Video Tracking In Dynamic Background

Visual tracking, especially in dynamic scenes, receives much attention because of its wide application. This technology is in ongoing progress in recent years, and lots of robust algorithms have been proposed for stationary scenes, but many challenges in a dynamic environment are still in the solutions.The circumstances in videos are always very complex, and illumination variation, occlusion, deformation and rotation make feature extracted difficultly and lead to tracking failure. This thesis gives a deep survey of visual tracking algorithm in dynamic scenes based on studying recent tracking algorithms. The main contents of this dissertation are listed as follows:1. An object tracking algorithm based on Kalman filter using Scale Invariant Feature Transform and Corrected Background-Weighted Histogram feature is proposed to track object in dynamic scenes. Kalman filter is used to predict object's possible position, extracting SIFT keypoints in the area of that predicted position afterwards, and match them to the keypoints in the first frame and in the previous frame. Each matched keypoints is cast votes for the candidate object centers. Using CBWH feature obtains another candidate object center. The weighted center is sent to correct Kalman filter's predicted position, and current object area is obtained. Experimental results demonstrate that the proposed tracking algorithm can track object accurately, and it is robust to occlusion and view changing.2. Taking advantages of covariance matrix's robustness to appearance changing and particle filter's ability of handling non-gaussian, we proposed an object tracking algorithm combining covariance matrix and particle filter. Particle resampling is based on the result of LK tracker, and the Perceptual Hash sequence of every particle is computed. Get rid of those whose distance to Perceptual Hash model is large. For each of the rest particles, compute their holistic covariance, block-division covariance and background covariance, and their corresponding subspace reconstruction error and weight, the region with highest weight is set as the current object. Experimental results demonstrate that the proposed tracking algorithm obtains accurate results, especially for image sequences with background clutter and motion blur.3. Temporal robustness evaluation, spatial robustness evaluation and one-pass-evaluation are performed on 2 proposed algorithms with some other excellent algorithms. Experimental results show the proposed algorithms can both track object accurately, the first proposed algorithm does well in out-of-plane rotation, deformation and occlusion, while the second proposed algorithm runs well under motion blur,fast motion,in-plane rotation and background clutter.