Research On Multi-feature Fusion Target Tracking Algorithm Based On Particle Filter In Complex Environment

Nowadays, video object tracking technology is greatly valued in both theoretical research and practical application for its importance in national defense as well as daily security system. Currently. Target tracking theory in simple environment is relatively mature, and successfully applied in medical diagnostics and other aspects. In practice, target tracking environment is more complex, the change of light shadows as well as the occlusion of similar interference and self-deformation that caused by projection from three-dimensional objects into two-dimensional deformation can all greatly affect the algorithm, There isn’t nevertheless a complete well-established theory in complex environment tracking under by now.On the basis of in-depth study on the particle filter algorithm,the whole subject research is carried out in three aspects as particle filter algorithm improvements,adaptive fusion tracking based on particle filter as well as tracking state judgment and processing:1. To settle the particle degradation that generated during multiple iterations, a reverse mapping sampling particle filter algorithm was proposed. The algorithm avoids the difficulty of directly sampling from the optimal density function, firstly, distribution function of the optimal density function was presented adopt numerical integration based on Evolution Strategy method. According to single-valued inverse mapping theorem of distribution function, the algorithm indirectly sample particles from the optimal density function through sampling from its distribution function, which inhibited the degradation of the particles.2. In order to solve the problem that single fusion algorithm can’t accurat ely characterize the target in complex environment, a characteristics measurement function was defined by the uncertainty degree of features, meanwhile, the mutual support of two features which was used to dynamically adjust the proportion of multiplicative fusion and additive fusion,was calculated by their relative entropy coefficient calculated. In this way, the proposed model achieve both adaptive weight adjustment and adaptive fusion, and the stability of this tracking algorithm in the case of light changes and angle changes, etc.3. Target tracking algorithm based on particle filter has a poor performance in target identification, to solve this, Ada Boost cascade classifier was introduced. Meanwhile, a tracking state criterion was set in the improved algorithm, thus the particles were guide to move to the classification center when the tracking confidence was at a low level. The target appeared again after longtime occlusion can be rapidly captured through the new method, and the accuracy and robustness is further enhanced.