An Ultra-High Speed Gesture Recognition Algorithm Based on MobileNetV2

Abstract

To solve the problem of interference in gesture detection and recognition under complex background, an ultra-high speed gesture recognition method based on a depth model is proposed. Generally, 12.5 frames per second (FPS) can achieve the requirements of real-time detection. This article reaches 2600FPS, which is defined as ultra-high-speed. The proposed algorithm makes up for the defects of the accuracy and robustness in traditional methods, as well as improves the low recognition speed of the general neural network under complex situations. Firstly, a neural network is designed and trained to accurately distinguish three different gestures. Secondly, a pruning and a merging operation is performed on the trained neural network, respectively. Without significantly affecting the detection and recognition results, the network is compressed and the efficiency is improved. Finally, the single-precision floating point data is quantized into integer data to further improve the detection and recognition speed. The experimental results show that the gesture recognition algorithm proposed in this paper can reach 2600FPS under the premise that the accuracy rate is 97.8%.