Multi-object Detection And Drivable Area Segmentation Based On Vehicle Vision System

With the rapid development of social economy,unmanned driving technology has become a hot topic of current research.Its core technology systems include environmental perception system,path planning system,decision control system and so on.The environment perception system is similar to human eyes,and realizes object detection through the real-time collection of surrounding environment information during the driving process of the intelligent driving car.Compared with other sensors,the vision camera sensor is inexpensive,and can obtain road ground indicator lines,traffic signs,obstacles and other information by installing in the front of the vehicle.Therefore,vision-based road multi-object detection and segmentation technology has developed rapidly in recent years.This thesis deeply analyzes the research status of road multi-target detection and drivable segmentation in the field of intelligent vehicle environment perception,and summarizes the problems of missing detection and fuzzy edge segmentation.Based on the above,a multi-task algorithm based on vehicle vision is proposed.The main work of this paper is as follows:(1)Based on the in-depth analysis of the principle of YoLov3,in order to solve the problem of missed detection and inaccurate positioning,change the DarkNet53 network into ResNet50 network,and redesign the ResNet50 network with deformable convolution to improve the flexibility of spatial sampling.Besides,replace the FPN with DFPN to capture more detailed feature information,and add an additional learnable weight branch of the structure,which can effectively reduce the probability of missing detection.Experimental results show that the improved algorithm can effectively improve the accuracy and robustness of multi-object detection,and can meet the needs of practical application.(2)Based on the in-depth analysis of the semantic segmentation DeepLabv3+ algorithm,in view of the excessive dilation rate in the ASPP structure,the edge feature of the image cannot be extracted well,which leads to the problem of inaccurate edge segmentation.The dual attention network is introduced to emphasize the region of interest and suppress irrelevant background regions,so as to obtain boundary-optimized segmentation results.Besides,the segmentation loss function is improved.The experimental results show that the improved algorithm can realize the drivable area segmentation more accurately.(3)Based on the in-depth analysis of the problems of intelligent vehicle environment perception,establish a multi-task network which can realize multi-object detection and drivable area segmentation,the network shares a feature extraction network.The training model is integrated into the ROS platform of the "Zhi Neng Xing" intelligent driving vehicle for verification.The verification shows that the multi-task network can accurately achieve object detection and drivable area segmentation in different complex traffic scenarios.