The Research For Object Tracking Methods With Complex Scene Clutter

Visual object tracking plays an important and practical role in Multi-Media Application,Secure Engineering, Traffic Surveillance, Human Action Recognition as well as VirtualReality. However, because of the dynamic scene, such as occlusion, appearance orillumination change, robust object tracking has been established as a challenge task fordecades. For tackling the above issues, this paper addressed that within three popular trackingframeworks:(1) Background updating framework;(2)“detection by grouping” framework(feature point-based trackers or discriminative patch-based trackers);(3)“detection byclassification” framework (some online learning based trackers).For the Background updating framework, with a detailed analysis of three popularbackground updating model (Gaussian Mixture Model (GMM), Codebook model, AdaptiveGMM), the Adaptive GMM is selected with respect to the extraction time of background andthe completeness of foreground. In order to address the clutter scene (waving tree, vibratecamera and fountain), motion template estimation and smoothness of trajectory to and fro isrespectively utilized to clear the inference scene. Nevertheless, Adaptive GMM could not beadopted when the camera isn’t a stationary one. So this paper tends to seek an adequatetracking method under “detection by grouping” framework.In terms of the “detection by grouping” framework, through a comparison of several newfeature point descriptors, such as SIFT, SURF, BRIEF, ORB, etc., SURF is chosen as thetracking point descriptor because of it accuracy for localization. Actually, SURF descriptorhas the character of scale invariant, no limitation of point correspondence within a few frames.In order to tackling the deficiency of Mean Shift, such as poor adaption for object scale, weakperformance for object with high speed owing to the fixed window, this paper combinesSURF and Mean Shift to track object. Accordingly, the Mean Shift is constructed by severalLocal Mean Shift-s (Each one is contributed by SURF point). So this paper propose a trackingalgorithm based SURF and Local Mean Shift (It is called SURF-LMS for simplicity).Through a comparison with Mean Shift, SURF-LMS outputs a superior performance.However, if the feature points at different locations have similar likelihood, they would roamaway each other. To address the issue of “detection by grouping”, some online trackers based on the“detection by classification” framework have achieved good performance. But problems ofthese trackers are still embodied in at least one of the three aspects:(1) The global-basedsamples (target region) have poor adaptability for occlusion, appearance or illuminationchange.(2) Lack of current sample estimation, which may cause “overfitting” issue.(3) Lackof adequate motion model to prevent target from drifting. For tackling the above problems,this paper presents a novel online boosting tracker embedded with a part based structure of thetarget. The contributions of this work are threefold:(1) A novel part-based structure is utilizedin the Online AdaBoost Tracking (OAB) tracker.(2) Attentional sample weighting andselection are tackled by introducing a Weight Relaxation Factor (WRF), instead of treatingthe samples equally as traditional trackers.(3) A two-stage motion model Multiple PartsConstraint (MPC) is proposed and incorporated into the part-based structure to ensure a stabletracking. The effectiveness and efficiency of the proposed tracker is validated upon severalcomplex video sequences, as compared against three popular online trackers. Theexperimental results show that the proposed tracker can achieve increased accuracy withcomparable computational cost.