Most people with disabilities due to accident injuries are patients after an arm amputation which can cause psychological disorders and even major trauma. The urgency of hand prosthesis functionality is increasingly needed. Hand gesture recognition (HGR) can be used to control hand prostheses, judging from the similarity in shape of objects/objects that tend to make the same hand movements. This development uses three common types of movement, namely pinch, pick, and grab. Developing a neural network model capable of implementing this concept is necessary. The neural network model developed uses the YOLOV7 and YOLOV7 tiny pre-trained networks with datasets collected through the public image data scrapping method. The dataset is 317 images and 2278 object labels with a training ratio of 80:20 testing. The training process uses the Pytorch framework with 300 epochs. The results of the loss values for each epoch show that the model is trainable in the given dataset. The training results are then evaluated by evaluating number of parameters, frame per seconds (FPS) and mean average precision (mAP) using a testing dataset. The overall results show the highest evaluation metrics in the model, with YOLOV7 pretrained with parameter number of 36,9 million, FPS of 161, and mAP of 98,11%. The model has the potential to be developed and implemented as a support for the control functionality of hand prostheses.
Copyrights © 2023