Abstract: A pedestrian detection algorithm based on mechanism of dynamic cost-sensitive parameters optimization is proposed. This algorithm introduces a cost-sensitive SVM classification algorithm by setting cost-sensitive parameters to process the problem of class-imbalanced between the sample size of pedestrian and non-pedestrian in an image. Taking into account the selection of cost-sensitive parameters’ values has a great influence on the detection performance, the paper proposes a chaotic particle swarm optimization algorithm based on T mutation that can improve the global search ability of particles by chaos algorithm and T mutation function, and using the best compromise between positive and negative samples correctly classified as the principle of optimization to optimize dynamically in the range of cost-sensitive weight value. Experimental results show that the pedestrian detection algorithm based on cost-sensitive parameters dynamic optimization mechanism is conducive to improve the detection accuracy.

Key words: computer vision, pedestrian detection, cost-sensitive, Particle Swarm Optimization（PSO）, T mutation

摘要： 提出一种基于动态代价敏感参数寻优机制的行人检测算法。该算法引入代价敏感的支持向量机分类算法，通过设置代价敏感参数处理图像中行人与非行人样本数量间的非均衡问题。考虑到代价敏感参数值的选择对检测性能影响很大，提出一种基于T变异的混沌粒子群算法，同时融入混沌算法及T变异函数提高粒子的全局搜索能力，并以正负样本正确分类的最佳折中作为寻优原则，在代价敏感权重值的取值区域内对参数进行动态寻优。实验结果证明，代价敏感参数动态寻优机制的行人检测算法有利于提高检测精度。

关键词: 计算机视觉, 行人检测, 代价敏感, 粒子群, T变异