The innovation is about a novel hybrid human body pose estimation method using RGB-D input. The method relies on the Openpose deep net architecture to get an initial 2D pose. Using depth information from the sensor a set of 2D landmarks on the body are transformed in 3D. Since the obtained 2D and 3D landmark positions contain noise (e.g. due to inaccurate localisation, depth errors) we employ a multiple hypothesis tracker. Each pose hypothesis is generated by using a subset of the landmarks. The subsets are obtained by random weighted sampling and the weight of each landmark is calculated using a set of geometric and temporal continuity criteria. Given the landmarks subset we estimate the pose of a parametric 3D human body model by a gradient based optimization scheme. Finally, the resulting 3D poses are evaluated by an objective function that densely calculates the discrepancy between the model and the observed depth.