A powerful and cost-efficient human perception system for camera networks and mobile robotics

In this work, we present a software library which enables the efficient use of the Kinect One, a time-of-flight RGB-D sensor, with the nVidia Jetson TK1, an ARM-based embedded system, for the purpose of people detection. Our software exploits nVidia CUDA to process all data necessary for robust people detection algorithm and other perception algorithms by parallelizing the generation of the 3D point cloud and many pixel-wise operations on both the raw depth and the infrared images coming from the Kinect One sensor. The library developed has been released as open-source and the whole system has been tested as a people detection node in an open source multi-node RGB-D tracking framework (OpenPTrack). The results gathered show that the proposed system can be effectively used as a people detection node, outperforming the state-of-the-art in terms of people detection frame rate not only with the nVidia Jetson, but also with non-embedded computers. © Springer International Publishing AG 2017.