connected through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0026] In the present application, unless otherwise clearly specified and limited, a first feature being "above" or "below" a second feature may include that the first and second features are in direct contact, or may include that the first and second features are not in direct contact but are in contact through another feature between them. Moreover, a first feature being "above", "above" and "above" a second feature includes that the first feature is directly above and obliquely above the second feature, or simply indicates that the first feature is higher in level than the second feature. A first feature being "below", "below" and "below" a second feature includes that the first feature is directly below and obliquely below the second feature, or simply indicates that the first feature is lower in level than the second feature.

[0027] The terms "including" and "having" and any variations thereof in the specification and claims of the present application are intended to cover non-exclusive inclusions. For example, a process, method, system, product or apparatus comprising a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products or apparatuses.

[0028] One of the preferred embodiments of the present application is as follows: Figure 1 As shown, a method for identifying vehicles in a service area based on three-dimensional laser radar data includes the following steps: S100: Collect service area environmental information and construct a colored base map line model including key ground features of the service area.

[0029] S200: constructing a local coordinate system and a world coordinate system; wherein the local coordinate system is used to aggregate and analyze vehicle driving behavior data, and the world coordinate system is used for cross-system data integration and information sharing.

[0030] S300: Divide the colored base map line model into regions; arrange laser radars in safe area positions of the colored base map line model based on the priority of vehicle identification positions, and use the minimum number of laser radars to achieve full coverage of the service area according to the approximation principle.

[0031] S400: Classify and identify the real-time status of vehicles in the service area based on the deployed laser radar, and upload the collected data to the server in real time.

[0032] It is understandable that in order to more accurately identify vehicles in the service area, it is necessary to fully collect environmental information in the service area. At the same time, in order to facilitate the subsequent layout of lidar, the map information of the service area can be presented in the form of a base map. Then, by traversing the layout of lidar on the base map, the layout efficiency of lidar can be effectively improved.

[0033] Specifically, the service area base map designed in this application covers key information such as parking spaces, lanes and no-parking areas in the service area. By combining with the subsequent two sets of coordinate systems, the base map information can be accurately calibrated, thereby providing comprehensive spatial data support for subsequent vehicle management.

[0034] In the construction of the coordinate system, the local coordinate system summarizes and analyzes vehicle driving behavior data, including traffic flow, speed, parking information, etc., which can facilitate the management and operation of the service area. The world coordinate system is mainly used to achieve the integration of lidar vehicle recognition and global positioning system, so as to provide support for cross-system data integration and information sharing.

[0035] As for the layout of LiDAR, since LiDAR is relatively expensive, in order to reduce the cost, this application designs a solution that can cover the entire service area and reduce the number of LiDAR deployments. This solution can achieve full LiDAR coverage of the service area to the maximum extent while reducing the number of LiDAR devices, thereby effectively reducing equipment costs and improving management efficiency.

[0036] For vehicle identification in the service area, this application can identify vehicles and analyze their behavior based on the point cloud data of the LiDAR. That is, through advanced algorithms, vehicles in the service area are classified, including cars, box trucks, and large trucks, and real-time tracking and behavior judgment of each vehicle are achieved. Through automatic detection of vehicle speeding, illegal parking, and reverse driving, the system can upload behavior data to the server in real time for service area managers to respond and make decisions quickly, thereby improving the safety and management efficiency of the service area.

[0037] For ease of understanding, each step will be described in detail below.

[0038] In this embodiment, step S100 includes the following specific processes: S110: Multiple cross targets are evenly distributed in the service area, and the three-dimensional coordinates of the cross targets are measured using GNSS equipment.

[0039] Specifically, the number of cross targets can be set according to actual needs, and generally 5-10 cross targets can be set. The cross target can divide the marking area into four sector areas, and the colors of the two adjacent sector parts are set differently, for example, black and white respectively. These cross targets are facing the sky so that the drone can accurately and clearly capture the positions of these cross targets during the measurement process. The selection of cross targets should ensure that the entire service area is covered to ensure calibration accuracy. After completing the arrangement of the cross targets, GNSS equipment, such as Beidou Positioning System, GPS and Galileo, can be used to accurately measure the three-dimensional coordinates of a single cross target.

[0040] S120: Use drones to take photos of the service area over a wide range and from multiple angles, and construct a colored 3D point cloud model of the service area based on the collected image data.

