利用检测特征空间的目标实时跟踪

Target Tracking Based on Feature Space of Detection

针对被跟踪目标尺度小、特征颜色与场景颜色相似的问题,提出一种基于实时检测结果的视频目标跟踪算法,即首先对背景进行高斯建模,利用背景减除法和帧间差分算法对前景区域进行有效提取,然后在提取的前景区域内进行基于均值移动算法的目标跟踪。基于像素级别的背景减除与帧间差分算法虽然精确和灵敏的优点,但是鲁棒性不强;而基于块级别的均值移动算法虽然鲁棒性强,但是弱化了特征颜色的空间信息,本文对两种机制进行了有效融合。通过该策略,跟踪系统在目标快速运动、有场景相似颜色干扰等情况下具有很好的跟踪性能,算法的计算量小,能够满足实时性要求。通过多组对比实验可以看出,新算法具有很强的抑制背景干扰、提高均值移动跟踪算法鲁棒性的能力。

To solve the problem of small targets or targets which have similar color with scene background in tracking, we propose an object tracking method which exploits real-time detection results. First, scene background is effectively modeled and foreground mask is obtained by background subtraction and frame differencing, then,mean shift algorithm is applied to objects tracking in the fused image space. Though precise and sensitive, pixel-level processing algorithms such as mixture of Gaussian and frame differencing are not robust. The mean shift algorithm which is a block-level processing one is robust whereas it weakens spatial information of feature space. This paper effectively combines the merits of both algorithms to achieve robustness and accuracy of object tracking. Through the method, our system shows very good tracking performance for targets which move fast or have similar color disturbance with scene background. Also, our algorithm is efficient and the computation of the novel algorithm is fast to satisfy real-time application. Several groups of comparative experimental results show that the new algorithm can effectively suppress the scene background disturbance and improve the performance of object tracking. Experimental results on video clips demonstrate the effectiveness and efficiency of our method.