CN-121982455-A - Point cloud recognition training data enhancement method for semi-closed scene

CN121982455ACN 121982455 ACN121982455 ACN 121982455ACN-121982455-A

Abstract

The application discloses a point cloud identification training data enhancement method for a semi-closed scene, which comprises the steps of extracting vehicle point cloud samples from marked original point cloud data and constructing a point cloud sample library, constructing a point cloud base map, dividing functional areas, determining the main driving direction of each driving area, determining the head direction distribution in each driving area according to the spatial position relation between the driving area and an adjacent parking area, determining the number of vehicle samples to be generated, randomly selecting a corresponding number of vehicle point cloud samples from the point cloud sample library, placing the selected vehicle point cloud samples at randomly sampled arrangement positions of the driving areas, and performing rigid body transformation corresponding to the head direction distribution on the placed vehicle point cloud samples. The application has the beneficial effects that the running area and the parking area are explicitly spatially divided, and the running area of the low-frequency vehicle is pointedly supplemented with the sample, so that the problem of the lack of training data on the surface of the spatial coverage layer is effectively solved.

Inventors

YANG YUCHEN
YING GUOGANG
WEI ZHENG
ZHENG RONGJUN

Assignees

宁波朗达科技有限公司

Dates

Publication Date: 20260505
Application Date: 20260407

Claims (10)

