Random Forest Distance Estimator Based on Monocular Egocentric RGB Camera in Human-Robot Collaboration

This paper presents a method for obtaining 3D gesture poses from monocular RGB camera images. A random forest based distance estimator is constructed and trained to generate depth information for hands. Furthermore, to improve efficiency, we employ a federated learning approach to reduce the model complexity of random forests. The RNN recognizes the topological structure of the hand, transmitting two-dimensional information of the hand keypoints to a random forest for the corresponding distance estimation. Experimental and application demonstration results show that this method can obtain stable 3D gestures from 2D images and reduce the complexity of the system implementation and registration. Because of input normalization, it can adapt to other cameras without the need for preconfiguration. Therefore, this method offers an effective and cost-efficient solution for obtaining 3D gesture poses, resulting in a significant cost advantage in the application of human-robot collaboration and making it highly promising for real-world applications.