Research Of Object Detection And Tracking Based On Sparse Representation And Compressed Sensing

Object detection and tracking in dynamic circumstance are the important topics of computer vision. They are widely used in civilian and military applications such as monitoring, aviation guidance, human-computer interaction, traffic monitoring and other fields. But the object detection and tracking are still faced with many difficulties, such as illumination change, clutter background, appearance change, occlusion and so on. Detecting and tracking object in complex background is still a challenging task.Sparse representation and compressed sensing have attracted much attention in recent years. And sparse representation in image applications is a hotspot. Base on the sparse representation and compressed sensing research, the main contributions of this dissertation are as follows.Firstly, an object detection algorithm based on hierarchical model of background subtraction is proposed. The sparse representation theory is addressed, especially in the dynamic group sparse (DGS) and its reconstruction algorithm. We adopt a dynamic group sparsity reconstruction algorithm with adaptive sparse degree. Then the background subtraction algorithm based on adaptive DGS is proposed. Then an online dictionary learning algorithm is used in background update scheme. And the accumulation of training frames is used for updating background model with online dictionary learning. In order to reduce the computational complexity, a hierarchical model of object detection algorithm is proposed. The proposed algorithm can reduce computational cost and work well in dynamic scenes.Secondly, a visual tracking algorithm is proposed based on compressive sensing MCMC sampling. In this algorithm, random measurement matrix is used to get compressed feature vector which can represent samples discriminatly. Markov Chain Monte Carlo method is used for the proposal distribution of particles, which avoids the degeneration of particles. The scores of bayesian classifier are integrated into Markov Chain Monte Carlo acceptance mechanism which differs from traditional particle filter sampling. In order to alleviate the drift problem, a two stage tracking scheme is proposed which combined with the initial model and online updated model. Extensive experiments on challenging sequences have demonstrated that the proposed tracking algorithm outperforms the state-of-the-art algorithms in terms of accuracy, robustness, and speed.