复杂场景下的实时运动目标检测

Real-time Motion Detection in Complex Scenes

提出一种改进的混合高斯模型算法对复杂场景中的运动目标进行实时检测.该算法首先在模型更新过程中提出一种相似模型调整策略,通过对模型值接近的模型的均值、权值、学习速率等进行自动调整,从而使算法更快地适应光照变化;然后基于尺度不变性局部三值模式纹理特征对检测结果进行校验,以快速有效地消除运动目标的阴影和光照渐变及突变的影响;最后设计一种图像尺度变换方法以提高算法的实时性.与现有算法相比,本文提出的方法能更好地在复杂背景中稳定检测运动目标,同时显著提高目标检测的效率.试验结果验证了本算法的有效性.

An improved Gaussian Mixture Model algorithm was proposed for real-time motion detection in complex scenes. First, the mean values, weight and learning rate of similar models would be adjusted automatically during models update. So the algorithm could adapt to illumination change faster. Second, scale invariant local ternary pattern feature was used to verify the candidate foreground points to get rid of the moving shadow and illumination variation. Finally, an image scale transformation method was designed to improve the efficiency of the algorithm. Compared with existing state-of-the-art algorithms, the proposed method could extract the moving object in complex scenes more effective and efficient. Experiments have shown the effectiveness of the proposed method.