Abstract:

In real-time target tracking of video sequence, aiming at the problem that the classical Mean Shift Tracking（MST） method can’t deal with occlusion and scale change, a tracking method combining MST, self-learning detector and particle filter is proposed. Firstly, the MST algorithm is used to track the object in the video frame, and the object is re-initialized when the object converges to the local minimum. Then, a detector based on online learning is proposed to update the object model of MST adaptively, so that it can automatically adjust the object scale. When complete occlusion occurs, the particle filter is activated to estimate the object position by probability calculation, so that MST can recover tracking when the object leaves occlusion. The experimental results on PETS video sequence datasets show that compared with several existing MST methods, this method has high tracking accuracy and can be used in real-time detection and object tracking applications.

Key words: video sequence, object tracking, mean shift tracking, scale estimation, detector, particle filtering

摘要：

在视频序列的实时目标跟踪中，针对经典均值漂移跟踪（MST）方法不能应对遮挡、尺度变化等问题，提出一种结合MST、自学习尺度探测器和粒子滤波的跟踪方法。采用MST算法在视频帧中跟踪目标，当目标收敛到局部最小值时重新初始化目标。提出一种基于在线学习的探测器，用来自适应更新MST的目标模型，使其能够自动调整目标尺度。另外，当出现完全遮挡时，启动粒子滤波器，通过概率计算来估计目标位置，使MST能够在目标离开遮挡时恢复跟踪。在通用数据集PETS视频序列上的实验结果表明，相比其他几种较新的MST方法，提出的方法具有更高的跟踪准确性，可以应用于实时检测和目标跟踪等应用中。

关键词: 视频序列, 目标跟踪, 均值漂移跟踪, 尺度估计, 探测器, 粒子滤波