Multi-camera collision detection between known and unknown objects

Today, real-time collision detection is a basic demand for many applications. While collision tests between known (modeled) objects have been around for quite a while, collision detection of known objects with dynamic, unknown (sensor-detected) objects remains a challenging field of research, especially when it comes to real-time requirements. The collision test described in this paper is based on several stationary, calibrated video cameras, each supervising the entire 3-dimensional space shared by unknown and known objects (e.g. humans and robots). Based on their images, potential collisions of the known objects in any of their (future) configurations with a priori unknown dynamic obstacles are detected. Occlusions caused by known objects (such as the robot or machinery set-up within the workspace) are detected and addressed in a safe manner by exploiting the geometrical information of the known objects and the epipolar line geometry of the calibrated cameras in a decision fusion process. The algorithm can be parameterized to adapt to different application demands. Experimental validation shows that real-time behaviour is possible in the presence of highly dynamic unknown obstacles as they occur when humans and robots share the same workspace for the accomplishment of a shared task. In effect, the vision-based collision test can safely be used for human-robot cooperation, intrusion detection, velocity damping, or obstacle avoidance.