Study Of Semi-automatic Labeling Tool For LiDAR 3D Targets

The development of autonomous driving can effectively solve the problem of traffic congestion caused by the increase in car ownership and reduce the chance of traffic accidents caused by human factors such as driver speeding,drunk driving and fatigue driving.Accurate environmental perception is the first key to the implementation of autonomous driving technology,which requires relying on vehicle sensors to obtain information about the surrounding environment and complete route planning to better participate in traffic.LIDAR’s natural advantage of being able to acquire 3D contours of targets and achieve accurate ranging and real-time trajectory tracking has made it a mainstream sensor commonly used in the industry and the primary choice for making true value labels for autonomous driving data in the field of deep learning.In addition to the influence of in-vehicle sensors,the performance of environment-aware models relies more on a large amount of annotated data,however,most of the mainstream annotated datasets in the field of autonomous driving are foreign urban road scenes,lacking data from roads with Chinese characteristics and college campuses with dense vehicles and vegetation causing severe occlusion.To facilitate the self-production of autonomous driving datasets,this paper combines algorithmic annotation and manual annotation from the perspectives of improving algorithmic annotation accuracy,reducing manual annotation cost and decreasing annotation time to study tools that can realize semi-automatic annotation of Li DAR 3D targets.The high line-count Li DAR is relied on to produce true-value labels to train detection models,and camera data is used to assist annotation to improve accuracy and achieve application innovation.In addition,this paper also makes improvements based on the Voxel Feature Encoding(VEF)module to improve the detection performance of the pure laser 3D target detection model Center Point,which can be used for algorithmic prelabeling.The main work of the study is as follows:(1)To address the problem of low detection accuracy of Center Point for long-range sparse point cloud targets,motion vectors are introduced to enhance target feature representation by exploiting the geometric changes of voxels occurring in a short period of time.A channel attention mechanism is added to weight the motion-sensitive channels of potential features to further enhance point cloud characterization.Ablation experiments show that the added motion vector module obtains a 1.25% improvement in Mean Average Precision(m AP)while the channel attention mechanism obtains a 0.28% improvement.(2)To address the problem that the open source tool 3D-BAT is not universally applicable,the data conversion module is added.Uniform processing algorithm annotates labels and raw data,and automatically calculates the spatial mapping relationship matrix from LIDAR to image coordinates according to the collected vehicle sensor internal and external parameters to realize the reception and annotation function of 3D-BAT for arbitrary data.Combined with the four coordinate conversion formulas of autonomous driving to realize the mapping of 3D target frame rotation around Z-axis on 2D images on point clouds.(3)To address the lack of complex road data sets in China,this paper builds a real vehicle data acquisition platform based on RS-Ruby LIDAR,Moriyun camera and TITAN4,completes the data acquisition work and sensor calibration for complex road scenes within two major universities in Wuhan and some gates and interchanges in Wuhan,and completes the data annotation on the improved 3D-BAT.The total number of collected point cloud data exceeded 4000 frames,and the total number of annotated frames exceeded2000.