Research On Occluded Pedestrian Detection And Lightweight Based On Component Matching

Pedestrian detection is an extremely important part of the vehicle’s advanced driver assistance system,and its task is to find pedestrian targets in pictures or video frames.However,in real scenes,due to the complex background,pedestrian targets often have various occlusions,and the task has high requirements for real-time performance,and the computing power of existing hardware terminals is often insufficient,so the comprehensive performance of pedestrian detection can not be well satisfied.Application requirements.Aiming at the problems of insufficient occlusion detection ability and low detection speed in current pedestrian detection methods,this paper deeply analyzes the research status and cutting-edge theories at home and abroad,and conducts in-depth research on how to improve the performance of occluded pedestrian detection.The main work of the paper is as follows:(1)Research on pedestrian occlusion detection: Firstly,design the overall idea of pedestrian occlusion detection and propose a detection method based on component matching in this paper;Secondly,construct a pedestrian detection dataset and analyze the data features;Further improvements have been made on the YOLOv5 s network to address pedestrian occlusion issues,including the use of improved deep separable convolution on the Neck end,the use of CBAM attention mechanism to enhance the network’s feature extraction ability,the use of repulsive loss to enhance pedestrian occlusion processing ability,and the use of improved non maximum suppression algorithms for occlusion issues to separate overlapping targets as much as possible Targeted design of pedestrian occlusion component matching algorithms for post processing;Finally,experiment and analyze the improved model.The experimental results show that the improved model effectively improves the detection ability of occluded pedestrians,increasing the Recall value of the target from 86.92% to 93.63%,and the m AP50 value from84.59 to 92.26.However,the speed has decreased,from an average of 83 frames per second to26 frames per second.(2)Research on model lightweight for pedestrian detection speed issues: Firstly,perform channel pruning based on BN layer sparsity on the improved model,including constant Ssparsity training,channel pruning,and fine-tuning training;Secondly,Tensor RT is used for model quantification acceleration,which improves the real-time performance of model detection and reduces the demand for hardware device computing resources;Finally,experiment and analyze the lightweight model.The experimental results show that after lightweight,the model reduces the Recall value of the target from 93.6% to 92.56%,the model size from 14.3MB to 4.4MB,and the speed from an average of 26 frames per second to 67 frames per second.While slightly reducing accuracy,it greatly improves detection speed and achieves good balance.(3)Research on the design and practical application of occluded pedestrian detection system: Firstly,design the overall structure of the occluded pedestrian detection system;Secondly,select hardware such as monocular cameras,embedded platforms,and industrial control computers in the system;Finally,the effectiveness of the pedestrian occlusion detection system was tested,including deploying the model to an embedded platform and installing hardware equipment on a real vehicle.Four sets of real test scenarios were selected for targeted experiments.The experimental results show that the Recall value of the model deployed on the embedded platform reaches 91.89%,and the final running speed of the model reaches an average of 36 frames per second.The occluded pedestrian detection system can be effectively applied on the embedded platform.In summary,the occluded pedestrian detection and lightweight research based on component matching in this paper provides a certain reference for the application of occluded pedestrian detection and model lightweight deployment.