In this study, we propose a new fingertip detection algorithm using two-stage random decision forest (RDF). In the first stage, local depth difference pattern (LDDP) and 3D geodesic shortest path (GSP) are adopted for training a finger pixel classifier. Two spatial and temporal features are then added into RDF to further distinguish fingertip pixels from finger pixels in the second stage. Finally, we utilize K-means clustering to re-identify fingertip candidates and limit the number of candidates to five. Our experimental result demonstrates that the proposed fingertip detection method is effective in complex gesture.