Hand Gesture Recognition Using Movidius Neural Compute Stick

Image classification aims to recognize the object detected to differ between different classes such as human being, vehicles, buildings and et cetera. In classifying any object, it is important to measure the accuracy to acknowledge its effectiveness on certain tasks specifications. Hence, this work is carried out to develop an application which can recognize the hand gestures on embedded system. There are four sorts of hand gestures used in this work which are for zooming in, zooming out, scrolling up and scrolling down. We proceed to measure the accuracy and performance on embedded system using Movidius Neural Compute Stick (NCS). In this work, AlexNet and GoogleNet models are trained in Caffe as it is compatible with OpenVino. The result obtained showed that GoogleNet delivers the best classification accuracy at 99.6%, while the performance of GoogleNet on NCS is acceptable for real-time implementation where it took around 110ms to infer the image. This is slower than CPU inference time which is at 68ms. Nevertheless, the result showed that the trained model can be used in embedded system with good performance.