Wrist Angle Estimation Using Depth Sensor and Convolutional Neural Network for Grasping with Trans-humeral Prosthesis

We propose continuous control of wrist angles of a prosthesis using a depth sensor mounted on forearm and a convolutional neural network (CNN). To confirm the feasibility of this method, we collected depth images and wrist joint angles of a person grasping objects and trained a CNN. First, a dataset for training was created by collecting depth images and wrist joint angles of the person grasping objects using three basic shape items placed in various orientations. This dataset was then used to train the CNN consisting of a total of sixty-one layers. As the result, we were able to estimate the angle of the wrist joint when a person grasps an object with an estimation accuracy of 6.80 degrees root-mean-square error.