Research And Analysis Of Object Tracking Algorithm

Object Tracking has played an important role in computer vis ion, navigation and Intelligent Traffic system(ITS). It's also a challenge because of the variety of scenery, complexity and real-time processing ability. The thesis focus on object tracking, analyze and conclude the classic tracking algorithm and propose some problem and future work.The thesis has a in-depth study of Tracking-Learning-Detection algorithm and Compressive Tracking algorithm. TLD realizes long term tracking by applying an improved on-line learning mechanism to update tracking feature points and detection model and parameters. CT is based on compress and cognitive theory. It reduces the dimension of image by random cognitive matrix which satisfies the RIP condition. However, both the two algorithm has its own insuffic iency.TLD lost track easily when targets undergo abrupt scale and appearance change. It costs large memory and has high complexity. CT has low complexity and mote tolerant than TLD in occlusion, appearance change and noise, but it's not robust to scale change.This thesis has proposed a multip le target tracking framework based on Gaussian Mixture Model and TLD mechanism. GMM is used to detect moving object for track initialization. TLD and multip le target association is applied for long-term and stable multip le target tracking. It is proved that this method is very effective and can be popularized.An improved algorithm based on compressive tracking is proposed in this thesis. Multi-scale and cascade classifier is adopted to choose the best scale and position. Experiment results ind icate this algorithm not only performs well with challenge scenes such as pose variation ? abrupt illumination?multi-target interfere?motion blur, but also prevail in scale variant sequences.