基于HOG与多实例在线学习的目标跟踪算法

Object Tracking Algorithm Based on HOG and Multiple-instance Online Learning

为实现在局部遮挡、光线变化等复杂背景下的目标跟踪,提出一种基于梯度方向直方图(HOG)与多实例在线学习的目标跟踪算法。利用已标定目标图像的HOG特征空间,结合局部二值模式(LBP)描述方法获取特征向量,构建初始随机蕨检测算子,采用随机多尺度采样方法跟踪每一帧的目标位置和尺寸,并基于多实例在线学习框架,通过检测到的目标样本以及附近的背景样本在线更新检测算子。将该算法与Online Boosting Tracker,MILTracker等在线学习目标跟踪算法在多个标准视频序列中进行比较,实验结果表明,该算法在局部遮挡和光照变化的环境下具有较好的跟踪稳定性,但在抗目标旋转方面有待优化。

In order to achieve effectively stabilized target tracking within partial occlusion,illumination changes and complex background environment,this paper presents an object tracking algorithm based on Histogram of Oriented Gradients（HOG）and Multiple-instance Learning（MIL）. Using the HOG feature space of the target block and the background in the first frame with Local Binary Pattern（LBP）descriptor to initialize the initial random ferns,it detects the target location and the objective scale of each frame with random multiple-scale sampling and uses the new target samples and the nearby background samples to update the appearance model within multi-instance learning after each detection. Through the experiments,the algorithm with multiple online tracking algorithms such as Online Boosting Tracker and MILTracker are compared and analyzed in a number of video sequences. The results show that it has a good target tracking stability under the complex environment,especially with partial occlusion and illumination changes,but in the anti-rotation of target,the algorithm has yet to be optimized.