无表观特征小目标检测与跟踪

Featureless small object detection and tracking

检测跟踪模糊的小目标是计算机视觉领域中难度极大,富有挑战的任务。由于被跟踪的目标过小或过于模糊,难以提取合适的可用于检测和跟踪的表观特征,使得现有的目标检测和跟踪算法不能解决上述问题。前景运动物体区别于背景随机噪声的一个重要特征是运动物体具有一定的运动规律,基于这个假设提出一种新方法,根据物体的运动规律对其进行跟踪。首先,提出利用运动物体的时空域关联性,对视频中的运动目标进行分割和去噪;其次,提出了利用动态规划得出并优化物体的运动轨迹。各种条件下的实验结果表明了上述方法的精确性和鲁棒性。

The detection and tracking of small and/or indistinct objects is a challenge in computer vision, since it is difficult for cun＇ent detection and tracking algorithms to extract appropriate apparent features from small and/or indistinct objects. We found by observing video data that the small moving objects can be distinguished from noisy background due to their regular motion pattern. Based on this finding, we proposed a novel method for solving the challenging problem of small and/or indistinct object detection and tracking. The contribution of this paper is the denoising algorithm for small object segmentation and the dynamic programming algorithm for route optimization in tracking. Extensive experiments show the promising results in both accuracy and robustness of the proposed method.