和声搜索粒子滤波视觉跟踪

为了降低粒子滤波精度对精确重要性采样函数的依赖性，提高粒子滤波的视觉跟踪效果，将和声搜索引入到粒子滤波框架中，提出了一种基于和声搜索的粒子滤波视觉跟踪算法.通过记忆考虑、基因变异、随机变异等和声搜索算子结合当前观测信息，改善了粒子滤波视觉跟踪算法的重要性采样函数，增强了重要性采样函数对系统状态转移模型的鲁棒性.同时，对和声搜索参数进行了优化，平衡了视觉跟踪实时性和精确性的要求，并对粒子的权重进行了补偿，使其符合粒子滤波的理论基础贝叶斯估计.实验结果表明：优化的和声搜索参数，比常见参数更适合和声搜索粒子滤波；与基于粒子滤波、和声搜索、Mean-Shift改进的粒子滤波、分布场、多示例学习等视觉跟踪算法相比，和声搜索粒子滤波视觉跟踪算法能够在光线变化、遮挡等复杂场景下获得了更精确的视觉跟踪效果.和声搜索粒子滤波算法较好地结合当前观测与历史信息，获得鲁棒的视觉跟踪性能.
This paper introduces the harmony search theory to propose a novel particle filter, and a visual tracking based on harmony search particle filter which can combine the current observation with history information to achieve a robust performance. Firstly, the importance sampling function is modified using such conceptions in harmony search theory as memory consideration, genetic variation, random variation and the current observation. These improve the robustness on system state transition matrix. Secondly, parameters of harmony search are optimized to balance the demand on timeliness and accuracy. Moreover, the weight of particle is compensated to further accommodate the Bayesian estimation.Simulations show that the optimized harmony search parameters are more suitable for harmony search particle filter than common parameters. Compared with classic visual tracking algorithms, the proposed algorithm demonstrates more accurate visual tracking ability under complex environments such as illumination changing and occlusion.