Research On Pedestrian Detection Technology Based On Shallow Learning And Deep Learning

Over the last decade,our lives greatly facilitated due to the rapid and steady development of science and technology,especially the development of computer technology,Internet technology etc.Intelligent robots,intelligent surveillance,intelligent transportation and automatic driving have become the current hot spots.As an important research field of computer vision,pedestrian detection technology has attracted more and more attention from domestic and foreign researchers and technological enterprises.Real-time pedestrian detection system with high accuracy has become an indispensable part of many intelligent systems.This paper aims to study a high-precision,real-time pedestrian detection algorithm,and based on this algorithm to develop a simple and friendly software for our lives.The main contents are as follows:(1)The traditional ACF+AdaBoost pedestrian detection framework was analyzed and improved.When the satisfied detection rate of traditional ACF+AdaBoost pedestrian detection framework reached,its error rate also increased rapidly,and the traditional ACF+AdaBoost pedestrian detection technique cannot meet the practical requirements.In order to resolve this problem,a hash code feature is proposed which adaptively weighted to increase the diversity of pedestrian features and improved the ACF features.What's more,an auxiliary network is added to reduce error rate of the system.The auxiliary network adopted a slight CNN structure to classify the classification results of the AdaBoost classifier again based on the real-time performance of the system.(2)The YOLO v3 pedestrian detection network based on the Darknet framework is studied.Firstly,the Darknet network framework was built.Then,the CrowdHuman data set is choose as the training collection of the pedestrian detection network.The network training parameters were configured according to the data collection and the running environment of network.At the same time,the GPU was used to speed up the network.The test results of network showed that our network can achieve more accurate real-time detection results.(3)A simple,accurate,and real-time pedestrian detection software based on MFC was designed and developed.We built the dynamic link library for YOLOv3 and used the MFC to design software interface according to the function and aesthetic requirements of the software.At the same time,the OpenCV platform was applied toperform the pedestrian detection task along with the created dynamic link library.Experimental results demonstrated that the designed software could stably run in realtime.