[0041] Specifically, the image data collection of the drone needs to ensure that there is enough image data to facilitate subsequent 3D reconstruction. For the construction of colored 3D point cloud models, the structure from motion algorithm (SfM) is preferably used. The position and orientation of the camera carried by the drone and the 3D coordinates of the objects in the scene are inferred by using the structure from motion algorithm (SfM). The SfM algorithm not only needs to identify the feature points in the image, but also can use the relative position relationship between the cameras to reconstruct the scene in 3D. The specific steps of the algorithm are as follows: Extract key feature points from each image and match the feature points between different images. Based on the known feature point matching results, the relative position and posture between the cameras are estimated through the basic matrix or essential matrix. This process is essentially to solve a linear system to determine the internal and external parameters of each camera. Using the previously obtained camera position and matched two-dimensional feature points, the algorithm solves the spatial points through collinearity equations. Collinearity equations are key mathematical tools in photogrammetry, describing the geometric relationship between points in image space, camera coordinate systems, and physical coordinate systems. Collinearity equations allow the position of an object in three-dimensional space to be inferred from two-dimensional image coordinates by establishing a relationship between image coordinates and object coordinate systems; the collinearity equation is in the following form: .

[0042] .

[0043] Among them, (x, y) represents the image coordinates, (x 0 ,y 0) represents the principal point coordinates, c represents the focal length of the camera, (X, Y, Z) represents the object coordinates in three-dimensional space, (X 0 , Y 0 , Z 0 ) represents the coordinates of the camera projection center in the object coordinate system, R ij Represents the rotation transformation from the camera coordinate system to the object coordinate system.

[0044] After the initial 3D reconstruction is completed, it will be further optimized through bundle adjustment. Bundle adjustment is to optimize the parameters of all cameras and the coordinates of 3D points by minimizing the reprojection error of image feature points at each viewing angle. The least squares method is used to adjust the positions of the cameras and 3D points until the error is minimized.

[0045] The 3D point cloud model obtained after the above 3D reconstruction and optimization is still a sparse model. Based on the sparse reconstruction, the Multi-View Stereo (MVS) technology is used to generate a dense point cloud. The dense point cloud represents the geometric structure of the scene more finely by adding more 3D points. Through the above steps, a dense colored 3D point cloud model of the service area can be generated. The point cloud can be converted to the world coordinate system through the four-parameter conversion formula by comparing the cross target coordinates in the point cloud with the pre-measured cross target coordinates; the specific conversion process is well known to those skilled in the art, so it will not be elaborated here.

[0046] S130: De-noising the obtained three-dimensional point cloud model, and using the de-noised three-dimensional point cloud model to construct a colored base map line model.

[0047] It is understandable that there may be a large number of noise points in the colored 3D point cloud obtained through the above steps, so it is necessary to perform denoising operation on the point cloud. Specifically, the denoising of the 3D point cloud model includes the following specific processes: S131: Determine the ground of the point cloud using the RANSAC algorithm.

[0048] It should be noted that in order to facilitate the subsequent steps, the point cloud space can be divided into multiple grid areas in the horizontal xy plane. The size of a single grid area can be set according to actual needs. For example, the size of a single grid is 3m×3m. Each grid area contains multiple points, and the z value of each grid is the height information of all points in the grid area.

[0049] S132: Determine the height from each point in the point cloud data to the ground. If the height from the point to the ground is within a set first range, the point is considered to be a low-range point and is retained. Otherwise, proceed to the next step.

[0050] It is understandable that the specific value of the first range can be set according to actual needs; generally speaking, considering the slope change of the road in the service area, it is preferred to set the first range to ±0.5m.

[0051] S133: Determine whether the height of the point from the ground is within the second range, if so, proceed to the next step, otherwise, determine that the point is a noise point and remove it.

[0052] It can be understood that the specific value of the second range can be set according to actual needs, taking into account that the height of general vehicles is below 3m, that is, objects above 3m will not interfere with the normal driving of the vehicle; therefore, the value of the second range is preferably 3m, that is, points with a height exceeding 3m are regarded as noise points and deleted.

[0053] S134: For a point whose height exceeds the first range but is within the second range, a range search with a set radius distance is performed with the point as the center of the sphere. If the number of other points within the spherical search range is greater than the set threshold, the point is considered to be a low-range point and is retained; otherwise, it is considered to be a noise point and is removed.

[0054] It is understandable that points whose height exceeds the first range but are within the second range may belong to points on buildings or street lamps. These points are theoretically part of buildings or street lamps and cannot be simply judged as noise points. That is, through step S134, the high points on buildings and street lamps are retained, and the noise points in the elevation are eliminated; the noise points in the high part of the parking area can be effectively removed, thereby obtaining a more accurate point cloud model, ensuring that the open area in the elevation direction is not affected by noise points. The specific value of the radius of the spherical search range can be selected according to actual needs, for example, the radius can be taken as 0.5m.

[0055] In this embodiment, after the point cloud data is denoised, the construction of the colored base map line model includes the following process: S135: Based on the denoised point cloud data, remove all points that are above a first range from the ground.

