RD-Hand: a real-time regression-based detector for dynamic hand gesture

Existing gesture-recognition methods have poor performance on the mobile terminal, because of their limited computing resources and noisy environments such as Spatio-temporal variations and light variations. In this paper, we first draw the idea of regression in object detection tasks into dynamic gesture trajectory recognition. We regard gesture trajectory recognition as an object detection task and propose a regression-based dynamic hand gesture detector named RD-Hand, which contains a special full connection layer for trajectory regression. RD-Hand only needs to perform classification and location once at the same time, and the target trajectory can be detected by regression. Firstly, we use a network combined with CNN and LSTM to extract the features of the trajectory. Then we use a full connection layer with special meaning to predict the classification and location. Finally, trajectory-based non-maximum suppression is used to eliminate redundant solutions. Experiments show that RD-Hand has high accuracy, good real-time performance without GPU.