A Preliminary Investigation into a Deep Learning Implementation for Hand Tracking on Mobile Devices

Hand tracking is an essential component of computer graphics and human-computer interaction applications. The use of RGB camera without specific hardware and sensors (e.g., depth cameras) allows developing solutions for a plethora of devices and platforms. Although various methods were proposed, hand tracking from a single RGB camera is still a challenging research area due to occlusions, complex backgrounds, and various hand poses and gestures. We present a mobile application for 2D hand tracking from RGB images captured by the smartphone camera. The images are processed by a deep neural network, modified specifically to tackle this task and run on mobile devices, looking for a compromise between performance and computational time. Network output is used to show a 2D skeleton on the user's hand. We tested our system on several scenarios, showing an interactive hand tracking level and achieving promising results in the case of variable brightness and backgrounds and small occlusions.