在线特征选取的多示例学习目标跟踪

目的 传统的多示例学习跟踪在跟踪过程中使用了自学习过程,当目标跟踪失败时分类器很容易退化。针对这个问题,提出一种基于在线特征选取的多示例学习跟踪方法(MILOFS)。方法 首先,该文使用稀疏随机矩阵来简化视频跟踪中图像特征的构建,使用随机矩阵投影来自高维度的图像信息。然后,利用Fisher线性判别模型构建包模型的损失函数,依照示例响应值直接在示例水平构建分类器的判别模型。最后,从梯度下降角度看待在线增强模型,使用梯度增强法来构建分类器的选取模型。结果 对不同场景的图像序列进行对比实验,实验结果中在线自适应增强(OAB)、在线多实例学习跟踪(MILTrack)、加权多实例学习跟踪(WMIL)、在线特征选取多实例学习跟踪(MILOFS)的平均跟踪误差分别为36像素、23像素、24像素、13像素,本文算法在光照变化、发生遮挡,以及形变的情况下都能准确跟踪目标,且具有很高的实时性。结论 基于在线特征选取的多示例学习跟踪,跟踪过程使用梯度增强法并直接在示例水平构建包模型的判别模型,可以有效克服传统多示例学习中的分类器退化问题。

Object tracking based on multiple instance learning with online feature selection

Objective Traditional multiple instance learning (MIL) tracking utilizes self-learning procedures in tracking systems. Once an object gets lost in tracking, the interior classifier easily degenerates. To alleviate this problem, we propose an improved multiple instance learning tracking based on online feature selection (MILOFS). Method First, a very sparse random matrix is constructed to facilitate the feature initial process. With this matrix, the intrinsic attributes of the features projected from a high-dimension image can be preserved. Then, the loss function of a bag model is built with the Fisher linear discriminant model. The discriminative model of the bag is formed directly at the instance level with the response of each instance. Finally, the gradient descend rule is incorporated into the online boosting framework, and gradient boosting is employed to construct the selection strategy for strong classifiers. Result Comparison experiments under different scenarios reveal that the center location errors of Online AdaBoost(OAB), Online Multiple Instance Learning Tracking (MIL-Track), Weighted Multiple Instance Learning(WMIL) and Multiple Instance Learning with Online Feature Selection(MILOFS) are 36, 23, 24, and 13 pixels, respectively. Hence, the proposed method is robust and accurate regardless of changes in the illumination, occlusion condition, and target appearance in the outer environment. Conclusion An improved MILOFS is proposed in this work. The proposed method integrated with a gradient boosting framework and online feature selection strategy effectively addresses the issue of classifier degeneration in traditional MIL tracking.