Object Detection Based On Neural Network With Feature Reuse In Traffic Environment

Object detection in traffic evironment is an essential part of the auxiliary driving and automatic driving system.Object detection based on deep learning is one of the most popular research directions in the field of computer vision.This paper is based on the “National Key R&D Program”(2016YFB0101102),“Electric Vehicle Intelligent Assisted Driving Technology R&D and Industrialization”,The main research is how to realize the target detection method based on neural network,no matter what the working condition,can use the vehicle camera to detect the position and category of dynamic target information.In order to balance the real-time and robustness of the target detection method,the most popular object detection algorithm YOLOV3 is improved from the perspectives of data set acquisition and processing,network structure construction and optimization,loss function,training scheme design,and so on.And compared with other mainstream target detection algorithms.The work Finished in this paper is as follows:Firstly,in order to meet the requirement of data amount in deep learning neural network training as much as possible,this paper adopts Coco data set to train the network model.Various data enhancement methods,such as random cropping,random filling,random inversion,random rotation,and random mixing,are adopted to expand the original data volume to 3-5 times,and the real annotation information is adjusted accordingly.The adjusted annotation information is divided into 9 categories by K-means algorithm.The clustering center of each cluster is taken to obtain the anchor frame anchor required in this paper for subsequent network training.Secondly,in view of the lack of feature extraction network used by YOLOV3 algorithm,This paper designs an improved DB-NET network.In order to further prevent gradient explosion and gradient disappearance of network in the training process,module residual structure is added.At the same time,in order to improve the operation efficiency of the network,the convolution layer of 1*1 kernel size is flexibly added for data reduction and reduction,and the deep separable convolution is used to replace the ordinary convolution.This paper found that the detection result of YOLOV3 was not ideal when the target size was more than 50% of the width and height of the image or the target size was less than 7% of the original image.Therefore,this paper used the pooling of spatial pyramid to build the receptive field change module to enhance the detection range of the predictor on the target size and achieve the effect of improving the detection rate.Thirdly,for YOLOV3,the central coordinate of the target is separated from the width and height of the boundary box and different loss function prototypes are adopted for training,which results in poor relevance of the target information and insensitivity of the loss function of square deviation to scale,which will affect the detection performance to a certain extent.Therefore,this paper uses the CIOU loss function to calculate the loss of the target’s positioning information as a whole.In order to further balance the imbalance between positive and negative samples,this paper uses Focal Loss as the prototype of the confidence loss function.Finally,this article is based on Windows 10,Python3.6,Tensorflow,Keras development environment,he performance improvement of the designed DB-NET is experimentally studied,Compared with the mainstream network structure Res Net and Darknet-53,a comparative experimental analysis is carried out,and replace the different loss function,the repetition of multiple target detection algorithm to verify the superiority of this method,in addition,this article also tested a variety of complex working condition of testing result,verify the robustness of this method.