A Study Of Tracking Algorithm Based On Vision

With the development of high performance computers, low cost video cameras of highquality become widely used, such that the demand in automated video analysis is greatlyincreasing. This leads to wide applications of object tracking in automated surveillance, videoindexing, human computer interaction, traffic monitoring and vehicle navigation. However,video tracking usually suffers from information loss caused by projection from3D space ontoa2D image, in addition; there are a number of factors which may deteriorate the trackingperformance, such as noise, complex object motion, non-rigid or articulated nature of objects,partial and full object occlusions, complex object shapes, scene illumination changes, andreal-time processing requirements. Because of its perspectives and challenges, object trackinghas drawn increasing attentions in last decades.Numerous approaches for object tracking have been proposed, and these methods can bedistinguished from each other by identifying how they approach the following questions:Which object representation is suitable for tracking? In what manner image features should beused? How should the motion, appearance, and shape of the object be modeled? The answersto these questions depend on the context/environment in which the tracking is performed. Themajor challenge of visual tracking can be attributed to the difficulty in handling theappearance variability of the object. Intrinsic appearance variability includes pose variationand shape deformation, whereas large appearance variation is inevitably caused by extrinsicillumination change, occlusions and so on. Handling the appearance variability properly isimportant for the stability of tracking algorithm. Therefore, the main research of this thesis isthe illumination compensation of tracking algorithm, articulated object tracking and multipleobject tracking.In the thesis, we propose some new methods to deal with the shortcomings mentionedabove. The main contributions of the thesis are listed as follows:(1) The distribution of color histogram is directly affected by environment illumination,which in turn affects the performance of the Camshift algorithm based on color histogram. Toalleviate the influence, a self-adaptive compensation method is proposed in chapter2. Thismethod can realize a self-adaptive compensation. Considering the computational complexityof the camshift target search, pyramid down and up sampling and PCA (principle componentanalysis) of color histogram are employed to locate the region of target. This method canavoid the target-searching in the whole image. It is shown that the proposed method can perform the tracking task with enhanced robustness as well as reduced computationalcomplexity.(2) To deal with the changes of the articulated object represented by subspace, anarticulated object tracking algorithm based on incremental learning is proposed in chapter3.In this algorithm, FFT(Fast Fourier Transform) and graph-cut algorithm are used to transformthe target image while the amplitude image is disposed by LBP(Local Binary Pattern). Theimages disposed by LBP are used to obtain the subspace representation by SVD(SingularValue Decomposition) and PCA. This algorithm is realized in the framework of particle filter,and it is able to reduce the location error. The method of the representation of articulate objectin this thesis is robust to the subspace representation under the appearance change of thearticulate object. Experimental results demonstrate that the algorithm is able to trackarticulated objects with higher accuracy.(3) The tracker based on color histogram and particle filter may be converged to a localregion of the object. In chapter4, an algorithm based on local discrimination and particlefilter is proposed. The local discriminations are also used to detect objects in multiple similarobjects tracking. Experiments demonstrat that this method can works well for both singleobject tracking and multiple similar objects tracking.(4) When the environment changes, the unchanged description of the object can notdescribe the object properly. Therefore a tracking algorithm with adaptive model updating ispresented in chapter5. The description of the object is composed of color histogram andhistogram of oriented gradients (CHOG). The color histogram and histogram of orientedgradients are updated by kalman filter and Gaussian model, respectively. The similar objectsare detected based on CHOG. Experiment results illustrate the efficiency of the trackingalgorithm.