Research On Important Issuses Of Video Pedestrian Detection And Tracking

Object detection and visual tracking are basic and hot research topics in the field ofcomputer vision, and serve as indispensable aspects in video surveillance, robot navigation,intelligent traffic, intelligent driving assistance, human-computer interaction etc. It is knownto all that cameras are prevalently used in public places for better management, surveillanceand safty purposes, such as subway stations, banks, airports and traffic intersections. Suchapplications focus on pedestrians and cars, and are seriously dependent on object detectionand visual tracking technology. These two topics are in close relationship, and there are stillmany problems for themselves. It is meaningful and valuable to do research on these topicsand develop application technology for video surveillance and intelligent traffic. For videosurveillance and intelligent traffic, we do research on key issues of visual tracking, multiplepeople tracking and pedestrian detection.The goal of visual tracking is to robustly track objects under common difficulties, such asocclusion, changing appearance, turning and background clutter. Single feature or model cancapture or learn some aspects of the whole visual characteristic of the targets, and resolvesome tracking difficulties. However, features and models are also with disadvantages forvisual tracking. We do research on two good tracking methods, and make improvements withother features or models. Some improvements and innovations are listed as follows.(1) A composite subspace based tracking method is proposed. Incremental principalcomponent analysis learns the principal pattern of object appearance effectively,leading to good tracking ability. However, it is difficult to capture the current patternwhen the appearance is changing rapidly. There is a linear subspace method whichcan learn latest appearance. We combine them artfully, resulting in a compositesubspace with merits of both. The composite subspace tracking method track robustly with high adaptability.(2) A multiple-patch tracking method with feature selection is proposed. Multiple-patchtracking is an effective method, and can deal with occlusion problem. The patchesand features are the keys of the method. And current best discriminative features canbe selected to improve the tracking performance. A patch generating method based onweighting features is also proposed. The proposed tracking method has goodperformance, and can track robustly under occlusion and complex backgrounddifficulties.Multiple people tracking and pedestrian detection have developed rapidly in recent years.They are the key technologies to video surveillance. Some improvements and innovations arelisted as follows.(1) A multiple people tracking method based on online two-stage association is proposed.Multiple people tracking based on tracklet is one of the best methods recently. Wepropose to improve tracklets by combining partical filtering tracking method insteadof only connecting detection responses. Scattering problem caused by missing andfalse detection is improved effectively, and long tracklets are obtained. Previousmethods make tracklet association after a long time. We propose a method whichassociates tracklets after each short interval, and enables near real-time peopleassociation. The experimental results indicate that the method obtains goodperformance, and can be applied in crowded scene.(2) For video surveillance, a Two-frame-filtering(Tff) processing based video pedestriandetection method is proposed. Although pedestrian detection has achieved prominentdevelopment in recent years, few works research detection using motion cue forsurveillance. Tff processing which exploit motion cue by the gray value variationbetween two frames is proposed. Tff gradient feature are proposed and used to train apre-detector to exclude most of the background regions. Histogram of Tff orientedgradient(HTffOG) is also proposed. The discriminative HTffOG is used to trainpedestrian detector. Experimental results show that this effective method achievesgood performance, and is suitable for real-time surveillance and intelligent trafficapplications.