Deep Learning Based Vehicle 3D Size Overrun Detection Technology

Vehicle overrun problem is one of the major factors endangering road safety.Despite the numerous fixed overrun inspection stations established in various regions,there still exist detection blind spots that can be bypassed.Mobile overrun inspection is an effective means of controlling overloading on key road sections.Traditional overrun detection techniques are mainly used at fixed detection stations,with expensive equipment and complex detection systems that cannot cope with mobile overrun detection scenarios.To address this problem,a single-camera imaging approach is utilized to obtain vehicle information,and overrun detection is completed based on depth technology.The main research contents are as follows:(1)Research on a vehicle overrun detection framework based on deep learning.Based on existing monocular 3D target detection methods for vehicles,and combined with the mobile overrun detection scenarios,a two-stage geometric constraint algorithm is used as the theoretical basis for overrun detection.A two-stage vehicle overrun detection framework is proposed: the vehicle type and position information is obtained through the vehicle detection and recognition sub-network,and the vehicle overrun measurement is completed through the vehicle overrun detection sub-network.(2)Research on a vehicle detection and recognition sub-network based on YOLOv5 sGhost.The YOLOv5 s network is used as the vehicle detection and recognition sub-network.To address the problem of large numbers of C3 modules in the YOLOv5 s network leading to excessive parameter volume,the traditional convolution within the bottleneck of the C3 module is improved to Ghost convolution,ensuring both high network detection accuracy and inference speed.(3)Research on a vehicle size overrun detection sub-network based on Mobile Net-SE.The lightweight Mobile Net is used as the backbone network for feature extraction,and the SENet attention mechanism is introduced to enhance the network’s feature extraction capabilities,with greater focus on key vehicle areas.A three-branch feature regression module is designed to regress multiple attribute parameters,further deepening the network depth while improving network expression capabilities.A reasonable vehicle size and heading angle regression loss is designed to regress more accurate vehicle size and direction information.Through testing the vehicle detection and recognition network on a self-built dataset,experimental results show that the vehicle detection m AP accuracy reaches 94.4%,the axle recognition AP accuracy is 91.6%,and the network inference speed reaches 55.5 FPS.Through testing the vehicle size overrun detection network on the KITTI dataset,experimental results show that the maximum detection error of the network within a 10 m range from the vehicle is 6.7%,and the inference speed reaches 47.6 FPS.The actual results of vehicle overrun detection show that the maximum detection error of vehicle size is less than 7%,and the vehicle overrun problem can be accurately identified.The proposed method has important significance and research value for strengthening the mobile vehicle overrun detection of key road sections.