[0056] It is understandable that the colored base map line model mainly records the ground information of the service area, and the denoised point cloud data also contains some point clouds of non-ground objects. For example, a semi-enclosed awning, etc., does not affect traffic below the awning. If all these points are projected to the ground in the next step, they may be regarded as ground structure points during data extraction, resulting in inaccurate subsequent base map establishment.

[0057] S136: Project the remaining point cloud data onto the ground to generate a colored base map model of the ground.

[0058] It should be noted that the generated colored basemap line model represents the ground information of the service area, including key ground features such as parking lines, zebra crossings and vehicle markings.

[0059] S137: Using computer vision technology, the spatial layout and lane information of different areas in the service area are extracted from the colored base map model generated in step S136 to obtain a colored base map line model of the service area.

[0060] It is understandable that there may be multiple computer vision technologies used in step S137, such as edge detection methods and threshold segmentation methods. The specific working principles are well known to those skilled in the art, so they will not be elaborated on in detail here. Computer vision technology can be used to extract line features such as lane lines and parking lines from the colored base map line model. These features represent the spatial layout and lane information of different areas in the service area. A JSON file containing the position, size and geometry of each part is generated through the extracted lane lines and other markings. The file contains spatial information of elements such as lane lines, parking spaces, and no-parking areas, and provides data support for future lane line extraction and vehicle position recognition.

[0061] Based on the above information, an accurate colored base map line model of the service area is generated. This model provides accurate ground layout information for subsequent vehicle identification and location analysis by integrating the three-dimensional coordinate system with spatial data such as lane lines. Using this base map line model, combined with future vehicle location recognition technology, intelligent management and monitoring of vehicles in the service area can be achieved. This technology can provide accurate basic data support for lane line management, parking space utilization, and traffic flow monitoring.

[0062] It should be known that this application uses drone photogrammetry technology to generate service area point clouds, so that the base map of the service area can be constructed quickly and at low cost. Compared with traditional manual measurement and high-precision ground laser scanning methods, this application not only greatly reduces equipment investment and labor costs, but also can quickly complete three-dimensional modeling of large areas, which is especially suitable for infrastructure planning and management of large-scale service areas. Through the technical solution of this application, it is possible to provide an accurate base map for the service area and ensure that various subsequent technical applications based on three-dimensional models have high-quality data support.

[0063] In this embodiment, in order to achieve accurate vehicle monitoring and information management, two sets of coordinate systems are designed in step S200, namely the local coordinate system of the service area and the world coordinate system; the specific construction process of the local coordinate system and the world coordinate system is a well-known technology for technical personnel in this field; for the sake of ease of understanding, the two coordinate systems will be briefly described below.

[0064] 1. World coordinate system.

[0065] The world coordinate system is mainly used to collect and integrate information about various parts of the service area and identified vehicles on a global scale. Through the world coordinate system, various types of information in the service area can be integrated with highway traffic flow data. For example, the world coordinate system can monitor the position and trajectory of vehicles in the service area in real time and exchange data with external systems (such as traffic control centers).

[0066] The construction of the world coordinate system relies on the global navigation satellite system (GNSS) and cross targets deployed in the service area. 5-10 cross targets are evenly deployed in the service area. These cross targets should have clear global positioning information, face the sky and be accurately captured by drones or other measuring equipment. By calibrating the positions of these cross targets in the 3D model and combining the coordinate information provided by GNSS, the base map and 3D point cloud model of the service area can be converted to the world coordinate system to achieve accurate global positioning and information fusion.

[0067] 2. Local coordinate system.

[0068] The local coordinate system is mainly used for vehicle monitoring, anomaly detection and behavior analysis in the service area. Compared with the world coordinate system, the local coordinate system is based on the specific scene of the service area and can more intuitively and simply represent the position and status of objects in the service area. Through the local coordinate system, vehicle entry and exit recognition, parking status judgment and other events can be efficiently monitored in real time. The use of the local coordinate system enables the system to achieve rapid response in a local area, improving processing speed and accuracy.

[0069] The establishment of the local coordinate system is based on the conversion principle of the world coordinate system. After obtaining the coordinate data of the world coordinate system, the local coordinate system can be obtained by decentralizing the coordinates within the service area. Specifically, the local coordinate system can be constructed by converting the world coordinate system data of all key objects in the service area (such as parking spaces, lane lines, traffic signs, etc.) into relative coordinates. This process can be simplified by setting reference points or fixed reference points, and ensure that subsequent data processing and management in the local coordinate system can proceed smoothly.

