The invention discloses a hand segmentation method based on region growing and machine learning. The method comprises the steps: firstly collecting a scene depth image; secondly completing the segmentation of a hand region by region growing, wherein the growing needs to meet the following three constraint conditions: the depth difference between a to-be-determined point and a current growing pointis less than a threshold value 1, the depth difference between the to-be-determined point and a seed point is less than a threshold value 2 and the three-dimensional distance difference between the to-be-determined point and the seed point is less than a threshold value 3; judging whether the area of the obtained growing region obtained by the growing meets the requirements or not, and then growing the growing region; and finally carrying out the binary classification of the growing region so as to determine if the growing region is a hand area or not. Because the priori knowledge of a hand is employed, the method reduces the classification complexity, and greatly improves the hand segmentation and detection accuracy and efficiency. In addition, the invention also proposes a training method for a hand segmentation classifier.