Research On Human Motion Analysis In Intelligent Video Surveillance Systems

Intelligent Video Surveillance (IVS) is a cutting-edge technology in computervision. It covers many subjects including image processing, pattern recognition andartificial intelligence. IVS is widely used in traffic monitoring, safety protection,medical care and many customer services, and deals with image sequences tounderstand object behaviors and make intelligent decisions. In recent years, with greatchanges taken place in world security situations, pedestrian detection, tracking andbehavioural analysis in videos using IVS become a more and more active researchtopic.Due to the complicated dynamic backgrounds in videos, variable appearance andwide range of poses of pedestrians, moving camera and size changing of objects,pedestrian detection, tracking and behavioural analysis in public scenes are allchallenging tasks. Therefore, the study of object segmentation, human classification,pedestrian tracking and abnormal behavior detection in crowds has a profoundtheoretical value and broad application prospects.Upon deeply understanding in computer vision principles and the latest researchresults, we first introduce a framework to analyze human motion, then propose severalalgorithms to segment object, classify human, track pedestrian and detect abnormalbehavior in crowd. The contributions of our work can be summarized as follows.1) A double-layer framework of background modeling and object segmentationis proposed. The single layer background model used alone in the conventioanlmethod is not accuracy enough to handle difficulties such as waving trees, lightswitching, complicated background and there will be holes and noise aftersegmentation. A modeling algorithm based on codebook(CB) is presented in thepixel-level layer, while a texture-based algorithm using center-symmetric local binarypattern(CSLBP) in the region-level layer. Experimental results on video sequenceshave demonstrated that our method can process effectively in dynamic situations withbackground disturbance, light switching and other complicated backgrounds. And ouralgorithm is fast enough for real-time applications.2) An advanced method of histogram of oriented gradients (HOG) for humanclassification is given. To avoid the heavy computation of sliding window detectionapproach in traditional object detection procedures, the regions of moving objects aresegmented. Only the segmented regions are scanned by HOG classifier. A fast way of calculating the HOG feature is achived by the integral image method. Experimentalresults show that our method can process compliated dynamic background effectively.The algorithm matches the performance of traditional HOG method with much fasterspeed.3) A pedestrian tracking algorithm to synthesize HOG detection and particlefilter is proposed. The algorithm takes the particle filter as the tracking framework,identifies the target area according to the result of HOG detection and modifiesparticle sampling constantly. Compared with the traditional particle filter algorithm,our method can track pedestrians in dynamic backgrounds more accurately. And ourmethod can handle situations with moving camera and changing size of objectsefficiently.4) A method to detect abnormal activity in crowds is proposed. Features ofcrowd-interest points are calculated first, then a model of crowd features usingmixture of Gaussian (MOG) is established and updated constantly. If the input featureis not close enough to any of the models, there must be an abnormal activity in thecrowd. Our algorithm does not need to segment and track individuals in crowds, andits training of the model is very simple. Experiments show our method can quicklydetect abnormal velocity and escape panic in crowds with a high detection rate and alow false positive rate.5) Finally, all of the above algorithms are implemented to a security monitoringIVS system. The framework of hardware and software are designed. We are in chargeof the implementation of human detection and motion analysis softwares.