Research On Feature Extraction And Object Tracking Algorithm Based On SVM

Object tracking is a well studied problem in computer vision and has manypractical applications both in military and civil field. It is a challenging task to developa robust tracking system in the complex Scenarios due to factors such as pose variation,illumination change, occlusion, and motion blur.Discriminative Methods pose the tracking problem as a binary classification task inorder to find the decision boundary for separating the target object from the background.This method is more accurate than generative way for integrating the information ofbackground. As the classifier is updated on-line with noisy and potentially misalignedexamples, this often produces a great Computational complexity and leads to thetracking problem. This paper applies the Support Vector Machine to the object trackingfield in discriminative method, I propose two efficient algorithms to solve the problemas stated above. My main work and contribution are as follows:①This paper proposes a feature selection method based on Random Projection andHaar-like feature which has been successfully applied in the field of face detection andobject detection. By projecting the feature vector from high-dimensional space tolow-dimensional space, Our appearance models preserve the structure of the imagefeature space of objects and do not lose any key information of objects, yet greatlyreducing the computational complexity.②This paper proposes a PCA-HOG based appearance model which use the PCAalgorithm to reduce the dimension of HOG feature extracted from object. Trackingobjects can be well described by this model with low computation complexity.③This paper designs a on-line learning classifier with SVM. I redesign the scorefunction to solve the problem of error accumulation and drifting. My algorithmsintegrating the discriminative method and generative method by adding the similarityinformation to the score function, and I employ a new factor to weight the classifierinformation and similarity information respectively.④This paper expands my tracking algorithm to multi-objects tracking Scenarios byParallel combination of SVM classifiers. Experiments show that my algorithm performswell in both single-object and multi-objects tracking scenarios.Algorithms proposed in this paper are based on the framework of SVT(SupportVector Tracking), the tracking results show that my algorithm are more accurate and robust than the SVT algorithm.