融合运动和颜色特征的视频目标粒子滤波跟踪

Particle Filter for Video Targets Tracking by Fusing Color and Motion Features

研究可见光的视频运动目标跟踪问题。运用粒子滤波进行视频图像的目标跟踪时,目标特征的选择较为重要。传统的基于颜色特征的粒子滤波跟踪,在背景与目标的颜色分布相同时容易出现目标丢失现象。运动和颜色的融合信息为特征,在包含目标的局部区域内进行光流计算,并定义运动观测为粒子的运动像素数量。在粒子滤波框架下将运动观测融入到重要性抽样函数中,扩大预测样本与观测似然峰值的重叠区域;以运动和颜色的融合信息形成联合观测似然函数,并根据它们单独观测的质量自适应地确定各自权重。实验结果表明,改进的跟踪算法在背景存在颜色干扰时的鲁棒性和目标发生机动时的准确性均有提高。

The tracking of moving video target is studied. Choosing the suitably features is important to target tracking by particle filter, the traditional feature is colored, which is poor in the background with similar color. The feature fusing the color and motion is chosen. First, the optical flow of the area contained the object is computed, and the particle＇s motion observation is defined in term of the sum of the moving pixels. Under PF framework, the lat- est motion observation is integrated into the importance density function so as to extend the overlapped regions of pre- diction samples and peak zones of observation likelihood. The color and motion features are fused in the observation likelihood function and their weights are adjusted adaptively according to the quality of individual observation. Two experiments show that the proposed method is accurate and robustly in the moving object tracking.