Research Of Motion Analysis For Tracking And Action Recognition Based On Sparse Model

The analysis of motion image is the basic facts and key techniques in computer vision and the importance branch of digit image processing.It provides the motion information to check out and retrieve for the users,and provides the basis for feature extraction,recognition and classification,whose result directly affect the subsequent high-level processing,such as behavior understanding,semantic description and reasoning decision-making.Recently,these technologies have been plays a great role in scientific research and engineering practice,such as military science,biomedical science,information science and meteorology,etc.This paper focuses on the three specific problems,i.e.,optical flow and occlusion computing,object tracking and action recognition.These three specific problems are important components for computer vision and intelligence analysis.However,the analysis of motion image is still not a well solved problem due to the complex of the motion and the insufficient understanding of the mechanism of sense movement of human visual system.Thanks to the stable recovery and sufficient dimensionality reduction of sparse signals,the essential,significant and key content of images/videos could be effectively preserved with sparse representation,which is helpful to reduce noise and occlusion effect.Also,sparse representation based minimization problem can avoid to set a manual threshold for classification,which has robust performance and improve the computational efficiency.Recently,sparse representation based method has been receiving a lot of interest for extracting and modeling features.This paper,firstly,computes optical flow and occlusion based on sparse model;then,develops a robust sparse model;and then,tracks objet based on another new sparse model;finally,proposes action recognition algorithm based on sparse model.(1)Lack of information in occluded regions leads to ambiguity inherent,which is a big challenge for motion estimation.Aiming at tackling the motion ambiguity efficiently,we propose to model the occlusion with sparse representations in the spatio-temporal domain and transform domain.We employ the Stein-Weiss analysis function acting as a novel regulariser and a sparsifying transform function respectively in variational and sparsity models,which is helpful to construct a unified optical flow framework for the sparse model without sparsifying transform and the sparse model with sparsifying transform.In order to deal with dictionary learning,we generate an overall dictionary directly via the sparse model without sparsifying transform,and then optimize for small size dictionaries over corresponding patches with the overall dictionary.Experiments on the Middlebury benchmark and Sintel benchmark show our method outperforms the existing estimation methods of jointing occlusion and optical flow.(2)For selecting sparse features appropriately from multiple tasks,we propose a robust multi-task feature learning model.We set the loss function as a non-smooth calibrated function to decrease the noise level of each task.The task relatedness is closely related to the regularizer.The non-convex regularizer has an appealing approximation to the original 0-type regularization,while the convex regularizer admits a globally optimal solution rather than a suboptimal solution.As such,we set the regularizer by combining non-convex and convex terms for effective leverage and compensation.Also,we develop an efficient algorithm to solve the challenging non-smooth and non-convex optimization problem.Experiments show our algorithm outperforms several state-of-the-art multi-task feature learning algorithms on both synthetic and real data sets.(3)To select sparse multi-task features effectively for visual tracking,we propose a robust multi-task feature learning model which is capable of both calibration and identification.We set the loss function as a non-smooth function to calibrate each task with respect to its noise level.We decompose the regularized matrix into two specified structures for exploiting common sparse pattern over the relevant task and identifying the relevant and irrelevant(outlier)tasks,respectively.Also,we develop an efficient algorithm to solve the challenging non-smooth optimization problem.Empirical evaluations demonstrate that our method has better performance than a number of the state-of-the-art trackers on available public image sequences.(4)We use lasso(least absolute shrinkage and selection operator)to develop a simple action recognition algorithm,which is formulated as a minimization problem with maximum average correlation height filter.Experiments on the real data sets demonstrate our algorithm achieves good performance.