基于递推估计和上下文更新的鲁棒单目标跟踪方法

针对杂乱背景和光照变化等容易使目标跟踪产生漂移的问题,提出一种基于递推估计和上下文更新的鲁棒目标跟踪方法,该方法是颜色粒子滤波目标跟踪的有效扩展.通过建立颜色粒子滤波跟踪的通用框架,利用上下文信息分配目标外观变化的置信度,在重采样阶段,采用递推估计从其外观相似度分数计算的权重选择粒子,并初始化异常粒子.形变和光照变化的视频测试表明,该方法可以克服光照变化和背景的影响,递推估计可以处理偏离整体估计的异常粒子.相比于标准颜色粒子滤波、粒子随机搜索法等方法,该方法在跟踪框中心误差和平均重叠方面均优于其他方法,在鲁棒性和准确性方面具有明显优势.

Robust single-object tracker method based on recursive estimation and contextual information updating

As the clutter background and illumination changing can easily make target tracking drift, a robust target tracking method based on recursive estimation and contextual information updating is proposed, which is an effective extension of target tracking based on color particle filter. Firstly, the general framework of tracking based on color particle filter is established. Then, the contextual information is used to allocate the confidence degree of changes of target appearance. Finally,in the period of re-sampling, recursive estimation is used to calculate the weight to choose particles from appearance similarity score, and the abnormal particles are initialized. The proposed method can overcome the influence of illumination changes and shape changes are verified by the video test on videos of shape changing and illumination changing. And the recursive estimation is proved to be able to deal with the abnormal particles that are deviated from the total estimation.