Urban Vehicle Detection And Recognition Based On Multi-Scale Features

With the construction of smart cities,Intelligent Transportation System as an important component of the construction must be paid attention to.The system can relieve traffic pressure,improve traffic management efficiency and ensure the construction of smart cities.As the key of ITS,vehicle detection and recognition can grasp vehicle and road information by detecting and recognizing vehicles in images.With technological innovation,vehicle detection and recognition based on deep learning has a qualitative leap compared to traditional detection and recognition.However,in the traffic environment,due to the shooting principle of monitoring equipment and the diversity of vehicle scales,the detection effect of vehicles with small pixel scales in the image is not good.Therefore,this thesis focuses on the difficulty of detecting and identifying vehicles with small pixel size in traffic images.To solve the problem.This paper improves the Feature Pyramid Networks,which is better for pixel-scale small target detection,and proposes two vehicle detection and recognition networks respectively.The main research is as follows:Through comparative analysis,Faster R-CNN and Res Ne Xt-101 are selected as the model framework for vehicle detection and recognition.Aiming at the fact that FPN only uses simple concatenation or addition operations when fusing multi-layer features,it fails to make full use of the feature information of different layers,so PFPN detection network is proposed.The network reconstructs the feature pyramid network from the four-layer feature map extracted by FPN through an adaptive fusion method.At this time,the network realizes the full fusion of features of different layers and detects vehicles with small pixel scales better.At the same time,in view of the high overlap rate and the unbalanced number of difficult and easy samples among vehicles,the introduction of Soft-NMS and Focal Loss alleviate and optimize this problem.Similarly,for FPN,only the deep feature information is introduced into the shallow layer and the help of the shallow feature information to the top layer is not considered,so the MPFPN detection network is proposed.The network first designed two structure FPNs.Among them,the Top-down structure FPN constructs feature maps in a top-down manner and inserts global average pooling in the top-level feature maps to better learn the feature information,while the Bottom-up structure FPN is the opposite of the former.Then the two structures of FPN are fused in corresponding layers.At the same time,in order to better utilize and retain the shallow edge shape information,the bottom-layer feature information of the fused feature pyramid network is re-fused and transferred to the top layer through a bottom-up enhancement path,thereby constructing a new feature pyramid networks.Both the PFPN and MPFPN detection and recognition networks are based on the FPN structure.Through the continuous fusion and use of high-level features and shallow-level feature information,the detection and recognition of vehicles with small pixel scales is finally realized.