Real-Time Gesture Recognition on Mobile Devices Using Deep Learning with Fully on-Device Inference

Real-time hand gesture recognition on mobile devices presents significant challenges due to the necessity of precisely capturing dynamic motions over time while minimizing latency, memory usage, and energy consumption. To attain high accuracy for deployment on mobile platforms, this research introduces a computationally efficient, lightweight, and on-device gesture detection system. The proposed system has a temporal shift mechanism that doesn't need any parameters, a compressed spatio-temporal attention module, and a convolutional backbone that works well on mobile devices. This lets it show motion accurately without using recurrent networks or 3D convolutions. Feature calibration using attention improves motion indicators that can tell the difference between different spatial, channel, and temporal dimensions. At the same time, redistributing temporal resources across channels makes it easier to be aware of time. Before it is put into use, the model is trained with quantization-aware optimization to make sure it will work with mobile neural accelerators. Testing was done on a large gesture dataset to show that the proposed method works well when mobile inference is limited, gets more than 98 % accuracy in classification, and is better than other lightweight and temporal models.