Abstract: This paper proposed a method of adaptive kernel density estimation (KDE) for motion detection. To begin with, an approach for adaptive selecting thresholds of foreground and background was proposed. By using the two thresholds, the approach can overcome defects of using only one threshold. More importantly, these two thresholds can be selected automatically and they are independent of scenes. Meanwhile, a background model updated according to probability was also provided. The background model of inter-frame difference incorporated with results of KDE can solve deadlock situations in background model. It can also be used to detect suddenly changed background. Experimental results were given to demonstrate that the proposed algorithms are suitable and effective for motion detection.