时空上下文学习长时目标跟踪

Long-Term Visual Tracking Based on Spatio-Temporal Context

针对在线学习跟踪算法中目标模型更新错误而导致跟踪漂移的问题,提出了一种简单但高效的解决方案。在目标区域均匀采样点跟踪器,基于纹理描述对前后两帧点跟踪器进行置信度评估并以此完成目标初步定位,由多维特征时空上下文模型输出目标位置置信图以完成目标精确定位,同时结合置信图决定模型更新速率并给出了一种多尺度更新机制。实验表明,该方法在背景干扰、快速运动、遮挡、光照变化及尺度变化下均能完成稳健跟踪,在320 pixel×240 pixel的视频序列中平均跟踪速度为55.1 frame/s,可以满足实时应用的需求。

Aiming at the tracking drift problem due to object model update falsely in the online learning tracking algorithms, a simple but efficient solution is proposed. In the target area point trackers are uniformly sampled, which are assessed based on texture description in two consecutive frames point trackers and then the initial location of target is completed. Multi-dimensional feature spatio-temporal context model is used to output the precise position of object by the confidence map, the model update rate is decided combining with the confidence map and a multi- scale update mechanism is proposed. Experimental results show that the proposed algorithm can complete the robust tracking under the condition of background interference, fast motion, occlusion, illumination changing and scale changing. In the video sequence of 320 pixel×240 pixel, the average tracking speed can keep in 55.1 frame/s, which meets real-time application requirement.