Robust Real-time Object Tracking Algorithm Based On Dictionary Learning

In recent decades, as the development of the country’s economic, science andtechnology, China had a high-speed pace of development. In many respects, for instance,computer vision, artificial intelligence, communication technology and multimediatechnology have a rapid development. As we all know, technology of computer vision hasbeen one of the most popular science and technology. In the video analysis field,targettracking is the most important field of computer vision and has been large applied. Targettracking is a multi-disciplinary and frontier technology and merge of pattern recognition,image processing, artificial intelligence. For many years’ development, target trackingtechnology is used in many field, for example, military, transportation security, intelligenttransportation, meteorological analysis, video coding, visual navigation and so on. At thesame time, target tracking technology obtain a large number of practical applications.The most difficult of video image analysis techniques is target tracking, it’s the coreof the video analysis system and plays a decisive role in intelligent monitoring system.However, the subject still exists same problems and difficulties to resolve in practicalapplications. In the complex and dynamic scenes, target could be affected by thesurrounding environment. In this situation, object will deform and bring the difficulty ofdetecting and tracking. Tracking algorithm based on image feature matching extractscale-invariant feature can obtain better tracking results, but it will take a lot of time.Through extract multi-scale feature(translation, rotation) to build a redundant dictionary,this paper presents a robust and real-time target tracking algorithm. Then using theredundant dictionary gets the feature of the tracking object and finding the nearestneighbor of the target template. This algorithm improves the robustness of the algorithm.The main contents and work is as follows:1. With the current status of research, summarized the existing target detection andtracking technology and analyzed of the various advantages and disadvantages ofdifferent algorithms. Typical target tracking algorithm has two categories: basedon motion analysis tracking algorithm and based on image template matchingtracking algorithm. Optical flow method and background modeling method belong to the based on motion analysis tracking algorithm. However this methodcouldn’t resolve object long-term keep stationary state. Based on image templatematching tracking algorithm compare and analyze target feature information totrack object. The result of feature extraction will affect the performance of thewhole algorithm.2. Contrary to the traditional image matching tracking algorithm, this paper presentsa tracking algorithm which use a set of image features to build redundantdictionary. This algorithm consists of classification and learning. The purpose ofclassification is to class the image which most possibly contains target. First of all,extracting multi-scale feature(translation, rotation). Then, searching the redundantdictionary to get target feature information. The purpose of learning is to updateredundant dictionary and target template. First, sampling positive samples in thetarget location. Then getting the scale, rotation, affine transformation of thepositive samples to build training samples. Extracting haar-like features oftraining samples and clustering this features to build new redundant dictionary.3. Realize the algorithm on the embedded system., the platform consists of hardwareplatforms and PC software platform. This paper selects the digital camera withCamera Link interface. The interface is a protocol for the digital camera and thestandard image data transmission which the highest transmission rates up to2.38Gbps. First, FPGA receives the high-speed and real-time image signal andtransfer to the DSP. Then, DSP processes the data for tracking target. Throuth PCI,tracking results are transmitted to the computer and showed on it.In this paper, target tracking algorithm has been detailed studied. A improvedtracking algorithm has been proposed in the paper. This tracking algorithm realize on anembedded platform and get a good tracking results.