结合排序向量SVM的视频跟踪

Visual tracking combined with ranking vector SVM

针对真实视频场景中复杂的目标外观变化问题,提出新的结合排序向量SVM(RV-SVM)的单目标视频跟踪算法.基于压缩感知理论,利用稀疏测量矩阵压缩多尺度图像特征.采用Median-Flow跟踪算法作为预测器,并为RV-SVM构建训练数据集,使算法能够适应真实场景中遇到的目标遮挡、3D旋转和目标快速移动等复杂情况.通过在线学习RV-SVM算法,对候选位置集进行排序,找到目标的真实位置.对不同视频序列的测试结果表明:该方法可以在目标运动、旋转以及光照和尺度发生变化的情况下实现准确的跟踪.

A novel single object video tracking algorithm with ranking vector SVM （RV-SVM） was proposed for complex changes of object appearance in realistic scenarios. A sparse measurement matrix based on compressive sensing theory could compress the multi-scale image features. A Median-Flow tracker algorithm was used as a predictor and to construct training data sets for RV-SVM algorithm, so that the algorithm could adapt complex conditions like object occlusion, 3D rotation and fast object motion. The real position of target was determined through training the RV-SVM algorithm online and ranking the candidate position set. Results of tests on variant video sequences show that the algorithm can achieve stable tracking either the object is moving, rotating or the illumination and scale is changing.