CAJ | 학술논문

针对复杂场景下目标跟踪中目标出现的外观特征变化和遮挡问题，提出一种结合超像素和广义霍夫变换的在线实时目标跟踪算法。该算法从上下文中提取局部特征作为支持因子，构建一个混合的判别产生式对象模型。利用该模型，通过霍夫投票预测目标的中心位置，再通过判别式投票对目标和背景进行概率估计。对图像进行超像素分割，将之前的投票结果映射到对应的超像素，生成基于超像素的概率分布图像。采用贝叶斯跟踪框架，根据后验概率最大化，在概率分布图像基础上确定目标的位置。实验表明，该算法在复杂环境下目标跟踪的过程中对目标发生的形变和遮挡现象有很强的鲁棒性，能够实现准确稳定的在线目标跟踪。
침대복잡장경하목표근종중목표출현적외관특정변화화차당문제，제출일충결합초상소화엄의곽부변환적재선실시목표근종산법。해산법종상하문중제취국부특정작위지지인자，구건일개혼합적판별산생식대상모형。이용해모형，통과곽부투표예측목표적중심위치，재통과판별식투표대목표화배경진행개솔고계。대도상진행초상소분할，장지전적투표결과영사도대응적초상소，생성기우초상소적개솔분포도상。채용패협사근종광가，근거후험개솔최대화，재개솔분포도상기출상학정목표적위치。실험표명，해산법재복잡배경하목표근종적과정중대목표발생적형변화차당현상유흔강적로봉성，능구실현준학은정적재선목표근종。
It is a great challenge to track an obj ect robustly when variations occur such as changes in illumination,appearance or partial occlusion. In this paper, we propose a target tracking algorithm combining superpixel and hybrid Hough voting.Local features are extracted from the context as supporters to construct a hybrid voting model.By this model,the target center is estimated by the Hough voting scheme.Local features are also distinguished to vote for the target and background,respectively.These voting results are combined into superpixels.Finally,the tracking task is formulated as the maximum a posterior estimate in the voting space.We demonstrate the performance of the algorithm on several public video sequences,which shows that our method is better than other online tracking approaches.