Study On Image Classification And Target Tracking Based On Sparse Representation

Sparse representation is a kind of data representation method which similar tohuman cortex, it is a fundamental theory for many signal processing problems, such ascomputer vision, audio signal analysis, and blind signal processing. Among manyresearch fields, the computer vision is well known for its tight connection to sparserepresentation, and it has drawn a lot of interests of researchers recently.Image classification and target tracking are branches of computer vision analysis;they have been used to construct many systems such as human-computer interface,intelligent transpotation, UAV navigation, intelligent security and shown greatimportance of practical value. Sparse representation is a newly devoleped signalrepresentation theory, it simulates data representation function of manmal cortex. Inrecent years, it has been wildly used in computer vision analysis applications. Theadvantage of sparse representation theory comparing to common methods for signalprocess is that it is more robust to errors and noises. This thesis mainly focuses on thetheory of sparse representation and its applications such as sparse representationclassifier，image classification and target tracking. Related work of other researchers isintroduced, based on those research results, further investigation are also carried out.The purpose of researching sparse representation theory is introduced; relatedtheories and applications are summarized; mathematical models of sparse representationfor signals, the methods for measuring sparsity and sparse representation principles forimage classification and target tracking are discussed in this thesis. Then, the existenceof optimal solution of sparse representation is proved; greedy methods and convexrelaxation methods are introduced. Furthermore, coordinate decent for solving Lassoproblems are introduced as well; a kernel non-negative sparse representationclassification method and a discriminative analysis based sparse representation methodare proposed, those two methods could be the fundamental basis for many sparserepresentation problems.The sparse representation based scene image classification method is an importantpart of this thesis. The specialty of scene images and a number of difficulties areintroduced. Frameworks of common scene classification methods, their merit anddemerit are discussed as well. Based on these discussions, a hierarchical kernel sparserepresentation classification and multi-scale block rotation-extension (HKSRC-MSBRE) based robust image recognition method is proposed to deal with the randompermutations and combinations of local images in image recognition tasks. At first, themulti-scale grids are used to cut the training image into pieces, and thenrotation-extended methods are applied to create a dictionary which is adapted to randompermutations and combinations of local images in test sets. To enhance the sparsity ofthe dictionary and increase the efficiency of the system, a new strategy is proposed toreduce the dimensions of the dictionary; then, a kernel random coordinate descentmethod is proposed to solve the convex optimization problem in the KSRC; at last, themethod for calculating the class label of each image is proposed. The experimentalresults show the proposed method has robust performances when dealing with randompermutations and combinations of local images, and it has outperformed otherstate-of-art image recognition methods.The sparse representation based target tracking method is an important part of thisthesis as well. In order to fully explore this problem, many tough situations arediscussed: such as target and its background sharing similar patterns, deformations andocclusions. Common target tracking framework such as Kalman filter and Particle filterare studied. By casting the tracking problem as a kind of sparse approximation problem,a novel algorithm based on STC and SRC is proposed for object tracking in complicatedscenes. Space constraint is given to ensure background patterns in the template havesmall weights, time constraint is given to handle the target appearance variances. Sparserepresentation classification method for tracking have been proposed to calculate lossvalues and experimental results showed that time-space constraints and sparserepresentation classification method have better results than standard l1-loss method.Proposed tracking algorithm outperforms better than other state-of-art trackingalgorithms in many difficult situations.At last, in order to solve tough problems such as varying lighting conditions,motion blurring and tracking failure in long-term tracking, a kernel parallel sparserepresentation classification and sparse classifier girds based cooperative target trackingmethod is proposed in this thesis. This tracking method is under normalhedgeframework, it combines a few novel technics such as kernel parallel sparserepresentation classification, automatic dictionary updating and sparse classifier grids,and those technics ensure the reliabilities for long term tracking. Unlike other computervision applications, tracking is a real-time task, and its algorithm must be highlyefficient. The proposed parallel method could fully utilize the multi-core CPU, thus solve the efficiency problem. Similar to other sparse representation based trackingmethods, classification confidence values equal to loss values; furthermore, in order toavoid drifting problems, the online updating method of target dictionary andbackground dictionary is proposed as well. The sparse classifier grids are designed toroughly detect the target after it reappears and reboot the main tracker. Experimentalresults show each part of proposed method could enhance the tracking performance,comparing to other state-of-art methods, the proposed method offers better real-timeperformance and reliability.