Mechanical Hand

The system consists of eleven steering gears to simulate the joints of the hands and one camera to obtain video information. With the cooperation of various sensors and related algorithms, the system can successfully reproduce human posture.

During the data collection phase, the dataset consisted of 5000 images with different skin colors, lighting, and backgrounds. It was divided into a training set, a validation set and a test set. The YOLOv5 model was used to predict five gestures on the test set. In the case of using the original model, experimental results showed that after 100 rounds of training, the model performance index was 81.1% precision, 81.4% recall and 0.531 MAP.

In terms of algorithm optimization, based on the original YOLOv5 model, the CBAM attention mechanism, multi-scale detection method and Gaussian filtering were introduced to enhance the model's efficiency further. After testing on the same test set, the experimental results showed that the model performance indexes were 87.2% precision, 84.1% recall and 0.564 MAP. Through optimization, model accuracy improved by 6.1%, recall rate increased by 2.7% and MAP improved by 3.2%. Experimental tests showed that the proposed enhanced method was feasible and could accurately recognize five basic gestures.

Finally, the model was deployed on a Raspberry Pi development board and was able to quickly reproduce the human pose with a communication delay of less than 1 ms.