A low-cost Raspberry PI-based vision system for upper-limb prosthetics

To achieve human like stable grasp with prosthetic, the correct grasp type along with the appropriate grasp aperture need to be identified from the targeted object. In most of the advance myo-electric prosthetics, the problem has been solved by using computer vision system in addition with the prosthetics. The vision systems are placed over the prosthetic itself and it needs to move towards the target object. Upon reaching the target object, the system capture the object image and process it to generate the correct grasp type. The processing time delays the initiation of the grasp process. To reduce this time delay and make the grasp instant like normal human hand, we have developed a vision system with Raspberry PI along with the PI camera. The system moves toward the targeted object and captures images of the focused region at two intermediate positions. OpenCV DNN model was used to identify all the objects within the captured frame. The employed model provided the efficiency of deep learning models in object identification without using of any GPU. The width of the objects in terms of pixels were also estimated from the images. The distance travelled by the system was measured by the SensorHAT (integrated with the system) which recorded the acceleration and direction of the hand movement. The device was tested in the cluttered environment in recognising four different grasp types- Compliant, Lateral, Tripod closed and Wrist-flexion aided grasp.