Detecting Pedestrians Using Motion Patterns: a Latent Tracking Approach

In this thesis we present a new method to detect pedestrian in video sequences. Unlike most of the common detection methods which only rely on the appearance of an object for detection, our proposed method uses the motion information of the object as well as its appearance to improve the detection quality. The idea is to capture the motion of each body part and use these motion patterns and add them to an appearance based detector to improve the detection results. We try to model the distinct motion patterns of pedestrians using a set of latent variables. The proposed method is tested on two separate datasets and the quantitative results are outperforming two of the commonly used pedestrian detectors in the literature and showing the capacity of motion patterns to improve the detections.