Superpixel tracking with color and texture features

• Main Authors: Dung-Han Yang, 楊東翰
• Other Authors: David Lin
• Format: Others
• Published: 2016
• Online Access: http://ndltd.ncl.edu.tw/handle/05970153127179645628

碩士 === 國立東華大學 === 資訊工程學系 === 104 === In recent years, there has been a significant growth of visual tracking in the related applications such as traffic monitoring, intelligent surveillance, and human-computer interaction. Many effective and efficient tracking algorithms were proposed to face various challenges, including large variation of scale, heavy occlusion and drifts. In this thesis, we apply the Bayesian theory and propose a visual tracking method using superpixel. This tracking algorithm contains three models: (1) appearance model, (2) motion model, and (3) online learning model. Our method has improved the previous superpixel tracking (SPT) method by adding texture feature to reduce the interference of neighboring similar color and increase the success rate of grayscale sequences. Three kinds of confidence maps (HSI color, LBP texture and color+texture combination) are generated in appearance model to separate the target from the background. We compute the probability of these three confidence maps separately, and then sum up their final results to obtain the most possible object position of the next frame. After that, the motion model is used to weigh the appearance model. Finally in the online learning model, we utilize the k-means clustering algorithm to characterize the features, and update the training set every ten frames to avoid tracking drift. We conducted the experiments using 27 videos from literature, and most of the sequences achieved better results than previous work. The system shows an improvement on the tracking drift caused by background clutter, occlusion, grayscale tracking, scale and pose variation. These results prove that the SPT-LBP is a robust tracking method for both color and grayscale images.