摘要
多实例学习跟踪算法在目标物经历较大姿势变化后,容易导致目标物跟踪漂移。针对这个问题,提出将粒子滤波与多实例学习模型相结合,利用多运动模型约束与K-means聚类构建分类器来在线检测与跟踪目标。特征模型的构建基于稀疏随机矩阵,使得图像采样得到的低维特征能保持来自图像的多尺度信息。多实例学习算法用来采样正负样本集合,并使用在线增强技术来构建强分类器。大量的实验结果表明,提出的算法在光照变化、遮挡、以及形变的变化下都能准确跟踪目标,并具有很高的实时性。
MuItipIe instance Iearning ﹙MIL﹚ tracking often suffers from tracking drift when objects undergo Iarge pose change.To soIve this probIem,a noveI MIL tracking with particIe fiIter tracking decomposition is proposed in this paper.The detection of target is proceeding with the constraints of muItipIe motion modeIs and K-means cIustering.The features modeI of tracking system is initiaIized with sparse random matrix,which makes Iow-dimension sampIed features maintain intrinsic information of origin high-dimension attributes.MIL method is used to construct discriminative modeI of the bag for finaI strong cIassifier.
出处
《工业控制计算机》
2015年第4期104-105,108,共3页
Industrial Control Computer
