The depth image and the target background image separation captured by Kinect real-time extraction exist obvious serrated and loud image noise. Proposes a denoising algorithm for Kinect depth image. Applies the auto-regulative thresholds edge detection algorithm based on Prewitt operator to obtain fine image edge, then classifies the depth image according to the edge information, and reduces noises of the edge region by the unidirectional multistage median filtering algorithm, while suppresses noises at non edge region by the bidirectional multistage median filtering algorithm. Finally, acquires high quality depth images of clear-edge and small noise by the proposed Kinect algorithm, and verifies the effectiveness of the algorithm.