中文摘要：

针对压缩跟踪不能适应目标姿态变化导致跟踪失败的问题，提出了一种基于二值随机森林的目标跟踪算法。该算法对实时压缩跟踪算法的特征提取和分类这两个部分作了改进。首先，在梯度图像上进行多尺度滤波，获得目标的高维特征描述，利用一个稀疏矩阵进行压缩，获得表征目标的低维信息；然后，通过比较图像块对的大小，获得二值描述符，利用随机森林构造目标表示方法；最后，计算汉明匹配、寻找汉明距离最小的候选样本作为当前帧目标的状态估计，并在此基础上提取目标的特征来更新目标特征模板。与原算法相比，该算法对旋转、折叠、遮挡等姿态变化的目标跟踪性能更好。

英文摘要：

For the failure of compressive tracking to the target whose guise was changing, this paper proposed a novel target tracking algorithm based on binary random forest. It improved the feature extracting part and classifying part in the real-time compressive tracking by proposed algorithm. Firstly, it obtained high-dimension feature of the gradient image by multi-scale filtering, and obtained the low-dimension feature by compressing. Then, it generated binary description by comparing a pair of image patches and formed target representation by random forest. Finally, it used computing Hamming match and the candi- date sample with the shortest Hamming distance as the state estimate of the target in the current frame, where the feature was extracted to update the target feature model. Compared with the original algorithm, proposed algorithm has better capability of the target tracking in these guises that rotation, fold, shelter and so on.