[0070] In the subsequent vehicle monitoring and behavior recognition process, the system will use the local coordinate system to process and judge the data. In order to achieve seamless conversion between different scenes, an interface will be provided between the local coordinate system and the world coordinate system to support real-time conversion and data update. Specifically, when it is necessary to associate the recognition results in the local coordinate system with global traffic data, the system will convert the coordinates of the local coordinate system into the world coordinate system through a preset conversion algorithm to facilitate efficient data sharing and information integration. Through the combination of the two sets of coordinate systems, the present application can provide local management and control while also being convenient for linkage with a wider range of transportation systems, Internet of Things platforms, and highway management systems. By using the world coordinate system, the data in the service area can be seamlessly connected with the surrounding traffic flow.

[0071] In this embodiment, when performing step S300, based on the principle of using the least number of laser radars to cover the entire service area, it is necessary to first confirm the installation height of each laser radar when deploying the laser radar. The installation height of the laser radar needs to meet the following conditions: meet the recognition of all types of vehicles, ensure the effectiveness of recognition when the vehicle passes obliquely below the laser radar, that is, for vehicles close to the laser radar, the laser radar should still be able to correctly identify its existence.

[0072] It should be known that the height of the vehicle shall generally not exceed 4.2m, and the length of the vehicle shall generally not exceed 18m. According to the working principle of LiDAR, the scanning angle and effective scanning distance of LiDAR are crucial to the recognition effect. In order to avoid recognition failure due to too small longitudinal angle, the scanning angle of LiDAR is usually required to be greater than 30 degrees. Therefore, according to the principles of trigonometry, it can be deduced that the minimum installation height of LiDAR is: 9×sin30°+4.2=8.7m. In order to leave redundant space, the installation height of each LiDAR in this embodiment is preferably 9 meters.

[0073] For ease of understanding, the local deployment process of the LiDAR can be described in detail below based on the simplified structure of the service area. Figure 2 As shown in the figure, the black shaded area represents the location suitable for setting up LiDAR. The shaded area not only includes the intersection between the lane and the parking area, but also specially marks the area above the parking area. It should be noted that these shadows are only schematic markings to show which areas are suitable for laying radars. When actually laying out, the specific location of the LiDAR control points should be considered. These areas are usually located at the intersection between parking spaces and the intersection of lanes. LiDAR can effectively monitor vehicles at these key locations. Figure 2It is intuitive to see which parts are suitable for placing LiDAR. The shaded area shown in the figure covers the key areas of the service area, including lanes, parking spaces, and entrances and exits. The deployment of LiDAR should be concentrated in these intersection areas to ensure that no important vehicle flow information and status monitoring are missed.

[0074] Specifically, in step S300, the region division of the colored base map line model includes the following process: S310: Divide the service area into three areas according to the colored base map line model, including the area that needs to be identified, the area for immovable objects, and the area where radar can be set up.

[0075] It is understandable that the areas that need to be identified include lanes, parking spaces, gas station areas, and charging areas. This area needs to ensure that the lidar can be fully covered, and lidar equipment must not be installed. The immovable object area includes areas such as server buildings, columns, and signboards, which are not suitable for placing radars. Areas where radars can be set up include green belts, roadside areas, parking areas and the middle part of the road, and intersections of multiple parking spaces.

[0076] S320: Divide the entire service area into grids and number the divided grids in horizontal and vertical directions.

[0077] It is understandable that in order to facilitate the subsequent calculation of the direction of the LiDAR deployment, the service area can be divided into spatial grids first. The size of each grid can be set according to actual needs, for example, the size of each grid is 15m×15m. In order to ensure effective vehicle identification and minimize the number of radar devices, the grids in the service area need to be numbered. For easy distinction, different types of numbers can be used in the horizontal and vertical directions, such as letters for horizontal numbering and Arabic numerals for vertical numbering. Figure 2 As shown, assuming that the size of the service area is 225m×120m, the service area can be divided into 15×8 grid areas, numbered from A to O horizontally and from 1 to 8 vertically.

[0078] S330: Obtain the numbered positions of the areas to be identified and the coordinates of the four corner points of each grid according to the divided grids combined with the colored base map line model.

[0079] It is understandable that after completing the grid division, it is necessary to further determine which grid area needs to be identified based on the base map and the grid position. Figure 2 As shown, the grids in the first row do not need to be identified, and the corresponding grid areas such as D2 to M2 do not need to be identified. While determining whether the grid area needs to be identified, the coordinates of the four corner points of each grid need to be obtained. These coordinates can be used to determine whether the point covered by the laser radar is located in the grid.

