Research On The Algorithm Of Pedestrian Detection Based On On-board Vision Systems

Traffic accidents occurs frequently,which bring serious harm to people’s lives,the road traffic safety issues have attracted widespread attention.Advanced driver assistance system such as automobile active collision avoidance system can avoid traffic accidents to a certain extent,which becomes a hot research direction especially in recent years.Pedestrians,as vulnerable groups in the traffic environment and special individuals with lives,so protecting the safety of pedestrians have great significance.Pedestrian detection technology captures road environment information through a camera,detecting pedestrians,and further to protect pedestrians.Therefore,camera-based pedestrian detection technology has great practical value,which will be a key technology in the realization of advanced driver assistance technology and driver-less driving technology.Due to the diversification of pedestrian targets and the complexity of the road environment,the research of pedestrian detection technology based on on-board cameras is a challenging topic.This paper summarized the research status of pedestrian detection technology at home and abroad,and summarized related technologies.With regard to pedestrian detection application for on-board camera,detection accuracy and detection speed are very important evaluation indicators.The ACF algorithm achieves a very high detection accuracy and achieves the real-time detection speed on the pedestrian detection problem.However,the ACF algorithm still has problems of missing detection,especially the detection of small targets and more false detections.Aiming at to further improve the detection accuracy and detection speed of pedestrian detection algorithms and meet the requirements of practical applications,this paper develops research based on ACF algorithm.The main research work of this paper is summarized as follows:(1)By studying the research status at home and abroad and technical difficulties of pedestrian detection,the existing technologies are studied and summarized.The pedestrian detection related technology was studied from the aspects of pedestrian candidate region search,feature extraction and commonly used classifiers.(2)Studied the ACF algorithm comprehensively.In view of the problems existing in the ACF algorithm,this paper proposed a phase-enhanced soft cascading Adaboost classifier,and to improve the detection speed,an improved method for non-maximal suppression algorithm is proposed.Based on Caltech training date set,a training set is constructed,on which the proposed improved classifier was trained offline,and compared experiments were performed on Inria data set,Caltech data set and self-acquired video set.The experimental results show that the improved algorithm proposed in this paper has fewer false detections and missed detections,better real-time performance,and better detection for distant small targets.(3)Taking into account the advantages of the ACF algorithm and the superiority of the convolutional neural network in image classification and target detection,this paper presents a pedestrian detection framework based on the soft cascading Adaboost classifier tandem convolutional neural network.Studied convolutional neural networks and other related theoretical knowledge.Based on INRIA data set and Clatech data set,a training and test set is established,and considering the three aspects of network depth,convolution kernel size and feature dimension used for classification finally,on which the convolutional neural network structure for pedestrian detection is designed,trained,tested and analyzed,and a relatively good network structure is selected.In tandem with the soft cascading Adaboost classifier,the pedestrian detection framework based on the soft cascading Adaboost classifier concatenated convolutional neural is constructed.Finally compared experiments were performed on Inria data set,Caltech data set and self-acquired video set.The experimental results show that the improved algorithm proposed in this paper has a good real-time performance with fewer false positives and fewer missed detections,also has room for further optimization and significance.