Scene Information Estimation And Object Tracking With Occlusion

This thesis focuses on scene information estimation and object tracking with oc-clusion for video surveillance with monocular stationary camera. Here,"scene infor-mation estimation"means to estimate the ground/non-ground region, the relative depthof ground region and the relative unevenness of ground."object tracking with occlu-sion"means to segment regions and track multiple objects with occlusion. Recently,monocular stationary camera based surveillance systems have been applied widely.But the practice of the above two points still have great diffculties. So, research onthese issues, as the hot points of computer vision, has been attracted more and moreattention.The main content and results are as follows.1. A new framework of scene information estimation based on semi-supervisedpedestrian detector is proposed in this thesis. With the statistical results of pedestriandetection for a period of time, the average width of a pedestrian and occurrence prob-ability of a pedestrian at each position of the scene can be learned. Then, the relativedepth of ground region, the ground/non-ground region and the unevenness of groundcan be estimated. Compared with traditional scene information estimation methods,new method can work well for video surveillance scene based on common monocularstationary camera with uneven ground and unknown camera parameter. In addition, theTri-training semi-supervised method is applied for pedestrian detector training in thisthesis. With small amount of labelled samples, satisfied detection result is achieved.2. A new method for segmenting and multi-object tracking based on local textureis proposed. Exploiting the characteristic that most local parts are distinguishable intexture and stable in spatial location relative to the center of the object in short time,local texture based part-appearance model is proposed. Then, with the online-learnedappearance information of non-rigid objects before occlusion, the locating and regionsegmentation of multi-object with occlusion can be obtained. New method can seg- ment and track multiple non-rigid objects through severe occlusion with various spatialdeformation, unpredictable movement, scale changes, illumination changes, invisibleground and unknown object type.3. A new method for segmenting and tracking multiple rigid objects based onMarkov Random Field (MRF) is proposed. The real-time appearance informationof rigid objects (contour and texture) is extracted with a kind of special structuringelements in Mathematical Morphology (MM) theory. Then, one-dimension spatial-temporal MRF model is built to segment and track multiple rigid objects with occlu-sion. As a result, the assumption on the vehicle shape model in prior is not needed, andthe sites in MRF are reduced, which saves computational cost greatly.All of the above methods are validated suffciently, and experiments demonstratethe effectiveness of the methods proposed in this thesis.