Research On Lane And Vehicle Detection Algorithm Based On Multi-task Convolutional Neural Network

Lane and vehicle detection are important tasks in the automatic driving environment perception system,as well as the basis of vehicle path planning,decision-making and control.While the convolutional neural network has been widely studied and applied to various visual detection tasks,it also provides a new approach for traffic target detection in intelligent vehicle vision system.Therefore,based on the theory of convolutional neural network,it is of great significance to develop an algorithm with high accuracy,high real-time performance and strong robustness to detect lane and vehicle.A multi-scene lane dataset is established,which contains 50,000 road images with lane annotations.A global lane detection method based on key points and gradient harmonizing loss is proposed for the elongated linear property of lanes.By designing the key points to represent the complete lane,a lane key points detection model is constructed combining lane classification and key point coordinate regression to realize end-to-end global lane detection.Aiming at the problem of unbalanced samples in network training,the gradient harmonizing loss function is designed to optimize the performance of the model.The experiment results show that the model applies to a variety of driving conditions with high robustness and positioning accuracy,as well as superior real-time performance,verifying the effectiveness of the proposed algorithm through multiple evaluation indicators.A vehicle dataset in road traffic environment is established,which contains 14,092 road images with vehicle annotations.By studying the multi-scale prediction structure and object detection principle of YOLOv3 framework,a vehicle detection method based on YOLOv3 network is proposed.To obtain the model parameters applicable to vehicle target,the influence of model parameters on vehicle detection performance is demonstrated through experiments,including input image size,loss weight coefficient and negative sample threshold.Experiments show that the final vehicle detection model has improved detection accuracy effectively without reducing the processing speed.In order to meet the real-time application requirements of lane and vehicle detection,a method based on multi-task network is proposed for joint lane and vehicle detection.The constructed multi-task network consists of basic network Darknet53,YOLO-Head branch network and Lane-Head branch network,detecting lane and vehicle simultaneously on a single network model.Network training and testing are carried out on the lane dataset and vehicle dataset with the alternate training strategy combining full parameter update with exclusive parameter update.Meanwhile,the road arrow sign recognition task is added to further verify the multi-task detection ability of the model.The experiment shows that 1)the training loss of multi-task network can converge stably,2)the vehicle detection performance is slightly lower than that of YOLOv3 model,and the lane detection performance is similar to that of the lane key point detection model,3)the increase of detection task does not reduce the multi-task detection ability of the model,which fully verifies the effectiveness of the multi-task network structure and training strategy.Compared with using two independent network models,the speed performance of the proposed joint detection algorithm is improved by 39.61%,which has high practical application value in the automatic driving engineering.