A geometric approach to multiple viewpoint human body pose estimation

People detection and re-identification is a crucial capability for mobile robots working in a human environment, as well as for human-robot interaction. Re-identification systems can be based on the observation of a number of cues, including the analysis of the human body pose, that can be accurately detected analyzing RGB-D data, currently widely used in robot vision. On the other hand, intelligent video surveillance is going towards multi-viewpoint RGB camera systems: skeletal trackers working on images are currently unable to provide performance similar to those based on 3D data. To overcome such flaws, this paper proposes a method for merging together the results provided by a body pose estimation algorithm observing the same scene from different viewpoints: this enhances the accuracy level, and lets the system recover 3D information, leading to a target representation which is more similar to the one obtained using 3D sensors. Such similarity is a first step to achieve a stronger cooperation between robots and camera networks, a capability that opens new scenarios in robotics.