基于Harris的遗传粒子滤波及其在车牌跟踪的应用

Genetic particle filter via Harris and its application in license plate tracking

为了有效解决传统粒子滤波算法所存在的种群多样性衰减问题,消除由此而带来的算法效率、精度下降的弊端,该研究提出利用遗传算法的交叉和变异遗传操作算子来优化其重采样过程。具体而言,在重采样后,对样本集中各个样本粒子依照适应度值排列顺序,再将适应度低于平均值的样本剔除,同时从留下的适应度较优的粒子中随机地选取同等数量样本用于对应补充被剔除样本,再引入遗传算法的遗传操作对粒子进行交叉、变异操作,来完成样本集的更新。同时考虑到传统视觉目标跟踪常用的灰度和颜色直方图特征极易受到背景颜色干扰、对光照变化极为敏感和计算量也较大等问题,提出引入具有容易提取、运算量小、抗旋转或倾斜角影响等优势的Harris特征,配合遗传粒子滤波跟踪框架,得到了一种鲁棒性较高的跟踪算法。将所提出的基于Harris特征的遗传粒子滤波跟踪器应用于高速公路上的车辆车牌定位,应用实验的结果表明经过遗传操作改进的使用Harris角点检测特征的粒子滤波算法精度、数值稳定性都得到了提高,在目标快速移动、光线和背景剧烈变化等场景都能够实现对目标车牌的有效跟踪。

To solve the problem of population diversity attenuation,low efficiency and accuracy of traditional particle filter,the genetic operation is applied to optimize its resampling process.Specifically,after resampling,each particle is arranged by the fitness value,and then the samples with fitness lower than the average value are replaced by the samples randomly selected from the particles with better fitness,and then the genetic operation is introduced to cross and mutate the particles to complete the update of the sample set.Meanwhile,the gray and color histogram features commonly used in traditional visual target tracking are very vulnerable to background color interference,very sensitive to illumination changes and has a large amount of calculation.So Harris features with the characteristics of easy extraction,small amount of calculation,anti rotation or inclination angle influence,strong anti-interference,uniform distribution,accurate positioning and high stability are introduced to cooperate with the genetic particle filter tracking framework,and a visual tracking algorithm with high robustness is obtained.The proposed genetic particle filter tracker based on Harris features is applied to the highway vehicle license plate tracking experiment.The application results prove that the method has high precision and high numerical stability,and can complete the precise tracking of the vehicle license plate in the case of rapid target movement,sharp changes in light and background.