Research On Detection Technology Of Traffic Flow And Violation Behavior In Intelligent Transportation

With the rapid development of science and technology in China,the national economic strength has significantly improved,and the number of vehicles has increased dramatically year after year,which is followed by increasing traffic pressure.These bring great challenges to the work difficulty and workload of traffic law enforcement personnel.At present,many scholars in China are actively engaged in the field of intelligent transportation.With the rise of deep learning and big data,traffic management has gradually become information-based.In this thises,the traffic flow detection and illegal behavior detection in intelligent transportation are studied,and the technology of target detection and target tracking is improved.The specific research contents are as follows:(1)For vehicle target detection,through the analysis of common target detection algorithms,it is found that YOLOv5 target detection algorithm is better for vehicle target detection,and YOLOv5 s can meet the real-time monitoring of traffic conditions under the surveillance camera.At the same time,this thesis improves the YOLOv5 s network model,and adds the dual channel attention mechanism CBAM-NET to different parts of the Backbone backbone network,which is the most important part for feature extraction.Through experimental comparison,the CBAM2＿YOLOv5s method with better effect is adopted,so that the network model can detect vehicles more accurately.(2)In the research of traffic flow statistics,this thesis proposes a virtual line detection traffic flow method based on target tracking,which uses cross multiplication to judge whether the vehicle passes the virtual line,and uses the improved YOLOv5 algorithm and Deepsort target tracking algorithm to complete the traffic flow statistics.In order to improve the vehicle detection effect of Deepsort,the appearance extraction model of Deepsort is reconstructed,and the improved network model is re recognized and trained;In order to solve the problem of inaccurate detection frame when Deepsort algorithm tracks the target,the Io U cross merge ratio algorithm of Deepsort cascade matching part is changed to CIo U.After the above improvements,the target vehicle will not be misjudged as a new target vehicle when it reappears after being blocked,and the detection frame is more accurate.Proved by experiments,CBAM2＿YOLOv5s combined with the improved Deepsort network model has improved the accuracy of traffic flow statistics by about 7% compared with the original network,which is better.(3)For the detection of vehicle violations,this thesis first proposes a method combining deep learning with manual work to extract lane lines and crosswalks,which is better than the traditional Hough transform method in accuracy and saves time.Then,based on the above Deepsort target tracking algorithm,the Kalman filter is used to predict the vehicle position.Then proposes a method to detect the illegal behaviors of vehicles such as lane changing,speeding,and not giving way to pedestrians on the traffic road according to the vehicle motion track,pixel position change,etc.The detection effect is very good and practical.