改进的基于多示例学习的目标跟踪方法研究

摘要

针对复杂场景下的跟踪问题，提出一种新的基于多示例学习的目标跟踪方法。该方法首先利用局部描述算子（ Harr-like特征）表征目标和周围背景区域，分别视为正负样本，然后利用基于Boosting的在线多示例学习（ MILBoost ）建立一种适应性的外观模型作为二值分类器。并提出一种修正的搜索目标位置算法，使haar小波和区域协方差矩阵相结合，取最大响应样本为新目标位置。该方法能够有效解决视频场景中目标受遮挡、旋转和光照变化等问题，具有鲁棒的跟踪性能。%For tracking issue in complex scenes , a novel target tracking method based on multiple instances learning (MIL) is proposed.In this method, a local descriptor (Harr-like features) is used to present the target and surrounding background area , and respectively, as the positive and negative samples .Then by using the Boosting-based online multiple instances learning ( MILBoost ) , an adaptive appearance model is established as the binary classifier .Moreover , a modified target location search algorithm is proposed , which makes the Harr-like features and region covariance matrix combined and takes the maximum response sample as the new target location .The method can effectively deal with the problems in video scene such as target occlusions , rotations and illumination changes , etc., and has robust tracking performance .