Event-based hand shadow recognition with varied light intensity and background subtraction

In this paper, we propose the use of event-based hand shadow images for the hand gesture recognition problem and we aim at a portable deep learning shadow-based detection application. Such an interaction-based application requires fast sensing, and limited data transmission (since classification is performed at the edge or back-end server). In addition, it needs to easily adapt to different testing environments with varied lighting conditions and environment backgrounds. In order to overcome these limitations, we introduce an image pre-processing step based on special features of event-based cameras, which reduces the information contained in ’gray-scale’ hand shadow images that are needed for classifying gestures, and at the same time reduces the impact of light illuminations and background interference while maintaining a high classification accuracy.