Object Tracking Algorithm Based On Sparse Representation Prototypes

In recent years, object tracking technology has obtained extensive and in-depth research anddevelopment in the domestic and overseas, and it also has broad application prospects in many areas.Currently, object tracking is still facing many difficulties and challenges, including the object shapechanges, background interference, illumination changes, occlusion, etc. At the same time, thetracking process needs to meet the requirements of accuracy and real-time. Generally speaking, atypical video tracking system in the particle filter framework consists of the following components:object initialization, the input video image, the appearance model, the dynamic model, theobservation model, object localization and the template updating model. Object tracking algorithmproposed in this paper are built in the particle filter framework, the main work is as follows:(1) A tracking algorithm based on improved sparse representation prototypes is studied. Theadvantage of this model is the ability to quickly perform iterative solution to improve the trackingspeed. Based on this, it improves the dynamic model of the object through introducing several stateparameters to achieve an effective tracking in the case of object size change and rotation change. Itdescribes the components of this algorithm in detail, including dynamic model, particle structure,observation model, template update. Experiments verify the effectiveness of the algorithm.(2) Atracking algorithm based on incremental2D subspace learning is studied. This algorithmdirectly extracts the two-dimensional image of the object, and then puts the image into the trainingset. It obtains the mean, the eigenvalues, the row-projected eigenvectors and the column-projectedeigenvectors of training set through incremental2D-PCA algorithm. Experiments verify theeffectiveness of the algorithm, but in the situation of severe occlusion, this algorithm has a seriousflaw.(3) Two tracking algorithms based on sparse subspace representation prototypes are studied.The first one combines incremental one-dimensional subspace learning with sparse representation,while the second one combines incremental two-dimensional subspace learning with sparserepresentation. It uses sub-sampling to filter the particles to improve the tracking speed. Beforeadding the tracking result into the training set, it adopts occlusion detection method to estimate.Experiments demonstrate the effectiveness of these two algorithms.