Research Of Robustness Target Tracking Under Complex Environments

There are many object tracking applications, such as human-computer interaction, security detection, military visual guidance, etc. Due to the impact of external environment, time performance and accuracy of the algorithm and the object itself, the application scenario of many object tracking algorithms has been limited. Therefore, constructing a robust tracking algorithm is of great significance. In this thesis, object tracking algorithms are studied systematically, including both generative object tracking method and discriminative object tracking method. The main contents are as follows:1. The difficulties in the object tracking are discussed. The application scenarios and research background of generative and discriminative object tracking methods are analyzed in detail. The algorithms of the two methods are discussed with the classical algorithms and related important details described, as the foundation of this thesis for theoretical research.2. Particle filter algorithm in the presence of disturbance may have a reduction in particle diversity and decrease in accuracy. For this problem, a new IP-MCMC-PF object tracking based on constraint knowledge is proposed and implemented. Firstly, the particle forecast accuracy is improved with constraint knowledge and the particles diversity is increased through a multi-strand parallel IP-MCMC method. Thus the particle degradation is effectively solved. Then on this basis, the sampling distribution of sample particles and the training samples of the detector are updated online through the PN learning algorithm. Thus an object tracking algorithm of online learning is achieved. The accuracy and adaptability of object tracking under complex background is improved.3. For the problem of tracking restrictions in compressed sensing object tracking algorithm, an adaptive compressed sensing tracking algorithm is proposed. The proposed algorithm firstly introduces wu’s method to solve the single scale problem in object tracking. Then, use the result of classifier to decide whether to update classifier. The interference is predicted with the response of classifier and the parameters are update adaptively. Feature sorting is realized according to the quotient of conditional probability distribution between positive and negative samples. The classifier is updated with the top features and the classifier performance is effectively improved.4. The two proposed algorithms are tested with a number of tracking videos under different scenarios. The tracking data and analysis results are given in the thesis. Experimental results show that both of the two tracking algorithms are robust with the presence of multiple interferences.