| Moving target tracking in video sequences, is the research hotspot and difficulty in the field of computer vision. The main task is to find the object in each frame, then extract the location information and mark it in the original image. It is widely applied to the military, intelligent monitoring, motion analysis and the weather forecast. However, most existing techniques are constrained by the specific application scenarios, and there is no an algorithm that can handle all problems in process of tracking. Robustness, real-time and accuracy are the main difficulties of target tracking, hence the tracking algorithm should be further researched.Through the analysis and study of the common problems in video sequences such as partial and severe occlusion, illumination change, Shape deformation, background clutter, target rotation and similar background, the main research results are as follows:1. A method based on sparse coding and multiple instance learning is proposed. It calculates the weights of all gaussian models to represent the importance of each model with coding, from which some good models can be selected to predict the target location, so that the selected models are typical and the predicted location is accurate. We determine whether the target is occluded according to the prediction accuracy of all models to labeled samples, then decide whether to update the gaussian models. The algorithm effectively solves the target loss caused by occlusion and shape deformation, which can achieve accurate and reliable tracking.2. A video tracking method based on superpixel with inter-frame constrained coding is proposed. The method presents a new inter-frame constrained coding based on superpixel model, which considers the interaction of corresponding superpixels between the adjacent frames and maintains the space consistency of video images, make the coding more stable. Due to the update of codebook and classifier parameter, the proposed method is robust for long-term object tracking, which alleviates the drift phenomena in the process of tracking. It can accurately track the object in the case of background clutter and the illumination change.3. A video tracking method based on multi-constrained non-negative coding(MCNC) is proposed. The multi-constrained non-negative coding based on superpixel is achieved, in which the manifold geometry of the local feature space from two adjacent frames and neighboring superpixels are incorporated. The proposed method enhances the stability of coding and makes the tracker more robust for object tracking. This coding method is able to describe the small misalignments or partial occlusions as an unlikely event other than an impossible event, which weakens the oversensitivity to spatial structure, which makes the tracking performance robust in case of illumination change, similar background and rapid movement. |
