多特征自适应均值漂移算法的目标跟踪

Human tracking based on adaptive multi-feature mean-shift algorithm

针对传统均值漂移(mean-shift)算法存在对目标特征描述不完整、目标模型不能动态更新、无法解决目标遮挡等问题,本文提出多特征自适应均值漂移算法的目标跟踪。首先利用人体躯干侧影改进模型核函数,采用目标颜色特征与纹理特征建立目标直方图模型,提高算法对目标描述能力;提出选择性模型更新策略,自适应地调整目标模型,改善了传统整体更新策略由于过度更新导致的跟踪发散;最后利用扩展卡尔曼滤波(EKF,extend Kalman filter)提取目标运动特征确定目标位置。与传统算法相比,本文所提算法能在背景干扰条件下准确跟踪目标;同时,图像处理平均速度达140frame/s,满足实时性要求。实验结果表明,本文算法可以实时准确地跟踪目标,对环境干扰、目标遮挡具有鲁棒性。

The mean-shift algorithm is generally used in tracking system,which is based on non-parametric density estimation method and produces less computational loads.The traditional mean-shift algorithm cannot effectively handle inadequate representation of target′s color distribution,invariable target model and occlusion.Aiming to resolve these problems,an adaptive multi-feature mean-shift algorithm is proposed.In this algorithm,an improved kernel function is obtained by combining the silhouette of the target.This method could eliminate the influence of the background,various illuminations and other related limitations.Furthermore,the color-texture histogram is extracted to achieve target model effectively.A selective update strategy is proposed to adaptively update the target model.The update scheme could not only eliminate the accumulated error,but also avoid drifting away.At last,extend Kalman filter（EKF）is used to estimate the person′s motion between consecutive frames.Compared with the traditional algorithm,the presented algorithm could track the target successfully when the background has similar color and texture.In addition,the average tracking time is 140ms/frame,which satisfied the requirements of real-time target tracking.Experimental results show that the proposed algorithm is robust against similarity distraction and occlusion.