This article describes a multiple feature data fusion applied to an auxiliary particle filter for markerless tracking of 3D two-arm gestures by using a single camera mounted on a mobile robot. The human limbs are modelled by a set of linked degenerated quadrics which are truncated by pairs of planes also modelled as degenerated quadrics. The method relies on the projection of both the model?s silhouette and local features located on the model surface, to validate the particles (associated configurations) which generate the best model-to-image fittings. Our cost metric combines robustly two imaging cues i.e. model contours and colour or texture based patches located on the model surface, subject to 3D joint limits and also non self-intersection constraints. The results show the robustness and versatility of our data fusion based approach.