Multi-object Detection Based On Sparse R-CNN Algorithm In Urban Traffic Scenes

At present,with the increase of urban transportation demand and the rapid development of computer vision technology,people’s requirements for the environmental perception ability of intelligent vehicles in traffic scenes are getting higher and higher,and the positioning and recognition of objects in urban traffic scenes is of great significance to promote the application of automatic driving and intelligent transportation.Aiming at the problem of object detection in urban traffic scenes,this dissertation proposes a multi-stage proposal sparse regions with convolution neural network features,which improves the accuracy of object detection while retaining the detection speed of the algorithm as much as possible,has good real-time performance,optimizes the detection of small and medium-scale traffic objects,and can excellently complete the task of object detection in urban traffic scenes.This dissertation compares and analyzes the detection performance of existing object detection algorithms,and selects the sparse regions with convolution neural network features(Sparse R-CNN)algorithm with excellent detection accuracy,detection speed,and model structure as the benchmark research model of object detection in urban traffic scenes.The test samples were constructed using the urban object dataset,and the shortcomings of the Sparse R-CNN algorithm in the urban traffic scenarios were investigated experimentally.Aiming at the problem that the Sparse R-CNN algorithm is limited by the number of proposed boxes in the object detection task of urban traffic scenes,the detection speed is limited by the number of proposed boxes,and the detection performance of small and medium-scale objects such as traffic lights and distant vehicles and pedestrians is poor,this dissertation proposes a multi-stage proposed sparse region convolutional network,namely the multi-stage proposal sparse regions with convolution neural network features(MPS R-CNN)algorithm,based on the Sparse RCNN algorithm.The MPS R-CNN algorithm mainly has the following characteristics:a multi-stage proposal box filtering update algorithm is designed,which improves the detection accuracy of the algorithm;a bidirectional parallel feature pyramid network is designed to enhance the feature fusion ability of the model.Aiming at the problem of urban traffic object detection,the copy-paste data enhancement algorithm and complete intersection over union loss function are introduced to enhance the effect of the model on the detection of small and medium-scale objects and improve the rationality of loss calculation.Experimental results show that the mean average precision of the MPS R-CNN algorithm reaches 77% on the urban object dataset,which is 7% higher than that of the Sparse R-CNN algorithm.The algorithm detection speed is kept at 37 frames per second,which is better than other object detection algorithms in urban scenes.