复杂场景下基于多特征融合的视频跟踪

为了解决常见视频跟踪方法在复杂场景中难以有效跟踪运动物体的难题,研究了在粒子滤波框架下基于多特征融合的判别式视频跟踪算法.首先分析了特征提取和跟踪算法的鲁棒性和准确性的关系,指出融合多种特征能有效地提升算法在复杂场景中的跟踪效果,然后选择提取HSV颜色特征和HOG特征描述目标表观,并在线训练逻辑斯特回归分类器构造判别式目标表观模型.在公开的复杂场景视频进行测试,比较了使用单一特征和多种特征的实验效果,并且将所提算法和经典跟踪算法进行了比较,实验结果表明融合多种特征的视频跟踪更具鲁棒性和准确性.

Multiple Features Fusion for Object Tracking in Complex Scenes

To address difficulties of traditional object tracking methods which can''t track moving object effectively in complex scenes, this paper proposes a multiple features fusion based discriminative object tracking algorithm in particle filter framework. This paper firstly analyzes relationship between feature extraction and robustness and accuracy of tracking algorithm, and points out that it can promote tracking performance largely by using multiple features in complex scenes. HSV color feature and HOG feature are selected to represent appearance of object, and the online trained logistic regression classifier is used to construct the discriminative appearance model. The method is tested in public videos with complex scenes. We compare results obtained by using only one kind of feature and multiple kinds of features. And we also compare our method with other classic tracking algorithms. Experimental results shows that the proposed object tracking algorithm with multiple features is more robust and accurate.