1. The point cloud recognition training data enhancement method for the semi-closed scene is characterized by comprising the following steps of: s100, extracting a three-dimensional bounding box of each vehicle from original point cloud data marked on a semi-closed scene according to marking information, and taking point clouds in the box as vehicle point cloud samples to obtain a point cloud sample library corresponding to all vehicles; s200, removing all extracted vehicle point cloud samples from original point cloud data to construct a point cloud base map, converting the obtained point cloud base map into a main axis coordinate system, projecting the main axis coordinate system to the ground, and dividing functional areas including a driving area and a parking area; S300, determining the main driving direction of each driving area according to the geometric form or lane line information of the driving area, and determining the head orientation distribution in each driving area according to the spatial position relation between the driving area and the adjacent parking area; s400, determining the number of vehicle samples to be generated according to a set proportion based on the total number of vehicles marked in the original point cloud data, and randomly selecting a corresponding number of vehicle point cloud samples from a point cloud sample library according to the obtained number of vehicle samples; S500, randomly sampling layout positions in a driving area, placing selected vehicle point cloud samples at the layout positions of the driving area, and carrying out rigid body transformation corresponding to head orientation distribution on the placed vehicle point cloud samples.
2. The method for enhancing point cloud recognition training data for semi-closed scenes according to claim 1, wherein in step S100, eight three-dimensional corner coordinates of all vehicles are extracted from original point cloud data according to the position, size and orientation of the center point of the vehicles in the global coordinate system to construct a three-dimensional bounding box; extracting all point cloud data in the three-dimensional bounding box from the original point cloud data in a three-dimensional mask mode, and performing decentralization to form a required vehicle point cloud sample; and independently storing all the extracted vehicle point cloud samples according to the category and the number to obtain a required point cloud sample library.
3. The method for enhancing training data for point cloud recognition of a semi-closed scene as claimed in claim 1, wherein in step S200, a point cloud satisfying a height threshold constraint is selected from the point cloud base map as a candidate ground point set; performing ground plane fitting through a random sampling consistency algorithm according to the selected candidate ground point set; constructing a main axis coordinate system parallel to the ground by taking the normal vector of the ground plane as a z axis and taking the main direction in the ground plane as an x axis and a y axis; Converting the point cloud base map into a main axis coordinate system, and then carrying out two-dimensional projection based on the ground to obtain a plane point cloud base map; According to the field region structural characteristics, the planar point cloud base map is divided into a plurality of different functional areas in the form of a two-dimensional envelope.
4. The method for enhancing point cloud recognition training data for a semi-closed scene according to claim 1, wherein in step S300, if there is no adjacent parking area in the driving area, the head orientation distribution in the driving area is a normal distribution with the main driving direction as a mean value.
5. The method for enhancing point cloud recognition training data for a semi-closed scene as claimed in claim 1, wherein in step S300, if there is an adjacent parking area in the driving area, the interactive behavior pattern of the vehicle in the driving area and the adjacent parking area is determined, and the specific determining process is as follows: Based on the central point coordinates of the laser radar capable of identifying the driving area and the adjacent parking areas, combining the central point coordinates of each parking space in the parking areas, constructing a direction vector of each parking space pointing to the central point of the laser radar, and unitizing the direction vector to obtain an orientation vector of each parking space; Calculating an included angle between the orientation vector of each parking space and the main driving direction of the driving area; If the calculated included angle is smaller than a preset angle threshold value, judging that potential interaction behavior exists between the parking space and the vehicle in the driving area, otherwise, judging that no interaction behavior exists between the vehicle in the driving area and the parking area.
6. The method for enhancing point cloud recognition training data for a semi-closed scene as recited in claim 5, wherein determining a head orientation distribution of vehicles within the driving area when there is potential interaction between the driving area and the adjacent parking area comprises: taking the long side direction of each parking space in the parking area as the whole direction to obtain a unit direction vector of each parking space; Aiming at four corner coordinates of each parking space in a parking area under a main axis coordinate system, combining unit direction vectors of each parking space to construct an extension ray of each parking space along the whole direction close to a driving area; Calculating intersection point sets of all extension rays and the driving area, and constructing an interaction area of the current driving area through line segment union sets between all intersection points and corresponding corner points; and assigning values to the head orientation distribution of the vehicles in the interaction area according to the mixed probability model.
7. The method for enhancing point cloud recognition training data for a semi-closed scene as claimed in claim 6, wherein the process of head orientation distribution by the mixed probability model is as follows: The vehicle head orientation distribution theta veh-1 of the parking behavior is generated by the probability p park , the vehicle head orientation distribution theta veh-2 of the normal driving behavior is generated by the probability 1-p park , and the specific expression is as follows: ; ; Wherein, theta j represents the orientation angle of the parking space under the spindle coordinate system, A random disturbance variance of the head direction of parking behavior is represented, theta lane represents the direction angle of a driving area under a principal axis coordinate system, A random disturbance variance of the head orientation representing normal driving behavior, Representing a normal distribution.
8. The point cloud identification training data enhancement method for semi-closed scenes according to any of claims 1-7, wherein in step S500, the rigid body transformation for the vehicle point cloud samples comprises the following procedure: for a vehicle point cloud sample placed at a layout position, acquiring a center point coordinate of the vehicle point cloud sample in a corresponding driving area; According to the orientation angle of the layout position of the vehicle point cloud sample, combining the obtained center point coordinates, and carrying out rigid transformation from a self coordinate system to a main axis coordinate system on the vehicle point cloud sample, wherein the specific expression is as follows: ; ; Where p' represents the coordinates of the vehicle point cloud sample after the rigid body transformation, p represents the coordinates of the vehicle point cloud sample under its own coordinate system, R (θ k ) represents the rotation matrix when the point cloud base map is converted to the principal axis coordinate system, θ k represents the direction angle of the driving area under the principal axis coordinate system, and c i represents the coordinates of the center point of the vehicle point cloud sample at the placement position.
9. The method for enhancing point cloud recognition training data for a semi-closed scene according to claim 8, wherein when the vehicle point cloud sample is placed in step S500, if the placement position is located at the boundary position of two adjacent driving areas, the correction of the head orientation angle is performed according to the length ratio of the vehicle point cloud sample in the two driving areas, and specifically comprises the following steps: setting a center point of a vehicle point cloud sample to fall into a first driving area, wherein a part of geometric structures are positioned in a second driving area; Acquiring eight three-dimensional angular point coordinates of a vehicle point cloud sample in a current gesture, and calculating the ratio alpha of the projection length corresponding to the angular point positioned in the second driving area to the total length of the vehicle point cloud sample; According to the obtained duty ratio alpha, the head orientation angle theta' of the vehicle point cloud sample is subjected to weighted correction, and the specific expression is as follows: θ'=(1-α)·θ 1 +α·θ 2 ; Wherein, θ 1 and θ 2 respectively represent the orientation angles of the vehicle point cloud samples corresponding to the first driving area and the second driving area.
10. The method for enhancing point cloud identification training data for semi-closed scenes according to claim 1, wherein after the layout of the vehicle point cloud samples is completed in step S500, the laid vehicle point cloud samples need to be checked for rationality, which specifically includes the following contents: If the laid vehicle point cloud samples have angular points falling into the non-driving area and the non-parking area, deleting the vehicle point cloud samples; If the three-dimensional bounding boxes of any two arranged vehicle point cloud samples are overlapped, randomly deleting one vehicle point cloud sample; Counting whether the total number of all the vehicle point cloud samples in the driving area meets the preset quantity constraint, if so, stopping the layout of the vehicle point cloud samples, and if not, performing secondary layout until the quantity constraint is met.

