Single Body Tracking Method Research

Pedestrian tracking plays an important role in the computer vision field, and has been widely concerned in recent years. Although many promising methods have been proposed, target tracking still remains challenging due to cluttered background, illumination mutation, target appearance change, occlusion, and the appearance similarity between the target and backgrounds. In this dissertation, to handle the above tracking failures, two tracking methods based on particle filter have been proposed for tracking a single target, the main contributions of this dissertation are as follows:(1) To handle tracking problems under occlusion, out of field of view and illumination mutation, a pedestrian tracking method based on patch-based features and spatial-temporal similarity measurement has been proposed. Inspired by the observation that the different parts in one pedestrian have different stable properties, I represent the target at patch level, and design a robust similarity measurement which simultaneously considers the spatial and temporal appearance information of the target. Besides, considering that tracking drift emerges due to illumination mutation between non-shaded and shaded regions, a shadow detection and removal algorithm is introduced to eliminate the effect of shadow on target tracking.(2) Aiming at the limitation that the target features extracted by traditional tracking algorithms have low robustness, a pedestrian tracking method based on deep learning has been proposed. Based on deep learning, the method combines the detection from deep learning and the result from LK, and uses PN learning to revise the parameters of the multi auto-encoder. It avoids the target updating error caused by the detection error from deep learning, and improves the robustness of the deep network.Through several tests, it is demonstrated that two proposed tracking methods can track the target under occlusions, out of the field of view, and the appearance similarity between the target and backgrounds, and outperform most state-of-the-art trackers.