Research On Vehicle Detection And Statistics Based On Improved YOLOv4

Intelligent transportation is the future development direction of the transportation industry.Traffic flow statistics are an important part of the intelligent transportation system.The use of image processing technology can directly count the vehicles in the traffic video,provide decision-making data for the traffic management department,and improve the efficiency of road transportation.It can also avoid the low efficiency of manual counting and the negative impact of laying ground sensing coils on the road surface.Based on the existing vehicle statistics algorithm,this paper has done corresponding research on vehicle detection,tracking and statistics.In terms of vehicle detection,this paper uses an improved YOLOv4 algorithm to detect vehicles.In order to improve the accuracy of the training model,we make a VOC data set that is more suitable for my country’s road conditions,and manually label a total of 35,000 cars from 4,523 photos collected on the flyover,and divide them into a training set and a test set to train the model later.Aiming at the problem of poor detection of small target vehicles in remote scenes,the K-means++ clustering algorithm is used on the vehicle training set to improve the matching degree between the prior frame and different feature layers and stabilize the clustering effect;To improve the YOLOv4 network structure,the lightweight CSP module is designed to replace the CSP module,by using the Dense Net module to replace the five-order convolution module in the feature pyramid.The feature network is simplified while the accuracy and detection speed are improved.Based on the verification of the test set,it is found that the average detection time of the improved YOLOv4 model is 5 milliseconds faster than the original model.In terms of average accuracy,the improved YOLOv4 model is 3% higher than the original model and 6% higher than the YOLOv3 model.The number of frames transmitted per second(FPS)can reach 25 frames per second,which can meet the requirements of real-time detection.In terms of vehicle tracking,a vehicle tracking model is established based on the Kalman filter algorithm;the IOU distance is used as the evaluation matrix,and the Hungarian algorithm is used to process the matching relationship between the vehicle detection frame and the tracking prediction frame.The detection model integrated with the tracking algorithm effectively suppresses the phenomenon of vehicle misdetection and missed detection.In terms of vehicle counting,a virtual detection line is drawn at a suitable position on the road,and the vehicle detection model of the fusion tracking algorithm is used to count the vehicles in the video.Using the improved model to count the traffic flow of a single lane,the accuracy rate is 96%,which is 4% higher than the original model;the traffic flow of the three lanes is counted,and the accuracy rate is 93%,which is 2% higher than the original model,which has certain practical application value.