People detection is a challenge task in the field of computer vision. This paper aims to detect the pedestrians with high precision by only using the depth information provided by TOF(time-of-flight) camera. When the TOF camera is overhead, the human head, as a part of the human body, has very rich feature information and it can be unobstructed for a long time. In view of this, a people detection method, combing blob detection with water filling algorithm, is proposed. The proposed method first uses the mixture Gaussian background model to find the region of interest in the image; then it uses the water filling algorithm to filter out the candidate heads; at last, it combines with the prior condition to determine the real heads in the scene. In order to verify the effectiveness of the proposed algorithm, two kinds of real-time depth image datasets were tested. The result turns out that, compared with two other algorithms, the proposed method has better performance, and can achieve real-time and accurate people detection.