为了处理运动目标跟踪中的遮挡、光照变化以及背景杂乱等问题, 提出了一种基于加权多实例学习的压缩感知目标跟踪方法。在提取图像块类Haar特征的基础上, 采用随机投影方法对高维特征进行压缩, 结合加权多实例学习策略, 在boosting学习框架下训练分类器, 根据分类器最大响应值得到跟踪目标图像块。使用矩形框手动标定第一帧图像的目标, 对后续帧采样的正负实例的类哈尔特征进行压缩, 通过这些样本特征训练分类器, 得到跟踪结果。实验结果表明, 所提算法具有较高的跟踪精度、实时性和鲁棒性, 对所选的四个具有挑战性的视频序列, 跟踪成功率都能超过89%, 帧率也超过26f/s。

In order to handle occlusion, illumination change and background clutter, a novel target tracking method adopting compressed sensing and weighted multiple instance learning is proposed. After extracting Haar-like features of an image patch, the dimension of these features is reduced by random projection. Combining weighted multiple instance learning, the tracked target patch is located as the classifier’s maximum respond value in the framework of boosting learning. The target in the first frame is labeled manually, and the Haar-like features extracted from the positive and negative instances is reduced for the sequence frames and the tracking patch is obtained by training classifiers. Experimental results illustrate the accuracy, real-time and robustness of the proposed method, for the selected four challenging video sequence, the tracking success rate can over 89% and frame rate over 26f/s.