融合显著性与运动信息的相关滤波跟踪算法

Correlation Filter Based Visual Tracking Integrating Saliency and Motion Cues

主流的目标跟踪算法以矩形模板的形式建立被跟踪物体的视觉表征,无法有效区分目标与背景像素,在背景复杂、目标非刚体形变、复杂运动等挑战性因素影响下容易出现模型偏移的问题,导致跟踪失败.与此同时,像素级的显著性信息与运动先验信息作为人类视觉系统有效区分目标与背景、识别运动物体的重要信号,并没有在主流目标跟踪算法中得到有效的集成利用.针对上述问题,提出目标的像素级概率性表征模型,并且建立与之对应的像素级目标概率推断方法,能够有效利用像素级的显著性与运动观测信息,实现与主流的相关滤波跟踪算法的融合;提出基于显著性的观测模型,通过背景先验与提出的背景距离模型,能够在背景复杂的情况下得到高辨识度的像素级图像观测;利用目标与相机运动的连续性来计算目标和背景的运动模式,并以此为基础建立基于运动估计的图像观测模型.实验结果表明,提出的目标表征模型与融合方法能够有效集成上述像素级图像观测信息,提出的跟踪方法总体跟踪精度优于多种当下最先进的跟踪器,对跟踪场景中的背景复杂、目标形变、平面内旋转等挑战性因素具有更好的鲁棒性.

Rectangle template is a popular target representation adopted by mainstream visual tracking methods.However,by including some background clutter as part of the target representation,the model is likely to drift away from the target gradually and result in tracking failure,especially in challenging situations such as background clutter,target deformation and complex motions.Meanwhile,motion and saliency cues,which play important roles in distinguishing targets from the background and identifying moving objects in the human vision system,have not been modeled into existing tracking methods.To solve these problems,we propose a foreground probabilistic inference formulation that collects pixel-level observations from different sources,and a unified framework integrating the pixel-level model with a widely used correlation filter based method.A saliency-based observation model is proposed by introducing background prior and a distance-based model,which provides reliable evidence to resolve confusion caused by appearance similarity between targets and the background.By taking advantage of continuity and inertia of both target and camera motion,we discover motion patterns in the spatial domain to distinguish targets from the background,and introduce a pixel-level motion-based observation model.Experiments demonstrate that the proposed method outperforms some of the state-of-the-art methods,and shows better robustness in challenging situations such as background clutter,target deformation and in-plane rotation.