Research On Online Object Tracking In Complex Scenes

Object tracking is currently one of the most popular topics in the computer vision field. After several decades of research, many successful theories and algorithms have been proposed, but tracking an object in complex scene is still a difficult problem. In this thesis, several classical tracking algorithms are investigated. And two improved algorithms are put forward on two affecting factors: illumination variations and occlusions. The main research progress is as follows:Illumination variation may result in extreme changes of the local or overall intensity of an image, thus leading to tracking failure. To solve this problem, a spatialtemporal context(STC) tracking algorithm combing with color information and retinex-enhanced gray information is proposed. The propose algorithm use multiple scale retinex(MSR) algorithm to reduce the extreme intensity change caused by illumination change. Besides, hue information is also used as a color feature to improve the accuracy and the stability of the algorithm. Experiment results indicate that this algorithm tracks well in complex scenes with illumination change and over performs the main stream algorithms.Aiming at solving the tracking problem in complex scenes with occlusions, a dual-model spatial-temporal tracking algorithm is proposed. In the proposed algorithm, an auxiliary object model based on binary pattern is introduce to evaluate the original STC's tracking result, and search object in the image when the tracking result is not valid. By introducing the auxiliary model, this algorithm can take advantage of the STC algorithm's high accuracy in short-term occlusions, as well as overcoming the wrong update problem of the STC algorithm in long-term occlusions. Experiments show that the proposed algorithm is able to track the object in complex occlusion scene, and more robust than other algorithm.