A Safety Preservation Type Vision System Based on Probabilistic Measures

The study concerns a safety vision system for human-robot collaborative work environments. The vision system is characteristic of having a freedom from structuring the environment by setting passive markings, i.e. neither painting on the floor nor hanging a camera up on a ceiling is necessary. To preserve safety states, we set the passive markings in the original image based upon Shannon information quantity and to monitor the states by the drastic changes in the measures of Kullback-Leibler divergence and Levy-distance. Our experimental result shows that all of the markings are successfully traced even when the camera view angle is changed and that the system detects any difference of an image from the original one as a human hand intrudes to the detection area. We discuss the performance of the applied measures from a detectability view point.