A Study On Pedestrian And Vehicle Detection Based On Deep Learning

With the development of artificial intelligence technology,pedestrian and vehicle detection has been widely used in many fields such as intelligent security,intelligent transportation,driverless and so on.And it has attracted more and more attention in academia and industry.The traditional target detection algorithm has insufficient feature expression ability and is not robust to the variety change,so it can not meet the needs of practical application.pedestrian and vehicle detection is still a challenging problem.In recent years,because of the wide application of convolutional neural network(CNN)in the field of machine vision,the performance of target detection has been greatly improved.Compared with traditional detection methods,convolutional neural network adopts an efficient target location method and can extract deep semantic features adaptively according to the category of objects,Thus,the real-time and high-precision target detection is achieved.This thesis takes the surrounding area of the campus as the research site and focuses on the detection of pedestrians and vehicles.Starting from the characteristics of surveillance video,the difficulties and problems in target detection of surveillance video are deeply analyzed,and performance improvement is also made for specific problems.It enables the detection method based on deep learning to be better applied to target detection in surveillance video scenes.The main research work of this thesis includes the following aspects:(1)Data set establishment The data set in this thesis consists of two parts.Part of the data is from the PASCAL VOC dataset,which contains 20 categories.Two kinds of pictures of pedestrians and vehicles are screened out for training of detection model;The other part comes from 10 video data sources,which are captured by cameras around the campus.After dividing the video into frames,the pedestrians and vehicles in the image are labeled manually.The labeled pictures are extended into the data set for network training to enhance the robustness and generalization ability of the model.(2)The main target detection algorithms based on deep learning(R-CNN,Fast R-CNN,Faster R-CNN,YOLO and SSD)are analyzed and the hardware platform is built.On the basis of experiments,the training and testing of Faster R-CNN,YOLO and SSD models are completed.Considering the performance of the above three target detection algorithms in detection accuracy and speed,finally the SSD network is selected as the basic detection network model.(3)As the camera’s shooting range is relatively large,the targets appearing in the monitoring video are mostly small targets,which brings difficulties to target detection.To solve this problem,this thesis proposes a new detection algorithmtarget The input image is divided into 4 sub-areas on average,then zoom in to the size of the original image to increase the proportion of the target.And the idea of metric learning is introduced on this basis.The test results show that compared with the original model,the detection performance of the proposed method is significantly improved.For pedestrians detection,the true positive rate(TPR)increased by about 28%in daytime and 26%in night.For vehicles detection,the true positive rate(TPR)increased by about 16%in daytime and 29%in night,and there were no cases of false positive.