The Method Of Research On Multi-target Tracking Based On Particle Filter And Mean Shift

With the continuous development of science and technology computer vision attracted more and more attention. Based on multi-target video sequence tracking it attracted a large number of experts and scholars to participate in research.Multi-target tracking applications are very wide. Whether in a modern defense system is still in the air and maritime traffic control system also or human-computer interaction video surveillance etc. multi-target tracking technologies are indispensable. Although the study of multi-target tracking technology is becoming a hot topic in recent years domestic and foreign scholars in this field have conducted extensive research made a lot of good algorithm. But the video track in the target increase in the number of unknown status tracking and time-varying and complex clutter target scattering and dynamic change in the distance and other conditions undoubtedly increased the difficulty of tracking multiple targets. And with the single target tracking it is different multi-target tracking and measuring the state there is no direct correspondence. These are the key issues in the study of multiple target tracking urgent need to address and optimization.Based on Mean Shift tracking algorithm has high real-time characteristics but at the time of a similar background and objectives occluded the goal is easy to lose track. Mean Shift tracking algorithm target tracking algorithm based on particle filter occlusion occurs when the target can be properly track the target with a strong anti-jamming capability but it has a large computation particle prone scarcity degradation compared to other shortcomings. Taking into account the limitations of mean shift algorithm and particle filter characteristics cannot be tracked in real time many scholars take advantage of two algorithms they were effective combination made a lot of high-performance algorithms.To solve some problems that exist in multi-target tracking.This is based on previous studies, the mean shift algorithm for real-time performance and particle filter anti-jamming features, while introducing a hotspot in recent years--- the theory of random sets particle filter, an improved the multi-target tracking algorithm.Firstly, a statement of the current multi-target tracking in research and development at home and abroad, illustrating the importance and value of multi-target tracking technology research. But also made some key technical difficulties that exist in the current study. In the following pages, the detailed description of the multi-target tracking methods and basic principles. For multi-target tracking in a generic sense, it describes the traditional multi-target tracking basic methods of key points. And different from the traditional multi-target tracking based on data association, based on the idea of filtering theory itself is proposed that this method can effectively avoid data association computational complexity, and low performance. Paper describes two aspects of this new way of thinking. Multi-objective model of a random set of filter theory.The basic principles and key concepts based on Mean Shift object tracking algorithm to analyze. Mean Shift real-time and good benefits, it is also faced with the plight of poor robustness. Mean shift algorithm has less computation, good real-time characteristics. In comparison with other algorithms, it eliminates the need to step in the process of global search, by finding the best set of pixels, giving greater weight and match to determine the target coordinates. At the same time, by selecting kernel algorithm, color histogram constructed by the mathematical model of the object, the algorithm enhanced anti-jamming capability. For the selection and configuration of the target model, the key point is that the color distribution range. Tracking, once the following conditions exist: such as the environment changed greatly, strong light or dark colors, when the object is displaced and other objects in the background color histogram similarity large, all have a greater impact on the algorithm. Especially in multi-target tracking this study, the scene with the observed increase in tracking objects become more complex, and very prone to health interleaved block displacement of the object, which mean shift algorithm performance, higher requirements.Target tracking algorithm based on particle filter, this paper introduces the theory part, that the key elements of the particle filter, including the basic principles and methods of Bayesian estimation, based on particle collection Monte Carlo sampling, sequencing, sampling process and the importance of tracking particle process prone to degradation phenomenon, which is one of the more frequently encountered defect tracking. Then also it introduces particle filter specific algorithm processes and procedures. Initialization stage of the algorithm is to extract urgent target tracking feature, in this step, we want to track artificial selected target, which specifies. Then the program will calculate the characteristics of the selected target. In determining the target feature, we enter the search stage, the use of the particles on the target search. The next stage, the particles will return tracking reports, to obtain the final position of the target, its weighted average step. How to get a new one image in goal? Enters resampling stage, or the use of particle properties to search, of course, we will select the high similarity of position a little more particles, similar to the low position is a little less, then continued heavy sampling, the process cycle will be.Finally, based on the previous chapters, the specific steps to improve the multi-target tracking algorithm is described in detail. Particle filter algorithm is based on the video moving target tracking process, a new filtering method based on random set theory, the use of probability density is assumed that the target method to predict and updated. Then the introduction of Kernel Density Estimation Mean shift algorithm iterates target state estimation process is completed, the results will be the entire multi-target tracking iterative optimization. Experimental results show that the improved algorithm can effectively handle multiple target tracking in case of occlusion, improved target tracking accuracy and robustness.