Performance Analysis of Body Tracking with the Microsoft Azure Kinect

In industrial environments, a proper trajectory planning of cooperative robots, or cobots, is a mandatory task to reduce the risk for workers by improving their safety. The task of cobot control always requires input data about the surroundings to enable planning procedures and proper reactions to unpredictable events, such as human actions. In this case, the exact position of humans can be easily inferred from RGB-D cameras, whose output can be processed by body tracking modules to produce exact pose estimations in real-time. This paper experimentally explores the performance of the affordable Microsoft Azure Kinect RGB-D camera and its bodytracking library. A parametric analysis of the uncertainty of the estimation of the skeleton joints is performed by changing the ambient light conditions, the presence of occlusions, the infrared camera resolution, and the human-camera distance. The output of this investigation proves the need for uncertainty management in the control of cobots working with humans.