Study Of Object Tracking With Online Methods Based On Video

Object tracking is an important issuein the field of computer vision, and is widely used in human-computer interaction, intelligent video monitoring, intelligent transportation and so forth. Appearance model is thecore part of object tracking framework, among which the ability of online learning is becoming a hot research area because the object appearances are always unpredictable. Although great progress has been made in the research of object tracking in recent years, it is still difficult to track an arbitrary object in real environments, due to influencesby complex deformations of the object, similar interferencesin the scene, unknowns of tracking reliability and so on. Researchers are trying to solve these problems from feature design of object’s appearance, representation of object, updating mechanism ofappearance model and fusion ofmulti-cues in the tracking framework. This dissertation belongs to this kind of research.This dissertation focuses on two parts of work:1) Hand, as an important interaction "device", is treated as the tracking object in the research, which is difficult to be tracked because oftoomany similar disturbances in the scene. The robust hand tracking algorithm in complex environment is studied from the appearance modeling by consideration ofobject’s characteristics, and comprehensive utilization ofbackground and motion information.2) Objects with rich features are still hard to be tracked because of its complex deformation. For objects belonging to this kind, the robust object appearance modeling algorithm is studied by imitating the shape perceptive mechanism in the ventral way of human primary visual cortex. The main contributions of the dissertation are summarized as follows:1. Online object tracking algorithms appeared in recent literature are studied and compared. Six typicalalgorithms areimplemented and explored. The testing videos areselected, containingcomplex deformation of the target, illumination with dramatic changes, similar disturbances in the scene and severe occlusion of the target. The comparative analysis of the six algorithms is conductedbased on qualitative tracking results and quantitative experimental data. The problemsof existing online tracking algorithms are summarized, and the research direction that may solve the problems is pointed out.2. By taking hand as tracking target, the tracking algorithm based on incremental subspace learning with fusion of multiple information clues is proposed:1) In consideration of hand color clustering character, incremental subspace learning algorithm is extendedto color images, and an enhanced incremental color subspace learning algorithm usingintensity and color informationis proposed;2) An online updating background model and a motion trajectory model are established, and two methods are presented for judgmentof whether the tracking targetis lost. The experimental results validatethe effectiveness of these algorithms.3. To enhancethe anti-interference ability of the object tracking algorithm, the tracking model is improved from the complementary use of different kinds of features and combination of detections oftarget motion trends:1) Considering the complementary role of holistic and local features, a new tracking algorithm combined with online fern classification based on incremental subspace learning is proposed, in order to improve the robustness of the algorithm, which adoptsfern classification to select good particles from those determined by incremental subspace model and particle filter;2) A new tracking algorithm combined with incremental subspace learning and optical flow motion detection of LK is proposed, which restrains the motion model of particle filter depending on the number and distribution of reliable optical flow points. The experimental results show that the anti-interference performance of the proposed algorithm and the computingefficiency of the incremental subspace learning algorithm have been improved.4. Objects with rich features, which generate difficultiesfor tracking because of complex deformations, are explored in this thesis. A novel online appearance modeling algorithm is proposed based on cognitive patches by imitating the perception mechanism of ventral pathway in visual cortex. The cognitive patches can be used to calculate C2features, which is invariant to position and scale, and can distinguish complex shapes. Experimental results indicate that the proposed model is robust and effective to complex appearance deformation and severe occlusion of object. In addition, the tracking result of the proposed model canbe easily driftedintracking target with symmetrical appearance because of the global location invarianceof C2feature.To overcome this shortcoming, a modified model combined with holistic feature perception of the target is proposed, which achieves satisfactory experimental results for face tracking.The studiesof this dissertation wouldbe helpful for the realization of human computer interaction via hand, the important "device", in practical applicationsas well asthe employmentand exploration of visual perception theory in visual tracing.