Research On Traffic Target Detection Technology For Autonomous Driving

Traffic object detection is one of the key research issues in the field of computer vision for autonomous driving technology.Traffic targets can be divided into dynamic targets represented by pedestrians and vehicles,and static targets represented by lane lines and traffic signs.At present,deep learning algorithms have achieved great success in the detection and recognition of traffic targets,and they are currently a hot research topic.However,there is still room for improvement in the performance of such algorithms in complicated weather such as rain,snow and fog,as well as under occlusion and dark night conditions.In addition,with the continuous improvement of detection accuracy,related deep networks have become more complex,and there are disadvantages such as huge training parameters and high hardware configuration required.In summary,this paper takes the static target detection represented by the lane line and the dynamic target detection represented by the vehicle as the research goal to improve the research.The main work is as follows:(1)A lane line detection method based on semantic segmentation and spatio-temporal correlation information is proposed.The common method of lane line detection is to detect a single frame of image.However,the images acquired by the vehicle-mounted camera are usually continuous and similar in structure,and the adjacent two frames of images contain a large amount of time-related information.The backbone network of this paper is the lane line detection network model of ERFNet backbone network.The Long Short-Term Memory(LSTM)model is added to the model,and the information of the current image frame is used to compare the lane line information of the current image frame.Make predictions.The method in this paper has been experimentally verified,and the accuracy of the public data set is higher than the original ERFNet network model,and higher than the commonly used lane line detection models(Lane Net,SCNN,etc.).A research on vehicle recognition technology based on lightweight YOLO model is proposed.With the continuous development of target detection and recognition technology,the accuracy of recognition has also been greatly improved.With the development of deep learning,the existing network models are becoming more and more complex,and the network is getting larger and larger,resulting in more and more memory occupied.In this paper,based on the YOLO(You Only Look Once)v4 target detection algorithm,its backbone network is redesigned with the lightweight network Mobile Net V2.The method in this paper has been verified by experiments,and the network parameters can be greatly reduced while maintaining high detection accuracy.The method proposed in this paper is trained and verified on known public data sets,and has a certain improvement in the accuracy and robustness of lane line and vehicle detection.