Tracking and proximity detection for robotic operations by multiple depth cameras

The maintenance of physical safety is of critical importance in an environment where there are human-robot interactions. This paper implements a proximity detection framework that protects objects, robots and humans using the point cloud information for the robot, objects and background in the workplace. The proposed algorithm tracks the sequential change in the point cloud for the robot to determine whether the robot is about to collide objects in the surroundings. The system is divided into three parts. Point clouds that are obtained from multiple RGB-D cameras are initially transformed with respect to the world coordinate system, and these transformed point clouds are then merged into a world point cloud. Using robot simulation software, a CAD model of the robot is used to generate an ideal point cloud that represents the robot itself. A random sample consensus algorithm is then used to accommodate the robot's point cloud, the shape of which is consistent with the ideal model and to separate it from the other point cloud, which represents the environment. The two respective point clouds are finally constructed using the Octree data structure. The point clouds in Octree are used to perform a fixed-radius vicinity search around each grid to classify the outlier points and updates Octree, in order to determine whether the robot is too close to an object in the environment. Experiments using an industrial 7-Axis manipulator verify the feasibility of the method that is proposed in this study. To ensure that the entire system can be used with other types of robots or sensors, the entire system is implemented using the protocols and framework of the Robot Operating System (ROS).