基于多区域的随机蕨在线目标跟踪算法

摘要

原始的随机森林跟踪算法,是以像素点的灰度值作为检测特征,在目标发生遮挡和旋转时,容易产生跟踪漂移,为此本文提出了一种基于多区域融合的随机蕨在线目标跟踪算法.首先将目标候选区域划分为多个子区域,然后采用基于积分图的随机蕨分类器对每个子区域的候选图像块进行分类,在跟踪过程中自适应地融合子区域分类结果以剔除被遮挡子区域对目标跟踪结果的影响,同时更新随机蕨特征和子区域图像块的选择.结合对TLD算法部分模块的改进,通过对不同视频序列进行测试,实验结果显示本文算法在跟踪大小320 pixel×240 pixel的视频序列时,跟踪速度达到20~30 frame/s左右,目标中心位置误差在30 pixels时,算法准确率可达到80%以上.%The random forest algorithm which uses gray value of pixel-pair to construct the binary features is apt to lead to the tracking failure (drift), especially the appearance change caused by occlusion, fast motion, and illumination variation. In this paper, an online tracking algorithm using the random ferns based on multi-region is proposed. First, the candidate area is divided into different sub-areas, then the random ferns based on the integral images is used to classify the patches, and the detection result being adaptively fused in order to decrease the influence of the occluded sub-regions. Combined with the improvement of TLD algorithm, the experiment is made on four different videos. The results show that this proposed algorithm performed better on some complex scenes. In the video sequence of 320×240pixels, the speed can keep on 20~30 frame/s or so, and the object center position error is in 30 pixels, while the accuracy can reach above 80%.