基于空时线索的TLD视频跟踪算法

视频跟踪是计算机视觉领域的研究热点之一，Tracking-Learning-Detection(TLD)是近年来提出的一种有效的视频跟踪框架。针对短时遮挡以及复杂可变背景环境下的目标跟踪问题，提出了一种基于空时线索的TLD视频跟踪算法。在该算法中，采用由局部图像块的多通道特征训练生成的霍夫森林进行检测，通过多个局部图像块引入目标相关的空间位置信息，提高了算法的区分能力；然后，根据图像块对光流跟踪初始位置进行随机化布置并利用空间位置信息对光流跟踪结果进行加权，改善光流跟踪的性能；最后，对光流跟踪输出置信度与霍夫森林检测输出置信度进行自适应空时融合，综合提高目标的跟踪精度。实验结果表明，与原始TLD算法相比，本文算法能够更有效地处理遮挡问题，实现复杂背景环境下的鲁棒目标跟踪。 

TLD Visual Target Tracking Algorithm based on Spatio-Temporal Cues

Visual target tracking is one of the hot areas of research in computer vision, and Tracking-Learning-Detection, abbreviated to TLD, is an effective visual tracking framework proposed in recent years. To improve the visual tracking performance in the context of occasional occlusion and time-variable complex background, a spatio-temporal cues based TLD visual target tracking algorithm was proposed in this paper. In the proposed algorithm, hough forests trained by multiple channels features of local image patches were used to detection. The spatial structure of the target object was introduced by multiple sampled local patches to improve the discriminating ability. Then, the initial places of the optical flow tracking were set according to the randomly sampled local patches and the tracking results were weighted by the spatial information. In the end, the output confidences of the optical flow tracking and the output confidences of the detections of the hough forests were fused in an adaptive spatio-temporal way. Experimental results demonstrate advantages of the proposed method in handling occlusion and tracking in complex environment, compared with the original TLD.