多通道Haar-like特征多示例学习目标跟踪

目的 提出一种基于多通道Haar-like特征的多示例学习目标跟踪算法，克服了多示例跟踪算法在处理彩色视频时利用信息少和弱特征不能更换的缺点。方法 首先，针对原始多示例学习跟踪算法对彩色视频帧采用单通道信息或将其简单转化为灰度图像进行跟踪会丢失部分特征信息的缺点，提出在RGB三通道上生成位置、大小和通道完全随机的Haar-like特征来更好地表示目标。其次，针对多示例学习跟踪算法中Haar-like弱特征不能更换，难以反映目标自身和外界条件变化的特点，提出在弱分类器选择过程中，用随机生成的新Haar-like特征实时替换部分判别力最弱的Haar-like特征，从而在目标模型中引入新的信息，以适应目标外观的动态变化。结果 对8个具有挑战性的彩色视频序列的实验结果表明，与原始多示例学习跟踪算法、加权多示例学习跟踪算法、基于分布场的跟踪算法相比，提出的方法不仅获得了最小的平均中心误差，而且平均跟踪准确率比上述3种算法分别高52.85%，34.75%和5.71%，在4种算法中获得最优性能。结论 通过将Haar-like特征从RGB三通道随机生成，并将判别力最弱的部分Haar-like弱特征实时更换，显著提升了原始多示例学习跟踪算法对彩色视频的跟踪效果，扩展了其应用前景。

Multiple instance learning based object tracking with multi-channel Haar-like feature

Objective A multi-channel Haar-like feature based object tracking algorithm with multiple instance learning(MIL)is proposed in this paper. It overcomes the disadvantages of the MIL algorithm such as using limited information and not replacing weak features for color videos. Method First,in the original MIL algorithm, the color video frame is tracked with a single channel's information or by simply converting it to grayscale images. This may lose some feature information. Therefore,we propose that the target is represented with Haar-like features generated from three channels of RGB with completely random location,size and channel to represent the target better. Next,Haar-like features could not be replaced in the original MIL algorithm,which has difficulty reflecting the changes of the target and the background. Thus, we replace some weakest discriminative Haar-like features with new randomly generated Haar-like features when weak classifiers are selected. It introduces new information to the target model and adapts to the dynamic changes of the target appearance. Result The experiment on eight challenging color videos shows that the proposed method obtains optimal performance compared with the original multiple instance learning algorithm,weighted multiple instance learning algorithm, and distribution field based algorithm. It not only obtains the minimal average center location errors,but also obtains a higher average accuracy rate by 52.85%,34.75% and 5.71% than the other three algorithms. Conclusion The proposed algorithm obviously promotes the tracking results compared with the original MIL algorithm on color videos by generating Haar-like features from three RGB channels and replacing some weakest discriminative Haar-like features in real time. It extends the application prospect of the MIL algorithm.