A novel occlusion detection approach is proposed for depth image by using Random Forest. The occlusion related features of each pixel in the depth image are extracted, and then the Random Forest classifier is used for detecting whether each pixel is an occlusion boundary point or not. All the occlusion boundaries in the input depth image are obtained. This work is distinguished by three contributions. A new occlusion related feature named depth dispersion is proposed and the Gaussian curvature feature is introduced, and both of them are used for occlusion detection by combining with other features. All the occlusion related features in depth image are analyzed and evaluated by using the importance and extraction time as criterion. On this basis, five features such as average depth difference, maximal depth difference, mean curvature, Gaussian curvature and depth dispersion are selected for designing the occlusion detection classifier. A new occlusion detection approach takes the Random Forest to solve occlusion detection problem in depth image. The experimental results show that, compared with the existing methods, the proposed approach has higher accuracy and better generality.