Research On Important Issuses Of Robust Visual Tracking

Visual tracking is a hot research topic in the field of computer vision, and serves as a indispensable aspect in intelligent video surveillance, robot navigation, intelligent traffic, national defense & security and etc. To meet practical applications, current visual tracking algorithms have to concern with many problems, such as the building of an effective appearance model to describe a target and meanwhile discriminate it from the background properly , the establishment of a fast and efficient tracking or inference algorithm to meet real-time requirements, and the association of a tracker with the right object during the course of multiple objects tracking. To solve these problems, visual tracking has to involve quite a few theoretical disciplines including image processing, pattern recognition, probability theory and mathematical statistics, optimization theory, control theory, etc. Therefore visual tracking is a research topic with high theoretical value and practical significance.In general, visual tracking algorithms can be categorized into two major groups, i.e., the deterministic method and the probabilistic method. The deterministic methods firstly build a target appearance model according to the vision characteristics of a target, then choose a proper optimization function based on the similarities between the measurement target and the candidate target, and finally use iterative optimization method to solve the local optimal solutions of the target states (including position, size, direction and etc). Such methods have the advantages of efficient computation and low dependence on parameters, but suffer from easily falling into local extremum which will lead to tracking failure. The probabilistic methods firstly predict the target's state, and then update the state through probabilistic inference with the vision observation models. One of the most popular probabilistic methods is the particle filtering based tracking method, which has the advantages of dealing with non-linear and non-Gaussian systems and providing robust tracking. But low computational efficiency and high dependence on parameters are two main disadvantages.To improve the deterministic tracking methods, the necessary work includes the building of an effective target appearance model and a corresponding similarity metric, as well as the derivation of the tracking method based on the similarity metric. Two methods are proposed as follows to fix the above two issues.(1) Aiming at the defects of the basic mean shift tracking algorithm that spatial information of the target is not integrated, which will induce inaccurate tracking and occlusions, we describe the target with multiple fragments by means of multiple histograms, and derive a similarity metric based similar mean shift iterative algorithm. Experiments on real video sequence prove that the proposed tracking algorithm has accurate tracking performance on target's size and position, and can solve the occlusion problem.(2) A differential structural similarity tracking method is proposed, in which SSIM serves as the similarity metric. The gradient with respect to the target states (position and size) is derived based on SSIM, and a tracking method is given based on steepest ascent optimization method. The proposed tracking method performs accurately and can cope with the change of target size, and is robust to the change of illumination.There are two significant issues for particle filtering based tracking methods, the improvement of sampling efficiency and the inference strategy in multiple objects tracking. These issues are addressed as follows,(1) A two hierarchical appearance model based parcel filtering tracking method is proposed. The color/intensity histogram serves as a rough model, and mean shift tracker is applied to shift each sample to its local optimal location. The eigenspace serves as a precise model, which is used to calculate weight coefficients of each sample and estimate the final state. The proposed method has fixed the problem of excessive ineffective samples. Less samples are used, and the color/intensity feature and eigenspace model are fused together to deal with rotation, intensity, and pose variations issues.(2) According to the characters of forward looking infrared imagery, a novel particle filtering method is proposed based on saliency detection and two step sampling. The target in IR image tracking is salient in most cases, so the proposed method firstly builds a saliency observation model on the saliency detection resulting images, and then an eigenspace model is employed as another observation model. The two models are combined together by a two step sampling method. The proposed tracking method presents accurate tracking results, and can deal with appearance changing and ego-motion problem incurred by the self motion of infrared sensors.(3) Aiming at multiple targets tracking issue, Markov random field is employed to build a probabilistic graphical model, and finally a kernel mean field Monte Carlo method is proposed. During the course of multiple targets tracking, occlusions between similar targets will cause wrong associations. This can be solved by mean field Monte Carlo method. In some cases, however, most samples will become ineffective due to proposal distribution inaccuracy. Fortunately, this can be solved by means of mean shift based kernel density estimation method. The proposed method has been applied to infrared image and head tracking applications.