Research On Traffic Flow Detection Algorithm Based On Target Detection And Tracking

With the accelerating process of urbanization in China,the problem of urban traffic congestion is becoming more and more serious.Traffic flow information can directly reflect the current traffic conditions.Obtaining traffic flow information in real time can facilitate the rational allocation of subsequent traffic resources,timely alleviating road congestion,and providing intelligent basis for relevant decisions of the transportation department.In order to accurately obtain the traffic flow information in the video image,this paper proposes a traffic flow detection algorithm based on improved SSD + Deepsort.The specific work contents are as follows:(1)Aiming at the problem that the SSD target detection algorithm is not effective in detecting distant small-scale vehicles in the traffic scene based on video surveillance,an improved SSD-based target detection algorithm is proposed.The improvement points mainly include two aspects.Firstly,the Inception sparsely connected convolutional structure is introduced into the two convolutional layers of Conv4-3 and Conv7 responsible for small target prediction,so as to widen the network width,and use residual structure connections to enhance the feature extraction capability of the network.Secondly,in the feature map processing stage,the channel attention mechanism-SENet structure is introduced,so that the algorithm model pays more attention to channel features with a large amount of information.After experimental analysis,the improved SSD algorithm improves the detection accuracy of small-scale vehicles in complex traffic scenarios,the recall rate is increased by 4.8%,and the average accuracy MAP is increased by 3.1%.In terms of detection speed,the improved SSD algorithm achieves 57 FPS.(2)Design and implement a vehicle tracking algorithm based on improved SSD+Deep Sort.When the Deep Sort algorithm is used in vehicle tracking,because the feature extractor of the algorithm is not enough to learn vehicle-related features,the effect of vehicle tracking is not good.Based on this,the vehicle re-identification dataset is used to retrain the network model.And a self-made test set is proposed to convert the UA-DETRAC data set into a tracking test set that meets the MOT16(multi-target tracking)format.After performance evaluation on this test set,the improved vehicle tracking model improves the MOTA(tracking accuracy)metric by5.8% while ensuring real-time performance.(3)Finally,in terms of traffic flow statistics,in view of the problem that the original virtual detection line algorithm only uses a single detection line,which leads to poor counting effect and easy to miss detection,an improved virtual detection line method is proposed,which can reduce the number of that vehicle detection is missed in a single frame image near the detection line,resulting in failure to count.A comparison experiment of traffic flow statistics is carried out in three traffic scenarios.The analysis results show that the improved SSD+Deep Sort combined with the improved virtual detection line algorithm’s traffic flow statistics method is more accurate in counting results.