[0080] In this embodiment, the vehicle identification priority at the entrance and exit of the service area is higher than that at other locations, that is, the laser radar needs to first ensure the traffic information at the entrance and exit of the service area, as well as dynamic information such as the driving mode of the vehicle after entering from the entrance and the entry direction before exiting the exit, so as to ensure accurate identification of the vehicle when driving in the service area.

[0081] Specifically, Figure 2 and Figure 3 As shown, in step S300, the laser radar arrangement at the entrance and exit of the service area includes the following process: S340: Locate the midpoint coordinates (x 1 ,y 1 ), and the coordinates of the center point of the service area (x 2 ,y 2 ).

[0082] S350: Calculate the placement vector v of the laser radar according to the obtained coordinates = (x 2 -x 1 ,y 2 -y 1 ).

[0083] S360: Determine a candidate area for radar placement according to the direction of the placement vector v.

[0084] S370: Traverse each grid in the candidate area to calculate the recognition range of the laser radar, select the local optimal grid and determine the precise layout position of the laser radar through the approximation principle.

[0085] For ease of understanding, the following will combine the layout of the laser radar at the entrance of the service area Figure 2 Provide a detailed description.

[0086] like Figure 2 As shown in the figure, the lane information of the service area is extracted through image processing technology to determine the widening, bifurcation or obvious turning point of the lane. Figure 2 It can be seen that the diversion area of ​​vehicles at the entrance of the service area is grid B3, so it is necessary to ensure that the laser radar at the entrance can cover this area and provide a certain degree of redundant coverage for the surrounding areas. The center point of the service area is located at the intersection of grids H4 and H5, with coordinates (x 2 ,y 2 ); the coordinates of the center point of grid B3 are (x 1 ,y 1 ). The specific values ​​of the two coordinates can be obtained according to the size of the service area, and then the placement vector of the lidar layout can be calculated.

[0087] According to the direction of the calculated placement vector v, determine the candidate area for the placement of the LiDAR, and ensure that the effective coverage range of the LiDAR can completely cover the B3 grid and the maximum coverage area of ​​the A and B columns of grids. In order to determine the best placement point for the LiDAR, traverse each candidate area, and calculate whether it can cover the B3 area and the A and B columns of grids based on the recognition range of the LiDAR. Finally, select a local optimal placement grid, and further use the approximation method to accurately determine the placement position within the grid. Assume that the effective recognition range of the LiDAR is 70 meters and the scanning angle of the LiDAR meets the requirement of 30 degrees. Assume that the LiDAR placement point is (x p ,y p ), the effective coverage area of ​​the laser radar is a circular area with this point as the center and a radius of 70 meters, which can be specifically expressed by the following formula.

[0088] .

[0089] It is understandable that using the center of the grid as the placement point alone will affect the placement accuracy of the LiDAR, and the center point of some areas may be on the road or parking space, which cannot meet the placement requirements. Therefore, it is necessary to find the best placement area in the placement area of ​​this grid; that is, to select the placement point in a single grid based on the approximation principle. The process of determining the precise placement position of the LiDAR based on the approximation principle is as follows: S301: Traverse all layout areas where laser radars can be placed in the corresponding local optimal grid.

[0090] S302: Divide the layout area into a plurality of placement points according to a set interval distance.

[0091] S303: Perform a hypothetical placement of the laser radar for each placement point and determine the coverage range of the laser radar, and select the placement point with the best coverage range as the precise layout position of the laser radar.

[0092] It should be known that the setting of the interval distance in step S302 can be selected according to the actual needs of those skilled in the art, for example, the interval distance can be set to 1 m.

[0093] In this embodiment, after completing the deployment of the laser radar at the entrance and exit of the service area, it is necessary to select the deployment points at other locations in the middle of the service area. Figures 3 to 5 As shown in the figure, the laser radar layout process for other areas of the service area except the entrance and exit positions is as follows: S381: Determine the grid corresponding to the edge of the coverage range of the laser radar at the entrance and exit of the service area, and then obtain the grid area number that is not fully recognized in the service area.

[0094] S382: Traverse each empty grid area in the horizontal or vertical direction and simulate the placement of the lidar.

[0095] S383: Record the empty grid area and edge position covered by the laser radar at each placement position.

[0096] S384: Select the placement point that covers the most empty grid areas and has the largest number of intersection grids with the grid area occupied by the laser radar at the entrance and exit of the service area as the optimal placement grid area.

[0097] S385: Determine the precise placement position of the laser radar in the optimal placement grid area by using the approximation principle.

[0098] It can be understood that the layout of the laser radar in the middle of the service area is mainly based on the coverage range of the laser radar at the entrance and exit positions, which can effectively reduce the number of laser radars deployed. For ease of understanding, the layout process of the laser radar in the middle area of ​​the service area will be described in detail with reference to the accompanying drawings.

