融合 SSD 目标检测的自适应手势分割方法

Hand gesture segmentation based on the combination of motion information and skin color features is the most popular method of hand gesture segmentation at present. This kind of method is easily affected by light changes, background changes, motion trajectory overlap and other factors, and the adopted skin color model has no adaption for different skin color gestures. So, this paper proposes a method to roughly detect hand gesture first and then segment hand adaptively. SSD (Single Shot Multi-Box Detector) is improved into a hand gesture detection model, where a backbone network based on dilated convolution and a set of Anchor setting scheme are designed, which can preliminarily segment the gesture ROI (Region Of Interest) to avoid the influence of skin color background on the gesture segmentation. Then, the YCrCb gaussian skin color model was built according to the gesture ROI, so that the skin color model could adapt well to different skin colors of gestures. The experimental results demonstrate the robustness and effectiveness of proposed method for gestures with different skin color in a variety of complex scenarios.