Research And Implementation Of 3D Point Cloud Target Recognition System In Highway Scenes

With the development of lidar technology and computer technology,the identification and application of 3D point cloud data have gradually become an important means of intelligent driving assistance.At present,most driving assistance tasks still rely on two-dimensional optical images,but the acquisition of two-dimensional optical images has certain requirements on lighting and environment.In some light control areas or areas with poor image acquisition quality,the use of three-dimensional point cloud data can effectively compensate for the defects of image data.However,there are some difficulties in point cloud data processing,such as high computing power requirements and slow processing speed.Therefore,the optimization and acceleration of the point cloud segmentation recognition algorithm are studied in this thesis.Based on the idea of edge computing,a 3D point cloud target recognition system that can be used for assisted driving in highway scenes is designed and implemented from the perspective of engineering.The main research work of this thesis is as follows:（1）Combined with the actual road data,the voxel sub-sampling algorithm is used to extract the sparse data,and the sparse road data set is constructed and analyzed.It was found that the proportion of different types in the data set was extremely unbalanced.Based on the squeezeseg V3-21 network,Focalloss was used to improve the loss function,and the m Io U of the improved model on the test set was improved by 2.2% compared with the model before improvement.（2）Given the shortage of computing resources of edge computing equipment,Tensor RT is used to accelerate the model quantitatively,and the acceleration effects under different conditions are compared.Finally,the inference time per frame is shortened from0.454 seconds to 0.125 seconds,and the acceleration ratio is up to 72.5%.The Tensor RT acceleration engine is packaged and the real-time inference module is designed and implemented.So that the prediction of 3D point cloud data can meet the requirement of real-time.（3）Referring to the general architecture of edge computing,a 3d point cloud target recognition system with a 1 center-N edge structure is designed and implemented.Docker container technology is used to realize the mass deployment of edge device environment,and realize the functions of the center and edge subsystems respectively.Finally,the system is tested by the black-box testing method,and multi-thread technology is used to further optimize the reasoning performance of the edge subsystem.The 3d point cloud target recognition system designed and implemented in this thesis can efficiently recognize and segment the point cloud data collected during driving in realtime and effectively reflect the road situation around the vehicle,to achieve the purpose of assisting driving.After testing,the system meets the requirements of accuracy and real-time performance and has high practical application value.