Real-time gesture recognition based on feature recalibration network with multi-scale information

Abstract

Gesture recognition is important in human–machine interaction. Current methods for solving gesture recognition have several disadvantages such as low recognition rate, slow speed and poor performance on recognizing multiple targets or long-distance targets in complex environments. In view of the above problems, we propose a gesture recognition approach that can recognize gestures quickly and accurately from complex background. This approach works on a deep convolutional network, which consists of a basic network module for extracting feature information, a squeeze-and-excitation networks for increasing feature channel affinity and a feature pyramid attention module for fusing context information with different scales. To test the proposed approach, we make a data set that contains 3289 images from difference complex scenes. Generally gestures in those images can be generally classified into 16 types. We have uploaded this data set for researchers use. Experimental results demonstrate that the recognition accuracy and speed of the proposed method can achieve 83.45% and 32.2 frames per second respectively, which has better comprehensive performance compared with other state-of-the-art recognition algorithms.