Learning-Based Collision Detection with Estimating Occluded Keypoint for Robotic Manipulators

The increasing importance of industry 4.0 expanded the applications of robotic manipulators. Due to the increase in the number of robotic manipulators at work, the probability of collision becomes a prominent threat. As a result, the urgency for real-time collision detection rises. Thus, we proposed a learning-based collision detection with estimating occluded keypoint for robotic manipulators based on the DeepLabCut toolbox keypoint model and depth camera. By identifying key points in robotic manipulators, we detect collisions through the distance formula and occlusion estimation. The experimental results indicate that method achieves 88.14% accuracy with a speed of 22.91 frames per second (FPS). Due to the characteristics of accuracy and high operation speed, it is suitable for real-time collision detection in realistic environments.