Research On Automatic Driving Vehicle Object Detection And Tracking Algorithm Based On Convolutional Neural Network

With the continuous development of the automobile industry in the world,people put forward higher requirements for vehicles.More and more people begin to hope that the vehicles they drive are more intelligent,and autonomous driving vehicles emerge as the times require.Thanks to the continuous improvement of the performance of a variety of sensors,the rapid development of artificial intelligence technology,and the substantial improvement of computer computing power,cars,as a traditional industrial product,are becoming more and more intelligent.Self-driving,a long-cherished dream since the dawn of the automobile,is becoming a reality.Among various key technologies of automatic driving,environment perception plays a pivotal role.As a very economical and feasible scheme,visual environment perception has always been the focus of relevant researchers.Object detection and tracking of obstacles,detection of traffic lights and traffic signs,detection of lane lines,division of road driving area,and other subdivision directions are inseparable technologies with the practical application of automatic driving.The main research content of this thesis is real-time target detection and tracking.The specific research contents are as follows:(1)Obstacle target detection.The current target detection algorithm based on the anchor box is complex,redundant,and slow in the detection process.In this thesis,Center Net,which has a very simple structure and is based on key points,is used as the basis for improvement.To make the algorithm more suitable for real-time detection requirements in autonomous driving scenarios,this thesis proposes the Center Net-IDAMobile Net V3 network,which uses the Mobile Net algorithm series of deepwise separable convolution,inverted residual connection,and other model lightweight ideas,which greatly reduces the model calculation and the number of parameters.At the same time,to make up for the decrease of model fitting ability caused by the decrease of calculation amount and the number of parameters,the idea of iterative deep feature aggregation is applied to carry out the iterative deep aggregation of features generated in different feature extraction stages in the backbone network,so that the features in the shallowest feature extraction layer with the largest resolution are continuously aggregated to the deeper layers.Through this operation,the detection accuracy of the model is improved without obviously increasing the amount of calculation and the number of parameters.(2)Obstacle target tracking.Aiming at the problem that the current target tracking algorithm based on the paradigm of "tracking-by-detection" cannot meet the real-time tracking of autonomous driving scenes,this thesis selects Center Track,the tracking algorithm of joint detection and tracking,as the basis for improvement.Because the backbone network used for feature extraction has a large amount of Hourglass calculation and a large number of parameters,it is difficult to achieve real-time multitarget tracking,which is not suitable for automatic driving scenarios.Center Track-IDAMobile Net V3 network proposed in this thesis,it draws on the Mobile Net series of lightweight convolutional neural network in which deepwise separable convolution replaces standard convolution,inverted residual module replaces standard residual module and other operations,and then uses the idea of iterative deep feature aggregation to carry out the deep fusion of features in different stages.To make up for the reduction of feature extraction ability caused by the lightweight operation of the model,a good trade-off between tracking accuracy and tracking speed is obtained.(3)Open data set and real vehicle verification.The obstacle target detection and tracking algorithm is tested on the public benchmark data set and the data collected by the real vehicle to verify the performance of the algorithm and analyze it.