[0099] like Figure 3 As shown in the figure, the two circles represent the coverage of the two laser radars at the entrance and exit of the service area. The fully covered area of ​​the two entrance and exit laser radar points is extracted, and the areas that are not fully recognized are demarcated as empty areas. Figure 4 As shown, the gray shaded part is the grid area range that can be fully covered by the laser radar at the entrance and exit. It can be seen from the figure that the areas that need to be identified in the current service area but have not yet been fully identified include E2, E7-E8, G3-G8, H3-H8, I3 and I6-I8.

[0100] After obtaining these areas, we first need to divide the span of the area; because each grid is an interval of 15m×15m, we can first determine the ratio and range between the horizontal grid span and the vertical grid span. Through this ratio and range, we can further estimate how many laser radars need to be deployed. Taking the current service area as an example, the vertical span of the unrecognized area is six grids, which is 90 meters, or 45 meters on one side. The horizontal span is four grids, which is 60 meters, or 30 meters on one side. Therefore, in theory, only one more laser radar needs to be deployed.

[0101] Next, it is necessary to determine the location of the additional laser radar. Select the middle grid that is not fully occupied in the horizontal or vertical direction, and then determine whether the grid can be used to deploy the laser radar. The specific judgment method can refer to the layout judgment method of the laser radar at the entrance and exit. If the laser radar can be deployed in the grid area, the laser radar can be virtually deployed and its coverage range can be recorded. After the virtual layout of the laser radar has been completed in all grid areas where the laser radar can be deployed, all recorded coverage ranges are compared, and the grid area that can occupy the most empty grid areas and has the largest cross-overlapping range with the laser radar at the entrance and exit is selected as the best grid area for placement. After carefully analyzing the uncovered areas and selecting the laser radar placement area, further precise adjustments are required. The selection of laser radar placement points is based on the same principles as the previous layout of the entrance and exit laser radars. In order to ensure that the coverage of each laser radar is maximized, after selecting the appropriate placement point, the effective coverage of the laser radar is verified again to ensure that it can cover areas that are not fully identified. Such as Figure 5 As shown, based on the current placement points of the three lidars, full coverage of all areas that need to be identified in the service area can be achieved.

[0102] In this embodiment, step S400 includes the following process: S410: Target annotation is performed on the three-dimensional point cloud collected by the lidar, including three types of targets: cars, box trucks, and large trucks. Other types are marked as negative samples.

[0103] It is understandable that in order to facilitate data management, the LabelCloud tool can be used to annotate the 3D point cloud and generate labels that conform to the KITTI dataset format. It should be noted that when annotating the vehicle, the direction of the vehicle head must be correctly annotated, otherwise it will affect the subsequent target recognition accuracy.

[0104] S420: Construct a vehicle recognition model on the server side and perform model training.

[0105] It is understandable that there are many specific methods for vehicle recognition model training. This embodiment uses the PointPillars model to perform model training, which specifically includes the following process: S421: Divide the three-dimensional point cloud data into multiple cylinders; if the number of points contained in a single cylinder exceeds a threshold, calculate the features of the cylinder; if the number of points in a single cylinder is less than the threshold, fill it with zero.

[0106] It can be understood that each column represents the point cloud information in a fixed area. The PointPillars model divides the space into three-dimensional grids of uniform size, namely columns; the grid size can be adjusted according to the actual scene. After completing the feature calculation of the column, each column is represented as a point cloud feature vector, including coordinate information, reflection intensity, etc.

[0107] S422: Aggregate the points in each column, calculate the mean feature of the column and compress it into a feature vector of fixed length; and concatenate the feature vectors of all the columns to form a two-dimensional feature map.

[0108] It should be noted that the size of the feature map is the number of bins multiplied by the feature dimension of each bin.

[0109] S423: Processing the two-dimensional feature map through the backbone network, gradually capturing the regional shape, texture and spatial relationship between different regions, so as to extract high-level spatial features in the two-dimensional feature map.

[0110] It can be understood that the 2D convolutional neural network is used as the backbone network, and then the feature map is processed through multiple convolutional layers to obtain the corresponding high-level spatial features. The convolutional layer usually combines BatchNorm and ReLU activation functions to enhance the nonlinear expression ability of the network.

[0111] S424: The extracted high-level spatial features are fused through the Feature Pyramid Networks network structure and passed to the head module, and then the head module generates the final detection results, including target category, location and confidence.

[0112] It should be known that the detection result generation of the head module includes the following process: first, the classification task is performed, that is, predicting whether each column contains an object; then the regression task is performed, that is, predicting the position and size of the object. By comparing with the labeled data, back propagating and adjusting the network weights, the vehicle recognition accuracy can be effectively improved after multiple rounds of training.

