Traffic Congestion Detection And Analysis Based On Deep Learning

The rapid economic development has triggered people's desire to buy private cars.The speed of car purchase far exceeds the speed of traffic construction,resulting in more and more serious traffic congestion.Reasonable analysis of road traffic status is an important basis for traffic management and control.Traffic information collection technology is a means to obtain traffic data,and it is also the basis for analyzing traffic status.Traditional collection technology is relatively mature and accurate,but it is easily affected by external conditions.The video detection technology is easy to install,maintain,detect a large amount of information,and provide image data.Therefore,analyzing the traffic status based on the video detection technology has certain research value.Based on the existing vehicle detection and tracking algorithms and congestion status discrimination methods,the thesis proposes deep learning-based traffic congestion status detection and analysis,studies related algorithms of target vehicle detection and tracking,selection of traffic parameters,traffic congestion status discrimination models and traffic jam state prediction model.The thesis first introduces the principles of several mainstream target detection algorithms,and conducts an experimental comparative analysis of these detection algorithms.The YOLOv3 algorithm with relatively good detection effect is selected as the basic network for vehicle detection.Then the principle,advantages and disadvantages of three target tracking algorithms are introduced.After analysis and comparison,after analysis and comparison,the recursive Kalman filter algorithm is used to track the prediction box output by YOLOv3,and uses the appearance characteristics of the target to match the tracking target.According to the change information between two adjacent frames in the detection video,the three traffic flow parameter information of traffic volume,traffic density and average speed are obtained.The Kalman filter algorithm makes full use of historical information,narrows the image search range,and improves system performance.Secondly,the classification level and basis of traffic congestion at home and abroad are studied.This paper proposes to combine the three traffic parameters of flow,traffic density and speed as the basis for the judgment of the traffic state,and at the same time use the threshold value to distinguish the traffic state.In addition,in terms of traffic state prediction,some mainstream prediction algorithms are introduced in detail.After experimental comparison and analysis,the Cuckoo Search Algorithm optimized Wavelet Neural Network Algorithm with better effect is selected as the basic model,taking into account the accuracy of the global search ability and local search ability of the basic model.As a result,the fixed step size and the probability of discovery are adjusted to change dynamically,which effectively improves the prediction accuracy.Based on the video image data of Shanghai City Road,the above methods are tested and verified.