基于高斯金字塔的运动目标检测

Moving object detection based on Gaussian pyramid

针对自然环境下运动目标检测背景动态变化问题,提出一种新的基于高斯金字塔模型的背景差分算法。首先将图像序列进行多尺度分解,得到不同分辨率下的当前帧和背景帧;然后,在不同分辨率下采用高低双阈值进行背景差分运算,得到双阈值产生的2帧前景图像,阈值根据环境自动获取;最后,将各层差分图像自顶向下融合检测感兴趣的运动目标,并在HSV空间中去除阴影。背景模型的初始化和更新方法基于2种假设:一是背景点出现的概率较大;二是距离当前帧越近的点越能真实地描述背景。研究结果表明:该算法能有效地应用于动态背景环境下,可以克服光照变化及阴影的影响。多个标准图像序列的测试证明了该算法具有较高的准确性、鲁棒性和自适应性,时间复杂度低,可以运用于实时检测系统中。

To solve the problem of dynamic background under natural environment when detecting moving objects,a new background difference method based on Gaussian pyramid model was proposed.Firstly,multi-scale decomposition was carried out for image sequence to get multi-resolution images.Then a high and low double thresholds background difference operation was used under different resolutions to get two foreground images by dual-threshold.All the thresholds were obtained automatically according to the environment.At last,difference images in each layer were fused top-down to detect the interested moving objects,and shadows were removed in HSV space.Background model initialization and update method were based on two assumptions,the first one of which is that background points appear with a larger frequency and the second is that the closer to the current frame,the more likely to represent the real background.The results show that the proposed algorithm can be effectively applied to dynamic background environments and can overcome the effect of illumination changes and shadows.Experiments on several standard image sequences demonstrate that the proposed method has high accuracy,robustness and adaptability.It has lower time complexity and can be applied in real-time detection systems.