[0113] S430: Calibrate the coordinate system of the laser radar to ensure that the coordinates of the point cloud data collected by the laser radar are unified.

[0114] It is understandable that the laser radar is calibrated through the above two coordinate systems to ensure that all point cloud data collected by the laser radar can be unified into the calibrated coordinate system. After that, the point cloud data collected in real time is transmitted to the server side, and combined with the trained model for detection, the vehicle can be identified.

[0115] S440: The point cloud data collected by the laser radar is uploaded to the server, and then detected in combination with the trained vehicle recognition model, and the vehicle is tracked and the behavior is recognized according to the detection results.

[0116] It is understandable that during the real-time monitoring stage of the LiDAR, the vehicle recognition results include information such as the 3D bounding box, vehicle center coordinates, and size. For each identified vehicle, the system assigns it a unique label and passes this information into the DeepSORT algorithm. DeepSORT combines the vehicle appearance features extracted by the convolutional neural network (CNN), predicts the object's motion trajectory through the SORT framework and Kalman filter, and uses the IoU (Intersection over Union) value between objects for target matching, thereby updating the object's state (such as position and speed).

[0117] Since the entire service area has achieved full coverage of the LiDAR layout, and the vehicle is identified in a calibrated coordinate system, the recognition results between two consecutive LiDARs can be matched by the vehicle position, thereby realizing cross-LiDAR tracking of the vehicle. Finally, the prediction results will be uploaded to the management center of the service area in real time for violation detection and service area management. In this way, the management and operation efficiency of the service area can be effectively improved.

[0118] It should be known that in step S400, by combining the PointPillars algorithm and the DeepSORT algorithm, accurate real-time identification and tracking of vehicles in the service area can be achieved. The PointPillars algorithm can accurately identify different types of vehicles (such as cars, trucks, etc.) through efficient feature extraction of point cloud data, and supports continuous tracking of vehicles in complex environments. The DeepSORT algorithm ensures high stability and high precision of vehicle tracking through Kalman filter and appearance feature extraction. On this basis, the identification data and location information of all vehicles will be uploaded to the service area management center in real time, realizing the intelligent and automated management of the service area, significantly improving the efficiency and safety of vehicle management, and providing intelligent operation and maintenance and data analysis support for the service area.

[0119] The above describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited by the above embodiments, and the above embodiments and the specification only describe the principles of the present application. The present application may have various changes and improvements without departing from the spirit and scope of the present application, and these changes and improvements fall within the scope of the present application for which protection is sought. The scope of protection claimed by the present application is defined by the attached claims and their equivalents.

Claims

1. A method for identifying vehicles in a service area based on three-dimensional laser radar data, characterized in that: The steps include: S100: Collect service area environmental information and construct a colored base map line model including key ground features of the service area; S200: constructing a local coordinate system and a world coordinate system; wherein the local coordinate system is used to aggregate and analyze vehicle driving behavior data, and the world coordinate system is used for cross-system data integration and information sharing; S300: Divide the colored base map line model into regions; arrange laser radars in safe area positions of the colored base map line model based on the priority of the vehicle identification position, and use the minimum number of laser radars while achieving full coverage of the service area according to the approximation principle; S400: Classify and identify the real-time status of vehicles in the service area based on the deployed laser radar, and upload the collected data to the server in real time.

2. The method for identifying vehicles in a service area based on three-dimensional laser radar data according to claim 1, characterized in that: Step S100 includes the following specific processes: S110: Multiple cross targets are evenly distributed in the service area, and the 3D coordinates of the cross targets are measured using GNSS equipment; S120: Use drones to take photos of the service area in a wide range and at multiple angles, and construct a colored 3D point cloud model of the service area based on the collected image data; S130: De-noising the obtained three-dimensional point cloud model, and using the de-noised three-dimensional point cloud model to construct a colored base map line model.

3. The method for identifying vehicles in a service area based on three-dimensional laser radar data as claimed in claim 2, characterized in that: In step S130, denoising the three-dimensional point cloud model includes the following specific processes: S131: judging the ground of the point cloud by using the RANSAC algorithm; S132: judging the height of each point from the ground in the point cloud data, if the height of the point from the ground is within a set first range, the point is considered to be a low-range point and is retained, otherwise proceeding to the next step; S133: Determine whether the height of the point from the ground is within the second range, if so, proceed to the next step, otherwise, determine that the point is a noise point and remove it; S134: Use the point as the center of the sphere to search within a range of a set radius. If the number of other points within the spherical search range is greater than the set threshold, the point is considered a low-level point and is retained; otherwise, it is considered a noise point and is removed.

