Pedestrian Recognition Under Vehicle Environment Based On Deep Learning Convolutional Neural Network

Under the trend of the deep integration of AI and real economy,many advanced AI achievements,such as speech recognition,natural language processing and image recognition,have been widely used in the upgrading and transformation of traditional automobiles,constantly improving the degree of intelligent automobiles.As an important branch technology in the process of automotive intelligence,image recognition driven by deep learning has made great progress in recent years.However,due to the limited resources of on-board computing,the deep learning algorithm can only be loaded on the hardware platform of industrial computer,while the deep learning algorithm for embedded hardware transplantation is rarely studied.In this study,pedestrian recognition in complex and uncertain traffic environment is taken as an object.Through comparative study of pedestrian recognition algorithms under various convolutional neural network frameworks,a pedestrian recognition method which consumes less hardware resources and is suitable for embedded hardware transplantation is proposed.The main work of this paper includes:(1)Pedestrian region extraction algorithm is studied from two aspects: first,through ROI extraction methods based on image bottom feature information and visual attention model,combined with morphological operations to remove noise and smooth image boundary,pedestrian feature region is extracted to create pedestrian data set;second,K-means clustering algorithm is used to preliminarily screen the target area of the image,combined with convolution sampling.The feature multi-scale combination method can obtain more accurate pedestrian feature regions.(2)Fast RCNN convolution neural network for pedestrian recognition is designed and constructed.The network is based on AlexNet network model and Caffe learning framework.Through improvement,four layers of RPN network are added to generate pedestrian recommendation area,and one layer is ROI pooling layer for generating fixed-size feature map for each region of interest.After the network training is completed,it is found that the overall recognition effect of Fast RCNN convolutional neural network can meet the requirements of high accuracy,but it occupies high computing resources.(3)To solve the problem of high occupancy of hardware resources in Fast RCNN,a YOLO convolution neural network with less computational resources is designed.Based on Darknet 53 network structure and ResNet layer structure,this network transforms layer-by-layer training into stage-by-stage training,alleviates the problem of gradient disappearance or dispersion in the process of network training,obtains image multi-scale feature map by convolution sampling,realizes pedestrian recognition two-class problem by using logistic regression classifier,and explores Dropout and DropCo.Nnect removes the problem of network fitting.The experimental results show that the Fast RCNN convolution neural network has higher accuracy and lower missing detection rate for pedestrian recognition,but its timeliness is poor and its calculation cost is high.The YOLO convolution neural network has higher accuracy for small-size pedestrian recognition,and has strong real-time performance,less computational resources and more practical value.