Research On The Method Of Pedestrian Detection In Video Images

With the development of technology, video surveillance, especially intelligent pedestriandetection, has become an important research topic in the field of computer vision. There areno effective solutions for pedestrian detection which both have high detection rate andreal-time processing speed. In this paper, prdestrian detection methods have been studied indetail, and a real-time detection system is constructed on the video images.The system designed in this paper is divided into three parts: classifier training module,moving detection module and classification module, using real-time, accuracy and robustnessas the evaluation rules. We search the whole video frames to find the interesting motionregions, which may contain moving pedestrian we need. Then these regions will bedetermined to pedestrian or non-pedestrain regions by using the frame information and trainedclassifier.In classifier training module, histogram of gradient (HOG) features are extracted from alltraining samples and the dimension of the features are reduced by principal componentanalysis (PCA) to form the HOG-PCA features. These features will be sent to support vectormachine (SVM) for training. We will find the best classifier parameter according to thebalance of speed and accuracy. Experiments show that the HOG-PCA+SVM, which arecompared with HOG+SVM, can both improve0.2%accuracy and42%detection time.In the moving detection module, three-frame differential and background subtractionmethods are combined to get the interesting regions, in order to overcome the disadvantagesof separate method. The improved selective background update and the threshold selectionalgorithm are used to improve the accuracy of getting the moving target, which can adapt tothe robustness of the environment change.In detection module, a moving object list is constructed to save moving areas thatcontain pedestrian. The target areas with high matching mearsure are determined to pedestrianareas by calculating the matching measure with the object list and updated in the list.Otherwise, areas with low matching mearsure will be sent to the classifier to find pedestrian.Experiments show that this method can speed up the detection time and improve the accuracy.Large of tests show that our system can meet the real-time, accuracy, robustnessrequirement. Our system has95.38%detection rate and detection time for98ms per frame,which lay the foundation for pedestrian tracking, analysis and abnormal warning.