4. The method for identifying vehicles in a service area based on three-dimensional laser radar data as claimed in claim 3, characterized in that: In step S130, the construction process of the colored base map line model is as follows: S135: Based on the denoised point cloud data, remove all points that are above a first range from the ground; S136: Project the remaining point cloud data onto the ground to generate a colored base map model of the ground; S137: Using computer vision technology, the spatial layout and lane information of different areas in the service area are extracted from the colored base map model generated in step S136 to obtain a colored base map line model of the service area.

5. The method for identifying vehicles in a service area based on three-dimensional laser radar data according to any one of claims 1 to 4, characterized in that: In step S300, the region division of the colored base map line model includes the following process: S310: Divide the service area into three areas according to the colored base map line model, including an area that needs to be identified, an area for immovable objects, and an area where radar can be set up; S320: Divide the entire service area into grids and number the divided grids in horizontal and vertical directions; S330: Obtain the numbered positions of the areas to be identified and the coordinates of the four corner points of each grid according to the divided grids combined with the colored base map line model.

6. The method for identifying vehicles in a service area based on three-dimensional laser radar data as claimed in claim 5, characterized in that: The vehicle identification priority at the entrance and exit of the service area is higher than that at other locations; then the laser radar arrangement at the entrance and exit of the service area in step S300 includes the following process: S340: Locate the midpoint coordinates (x1, y1) of the entrance and exit of the service area in the divided grid, and the center point coordinates (x2, y2) of the service area; S350: Calculate the placement vector v=(x2-x1, y2-y1) of the laser radar according to the obtained coordinates; S360: Determine a candidate area for radar placement according to the direction of the placement vector v; S370: Traverse each grid in the candidate area to calculate the recognition range of the laser radar, select the local optimal grid and determine the precise layout position of the laser radar through the approximation principle.

7. The method for identifying vehicles in a service area based on three-dimensional laser radar data as claimed in claim 6, characterized in that: The laser radar layout process for other areas of the service area except the entrance and exit locations is as follows: S381: Determine the grid corresponding to the edge of the coverage range of the laser radar at the entrance and exit of the service area, and then obtain the grid area number that is not fully recognized in the service area; S382: traverse each empty grid area in the horizontal or vertical direction and simulate the placement of the laser radar; S383: Recording the empty grid area and edge position covered by the laser radar at each placement position; S384: Selecting a placement point that covers the largest number of empty grid areas and has the largest number of intersecting grids with the grid area occupied by the laser radar at the entrance and exit of the service area as the optimal placement grid area; S385: Determine the precise placement position of the laser radar in the optimal placement grid area by using the approximation principle.

8. The method for identifying vehicles in a service area based on three-dimensional laser radar data as claimed in claim 7, characterized in that: The process of determining the precise placement of the LiDAR based on the approximation principle is as follows: S301: traversing all layout areas where laser radars can be placed in the corresponding local optimal grid; S302: Divide the layout area into a plurality of placement points according to a set interval distance; S303: Perform a hypothetical placement of the laser radar for each placement point and determine the coverage range of the laser radar, and select the placement point with the best coverage range as the precise layout position of the laser radar.

9. The method for identifying vehicles in a service area based on three-dimensional laser radar data as claimed in claim 1, characterized in that: Step S400 includes the following process: S410: Target annotation is performed on the 3D point cloud collected by the LiDAR, including three types of targets: cars, vans, and trucks. Other types are annotated as negative samples. S420: Building a vehicle recognition model on the server and performing model training; S430: Calibrate the coordinate system of the laser radar to ensure that the coordinates of the point cloud data collected by the laser radar are unified; S440: The point cloud data collected by the laser radar is uploaded to the server, and then detected in combination with the trained vehicle recognition model, and the vehicle is tracked and the behavior is recognized according to the detection results.

10. The method for identifying vehicles in a service area based on three-dimensional laser radar data as claimed in claim 9, characterized in that: The process of model training using the PointPillars model in step S420 is as follows: S421: Divide the three-dimensional point cloud data into multiple cylinders; if the number of points contained in a single cylinder exceeds a threshold, calculate the features of the cylinder; if the number of points in a single cylinder is less than the threshold, fill it with zero; S422: Aggregate the points in each column, calculate the mean feature of the column and compress it into a feature vector of fixed length; splice the feature vectors of all columns to form a two-dimensional feature map; S423: Processing the two-dimensional feature map through the backbone network, gradually capturing the regional shape, texture and spatial relationship between different regions, so as to extract high-level spatial features in the two-dimensional feature map; S424: The extracted high-level spatial features are fused through the Feature Pyramid Networks network structure and passed to the head module, and then the head module generates the final detection results, including target category, location and confidence.

Images