Description

Point cloud recognition training data enhancement method for semi-closed scene Technical Field The application relates to the technical field of intelligent transportation, in particular to a point cloud recognition training data enhancement method for a semi-closed scene. Background In recent years, a fixedly deployed LiDAR (LiDAR) has the advantages of all weather, high precision, interference resistance and the like, and is becoming an important sensor for vehicle monitoring and environment modeling in scenes such as service areas, parking lots and the like. Compared with the traditional vision scheme, the laser radar can provide dense and high-precision three-dimensional point cloud data, and is convenient for realizing target detection and behavior tracking under the condition of not depending on illumination conditions. However, in the process of acquiring actual point cloud data and constructing training samples, the phenomenon of unbalanced distribution of vehicles in space positions often occurs for semi-closed areas such as parking lots, service areas and the like. In particular, in the above-described scenario, the vehicle is usually in a stationary or slowly moving state for a long time in the parking space region, while the frequency of occurrence in the traffic lane region is low and the duration is short. This results in a significantly higher number of vehicle samples in the parking space region than in the traffic lane region in the point cloud data acquired over a long period of time. When the point cloud data are marked and used for model training, the training data have obvious defects on the surface of the space covering layer due to insufficient number of vehicle samples in the traffic lane area. When the existing point cloud target detection or recognition algorithm is adopted for training, the model is easy to generate over-fitting for the frequently occurring vehicle features in the parking space region, and effective learning for the spatial form, the posture change and the point cloud features of the vehicle in the traffic lane region is lacking. In the actual application process, when vehicles exist in the traffic lane area, the model is easy to generate the condition of missing detection or misjudgment, so that the accuracy rate of overall vehicle identification and the reliability of the system are reduced. Disclosure of Invention One of the objectives of the present application is to provide a method for enhancing point cloud recognition training data for a semi-closed scene, which can solve at least one of the above-mentioned drawbacks in the prior art. In order to achieve at least one of the above purposes, the technical scheme adopted by the application is that the method for enhancing the point cloud identification training data aiming at the semi-closed scene comprises the following steps: s100, extracting a three-dimensional bounding box of each vehicle from original point cloud data marked on a semi-closed scene according to marking information, and taking point clouds in the box as vehicle point cloud samples to obtain a point cloud sample library corresponding to all vehicles; s200, removing all extracted vehicle point cloud samples from original point cloud data to construct a point cloud base map, converting the obtained point cloud base map into a main axis coordinate system, projecting the main axis coordinate system to the ground, and dividing functional areas including a driving area and a parking area; S300, determining the main driving direction of each driving area according to the geometric form or lane line information of the driving area, and determining the head orientation distribution in each driving area according to the spatial position relation between the driving area and the adjacent parking area; s400, determining the number of vehicle samples to be generated according to a set proportion based on the total number of vehicles marked in the original point cloud data, and randomly selecting a corresponding number of vehicle point cloud samples from a point cloud sample library according to the obtained number of vehicle samples; S500, randomly sampling layout positions in a driving area, placing selected vehicle point cloud samples at the layout positions of the driving area, and carrying out rigid body transformation corresponding to head orientation distribution on the placed vehicle point cloud samples. Preferably, in step S100, eight three-dimensional corner coordinates of all vehicles are extracted from the original point cloud data according to the position, the size and the orientation of the center point of the vehicle under the global coordinate system to construct a three-dimensional bounding box, all the point cloud data in the three-dimensional bounding box are extracted from the original point cloud data in the form of a three-dimensional mask, the required vehicle point cloud sample is formed after the decentralization, and the extract