A Variational Framework for Pedestrian Segmentation in Cluttered Scenes Using Bag of Optical Flows and Shape Priors

In this thesis, we present a framework for segmentation of pedestrians in videos. Traditional image segmentation algorithms have relied on learning an appearance model for the regions of interest in an image and using that model to segment new images. However, large variations in the appearance of pedestrians due to differences in clothing makes it impossible to learn a reliable appearance model for pedestrians from training videos and using it to segment new videos. Motion has also been used to segment regions of interest from videos. However, the motion of pedestrians is highly articulated, which makes the modeling of pedestrian motion very challenging. For example, the hands of a person may have a very different motion from that of his torso or legs. Therefore, in general, a pedestrian cannot be characterized as a coherent region of intensity, color or motion in a video. In our approach, we first assume that we have a prior knowledge of the type of motion of the pedestrian in front of the camera and model a pedestrian with a joint appearance and motion model. The appearance model is learnt online and evolves temporally, while a bag of optical flows models the pedestrian motion. We augment this model with prior knowlii