传统的压缩感知目标跟踪算法在光照变化剧烈、目标与背景存在一定相似性的情况下容易产生跟踪偏差，故本文提出了一种基于梯度方向直方图特征进行压缩感知跟踪方法。该方法用梯度方向直方图特征替换原来的广义类Haar特征进行压缩感知跟踪。首先，将梯度方向直方图作为原始特征，并利用压缩感知理论得到稀疏表示的特征子空间；然后，在后续帧中用朴素贝叶斯分类器进行目标位置的搜索；最后，对分类器进行在线更新。由于梯度特征能更稳定地表示目标，所以这种跟踪方法具有更好的鲁棒性；另外在计算时采用了积分直方图技术，有效克服了计算量大的问题。对不同视频的实验结果表明，该方法在实验环境Intel Core2 2.93 GHz，matlab R2010a，图像大小320×240下，跟踪速率可达到10 frame/s。在目标姿态、环境光照变化剧烈，背景中存在与目标有一定相似性的物体等情况下跟踪准确。

As traditional compressive sensing tracking algorithm will produce tracking errors in circumastances when illumination has dramatic change or there exists a object similar to the target in background, this paper proposes a sparse representation object tracking algorithm by taking the histogram of gradient feature to replace the generalized Haar feature. The algorithm uses the histogram of gradient feature as an original feature firstly, and gets the sparse representation of object feature subspace by using compressive sensing theory. In the subsequent frames, the naive Bayes classifier is used to search the target location and the classifier is online updated finally. As the histogram of gradient feature can represent the target more stably, this algorithm is more robust than original compressive tracking algorithm. Furthermore, the integral histogram is adapted to effectively reduce computational load when the gradient feature is computed. Experiments on different videos show that the tracking algorithm can reach the tracking rate of 10 frames per second in an experimental environment of Intel Core2 2.93 GHz, matlab R2010a, image size 320×240, and it achieves stable tracking in some special conditions as mentioned above.