Intelligent Traffic Monitoring System Based On Small Target

Nowadays,the situation of urban traffic congestion and frequent accidents are becoming increasingly serious.Combined with the booming target detection and tracking technology in recent years,this paper builds a set of intelligent traffic monitoring system to monitor the traffic flow,speed and abnormal parking situation.Combined with the functional requirements of the system,this paper studies and implements a set of small target detection and tracking algorithm with high detection accuracy and high tracking accuracy:(1)In this paper,the YOLO＿V3 algorithm is selected as the basic detection algorithm and a targeted optimization scheme is proposed by weighing the detection accuracy and speed through various experiments and performance analysis,combining with the characteristics of small target data.(2)First of all,excessive clipping and scaling will lead to the loss of small target features.This paper proposes a preprocessing scheme that is more suitable for small targets.Multi-size with different probability for cutting;At the same time,a screening mechanism is added to ensure that each image contains at least one target and less negative sample number.In the case of truncation of boundary target boxes,the maximum IOU of the remaining target boxes and anchor boxes was calculated.If the value was less than the threshold,the maximum IOU of the remaining target boxes and anchor boxes was re-clipped to improve the target training utilization rate.(3)To solve the problem that deep network will swallow small target features,this paper adjusted feature extraction module,reduced the number of convolution layers,and increased the number of residual units in shallow network.The original 32 times down sampling output was changed to 8 times down sampling output,and the content of small and medium-sized target features in the output feature information was increased.The improved algorithm was applied to DOTA data set.Compared with the original YOLO＿V3 algorithm,the improved algorithm improved the recall rate and accuracy of small targets,and increased the MAP value by 6.93%.(4)In this paper,the improved YOLO＿V3 algorithm combined with the tracking algorithm to achieve accurate tracking of small targets,combined with the Kalman filter algorithm to predict the trajectory of the movement,and for the actual complex tracking environment,a more comprehensive target correlation algorithm is proposed.Finally,the tracking accuracy can reach 81.5% in the UAV123 data set.Compared with other four tracking algorithms,the tracking success rate and accuracy of the algorithm in this paper are the best.(5)To solve the occlusion problem in target tracking,the occlusion discrimination mechanism and trajectory correction strategy are proposed in this paper.By judging whether there are close adjacent disappearing targets around the new target,the target ID before and after occlusion is matched.When the predicted trajectory drift occurs,the trajectory will be corrected according to the change of the X and Y direction of the target.Experiments show that the tracking algorithm in this paper has strong ability of occlusion discrimination and trajectory correction.