自适应学习的混合高斯模型运动目标检测算法

Adaptive learning algorithm for moving target detection based on Gaussian mixture model

针对传统混合高斯模型采用固定学习速率带来的模态残留和拖影等问题,提出了一种自适应学习速率运动目标检测算法。对图像序列像素变化特性和模型控制参数性能进行了分析,将模型学习过程分为背景初始形成和背景维护更新两个阶段,不同阶段采取不同的学习策略,初始形成阶段采用较大递减学习速率加速背景模型的形成;维护更新阶段根据像素点匹配次数与不匹配次数作为反馈量来调节学习率,实现模型的自适应学习。实验结果表明,该算法能够有效改善原始模型收敛速率慢导致背景模型更新不及时的问题,可以更准确地检测出运动目标,并具有较好的自适应性和鲁棒性。

Aiming at the problem of modal residual and ghosting for traditional Gaussian mixture model which has a fixed lear- ning rate, an improved algorithm of moving target detection by using adaptive learning rate is presented. The variation characte- ristic of pixel on the image sequence and the model performance of controlling parameters are analyzed, model of the learning process can be divided into initial formation and background maintenance updates in two stages. To adopt different learning stra- tegies in different stage, the initial formation stage, a bigger decreasing learning rate is adopted to accelerate the background modeling. The maintenance updates stage, according to the number of the pixel matching and mismatching as feedback to adjust and implement model of adaptive learning. Experimental results show that the improved algorithm can effectively improve the original model of the slow convergence rate lead to problems in terms of background model update not in time, which can more accurately in detecting moving target. It is characterized by good adaptability and robustness.