Fast Video Vehicle Detection Method In Complex Traffic Scene

Intelligent transportation system(ITS)is a solution to the transportation problems caused by current economic development.Real-time and accurate vehicle object detection is an indispensable function in intelligent transportation systems.However,vehicle object detection in traffic scenes is susceptible to factors such as differences in appearance,differences in imaging scale,partial occlusion,weather conditions,camera shooting angles,large amount of image data collected,and high computational complexity.As a result,the real-time,accuracy and robustness of existing vehicle object detection methods are not satisfied.In order to solve the above problems,this paper improves the object detection model YOLOv3 based on deep learning technology.First,simplify the backbone network to reduce the calculation;then design an appropriate multi-scale detection strategy;and finally improve the prediction frame screening mechanism,and obtain more ideal detection results when processing vehicle detection in traffic scene.(1)In order to prevent the problem that the number of parameters and the calculation are too large,the running speed of the vehicle detection model is too slow and the real-time performance is poor,this paper proposed a shuttle residual structure.Compared with the ordinary residual structure,this structure has a smaller model size and lower calculation,but at the same time can maintain the accuracy of model detection.A new feature extraction network is constructed by using reasonable stacking of shuttle residual structure,which achieves the purpose of speeding up the calculation and reducing the amount of floating-point calculation while maintaining accuracy.(2)Aiming at the reduction of detection accuracy caused by large scale changes in vehicle target detection,and the problem of missed detection of vehicles due to too small scale.This paper proposes a multi-scale detection strategy suitable for vehicle target detection in traffic scenarios by constructing a multi-level and multi-layer feature pyramid structure.The adaptability of the vehicle target detection model to scale changes is improved,and the problem of missing detection due to the small scale of the vehicle is solved.(3)In view of the problem of vehicle missed detection caused by vehicle congestion and even overlap,this paper proposes to use Soft-NMS instead of NMS to filter the prediction region.Soft-NMS will not directly delete the overlapping prediction region but give it an opportunity to compare with other prediction region.Using this method greatly reduces the probability of vehicle missed detection due to overlap.Comparing proposed method with the existing vehicle object detection method under different data sets,the experimental results show that the proposed method has the advantages of strong adaptability to scale changes,high detection accuracy of small target vehicles and fast algorithm speed.Proposed method has achieved more ideal detection results when dealing with vehicle object detection in traffic scene.