基于自适应特征融合的均值迁移目标跟踪

提出了一种基于自适应多特征融合的目标跟踪算法.分别利用RGB颜色和LBP纹理特征建立目标模型,通过线性加权将两类目标子特征模型代入目标相似性函数并用均值迁移算法进行目标位置优化计算.在跟踪过程中,引入S igmoid函数动态调整两类子特征权重,并利用子特征相关系数和可靠性指数对目标特征模型选择性自适应更新.实验结果表明,该算法能在跟踪场景和目标外观变化时自适应调整两种子特征权重,避免了特征失效导致的跟踪失败;特征模型选择性更新策略有效抑制了模型漂移.与单一特征和模型直接更新的跟踪方法相比,该算法在复杂跟踪环境更具有鲁棒性,能进行准确稳定的实时跟踪.

Object Tracking Based on Adaptive Multi-Cue Integration Mean Shift

A tracking algorithm based on adaptive multi-cue integration mechanism is proposed. The RGB color cue and local binary pattern (LBP) texture cue are utilized to represent the target, and then they are combined by linear weighting to the similarity function. By expanding the similarity function, a new expression consisting of the data items with local information is obtained, and mean shift algorithm is used to find out the optimal location by iterative computation. Sigmoid kernel are used to adjust the feature weight adaptively in the tracking procedure, Bhattacharyya coefficient and reliability index are used as criterions for selective sub-model update.Experimental results show that the tracker based on multi-cue integration mean shift works even more robustly. With adaptive multi-cue integration mechanism and selective model update strategy, the problem of tracking failures caused by using single cue or model drift in complex scenes can be solved.