Research On Object Detection And Segmentation Technology For Autonomous Driving Visual Environment Perception

With the development of the automotive industry,self-driving cars have become one of the new research hotspots.With assisted driving or autonomous driving functions,selfdriving cars not only provide consumers with a better driving experience,but also improve driving safety and alleviate traffic congestion problems.The self-driving car consists of an environmental sensing system,a decision-making system and a control system.The environmental sensing system is responsible for collecting information about the road environment around the vehicle,so the performance of the environmental sensing system is crucial to driving safety.Among them,the visual environment perception system is an important part of the autonomous driving environment perception system,so it is important to study the autonomous driving visual environment perception algorithm.Based on deep learning methods,this paper presented an improved study of object detection and semantic segmentation algorithms for autonomous driving vision tasks,and on the basis of this study,proposed a multi-task network AMTNet that can simultaneously complete object detection,drivable area segmentation and lane line detection tasks.The specific research in this paper can be summarized into the following aspects.Firstly,this paper conducted some research on existing object detection algorithms for autonomous driving vehicles,and built the autonomous driving object detection network of this paper on the basis of YOLOv5.Improvements have been made to the CSP module in the network by eliminating one of the convolution modules and replacing another convolution module with a CBS module,and adjusting the order of the output layers to reduce the computational effort.The network was trained and validated using the BDD100K datasets,and the network improved in the average detection accuracy of 13 classes of traffic objects.Secondly,this paper investigated and compared existing autonomous driving semantic segmentation algorithms and designed this paper’s autonomous driving semantic segmentation algorithm based on the DeepLabv3+ architecture.The backbone network has been replaced with the lighter MobileNetv2,and an auxiliary loss branch was added to optimize the network training.The network was trained and validated using the Cityscapes dataset.Experiments showed that the improved network achieves a good balance between accuracy and real time performance,with a large improvement in detection real time performance.Finally,based on the previous research,this paper proposed a multi-task network AMTNet for autonomous driving visual environment perception,which used an encoder and decoder structure to build the network for simultaneous object detection,drivable area segmentation and lane line detection tasks,and used the BDD100K dataset for end-to-end training and validation of the AMTNet network.As the multi-task network shares commonalities between several tasks,the performance of the network is improved on all three tasks and the real-time performance of the network is improved.The experiments showed that AMTNet achieves excellent performance and real-time performance.