Research Of Methods For Object Detection And Tracking Based On Sparse Representation

In the process of object detection and tracking, environment interference and object itself variation often appears, and these increase the difficulty of detection and tracking, and results in performance reduction of algorithm. This thesis studies the sparse representation theory based on the in-depth discussion of present domestic and international research status of the object detection and tracking, and applies it to occlusion-case object detection. Further more, a novel object tracking method is proposed based on robust sparse representation model. The main contents of this thesis can be listed as follows:1. The thesis first deeply explores the fundamental theory of sparse representation, describes The convex optimization algorithm for sparse representation and its application in object detection and tracking is introduced afterwards. These are theoretical preparation for subsequent research.2. For the problem that face image detection is difficult with occlusion interference, an occlusion face detection method is proposed based on sparse representation. Adding the basis vectors which can represents the occlusion to the dictionary, the test samples are represented as the linear representation of training samples and interference samples, The dictionary is trained using the K-SVD methods so that the sample can get its sparse representation of the over-complete dictionary even under strict sparsity constraints. This thesis uses the sparsity concentration index to discriminate the face or non-face images, and further facial image can be identified by coding residual. Random projection is used to reduce the dimensions of the image and computation amount.3. For the problem that the reconstruction of sparse coding has not a high accuracy, a modified sparse coding model is proposed. This model does not require the construction error following a particular prior probability density function. It adds an adaptive restraint of the coding coefficient to find the maximum likelihood estimation solution of the model. The model is more robust to interference and the get more accurate tracking results. Then a modified object template updating method is used to exclude the large interfered sample from the template library, which makes the current target templates can describe the latest object state much better. Finally, it is applied to the particle filter framework to achieve object tracking.4. This thesis selects number of partial occlusion face images and representative object track video, and the experiments are carried out by using the modified algorithm. The experimental results show that, the proposed detection algorithm is robust in dealing with the object partial occlusion, and can handle variety of occlusion types. The tracking algorithm can deal well with the corruption like occlusion and illumination variation in object tracking.