基于稀疏表达特征选择的压缩感知目标跟踪算法

Compressed Sensing Target Tracking Algorithm Based on Sparse Expression Feature Selection

目标跟踪是计算机视觉领域重要研究方向之一。压缩感知跟踪速度快、精度高,但是跟踪被遮挡目标时使用被遮挡的哈尔特征构建分类器,导致分类器性能降低,目标容易丢失。为了解决该问题,提出了根据l1稀疏表示判断哈尔特征是否被遮挡,然后使用未被遮挡的特征构建贝叶斯分类器。首先对每一帧跟踪结果运用稀疏表示提取出未被遮挡特征的集合,在构建贝叶斯分类器时仅使用未被遮挡的特征。然后使用训练好的分类器对下一帧候选样本进行分类,选取具有最大分类响应的候选样本作为跟踪结果。实验结果表明,该算法在跟踪目标部分遮挡时相比CT算法有更高跟踪准确度,算法能够实时得到高效、准确的目标跟踪结果。

Target tracking is one of the important research directions in computer vision.Compressed sensing tracking（CT）is fast and of high precision,but easy to lose track of the targets which undergo occlusion.This is mainly due to use of occluded Haar features in the tracking process to construct the classifier.In order to solve this problem,this paper proposes to judge whether the Haar feature is occluded according to the sparse representation,and then construct the Bayesian classifier using the non-occluded features.Firstly,the sparse representation of each frame is used to extract the unobstructed feature sets,and only the unobstructed features are used in constructing the Bayesian classifier.Then the trained classifier is used to classify the candidate samples of the next frame,and the candidate sample with the largest classification response is selected as the tracking result.The experimental results show that the algorithm has higher tracking accuracy than the CT algorithm when tracking the targets with partial occlusion,and the algorithm can get efficient and accurate target tracking results in real time.