Research Of Compressive Tracking Based On Local Sparse Features

The video object tracking technique is an important and hot research content in the field of computer vision, and it is also the key application technique in the aspect of intelligent security and protection, intelligent transportation, behavior analysis and human-computer interaction. Although the computer vision had a rapid development in recent years, and a large number of excellent object tracking algorithms were proposed by many domestic and international scholars, but the object tracking technique is still a challenging subject due to the influences of illumination change, occlusion, background interference and rotation.Compressive sensing theory is a new signal processing theory proposed by Donoho et al. More and more researchers pay more attention to the theory since the published of compressive sensing. The theory was widely used in the signal processing like image processing. And good results have been achieved in the aspect of object tracking for the past few years. Using the theory and the frame of traditional compressive tracking method, we proposed a robust compressive tracking based on local sparse features. The tracking algorithm we proposed is called ICT (Improved Compressive Tracking). In this paper, the main research contents and innovation points are as follows:(1) In order to solve the object drift problems which the traditional compressive tracking method had met in the complex environment like illumination change, motion blur and rotation, we proposed an algorithm named a robust compressive tracking based on local spars features. In the algorithm, we use the SURF for the feature extraction of the target and background. Practical application proved that an effective appearance model is the important precondition for success of the object tracking algorithm. According to the appearance model, the object tracking method can be divided into two categories which called generated and discriminant tracking algorithm. Under the model of generated method, we compute the minimum reconstruction error to obtain the target location. On the other hand, the object tracking task is converted into a binary classification problem in the model of discriminant.(2) The feature extraction always causes a problem of high-dimensional which caused expensive calculations. In this paper, we compress the high-dimensional feature space to low-dimensional space on the basis of compressive sensing theory which can effectively reduce the feature dimension while retain the information of the original high-dimensional features. Meanwhile we proposed a coarse-fine strategy to decreases the computation cost. All of these ensure the algorithm has a better tracking result and satisfies the real time request. Next we classify the positive and negative samples according to naive Bayes classifier with online update. So, the method we proposed belongs to the discriminant model of the object tracking.(3) In this paper, we choose 3 mainstream object tracking methods and discuss their basic idea and realization method from generated and discriminant model. In the last of paper, we compared the 3 tracking methods and ours. And the experimental results based on different test video sequence show that the algorithm can effectively and robustly realize the object tracking task in the scene of illumination change, rotation and motion blurring. With the configuration of the computer, the tracking algorithm we proposed can reach 31 frames per second which satisfies the real time requirements in the operating environment.