Gesture Recognition System Based on Improved Stacked Hourglass Structure

Abstract

Image-based human gesture recognition technology has been extensively studied. However, in the actual application of human gesture recognition, when the complex reality environment is encountered, the recognition accuracy and real-time performance of the existing methods are degraded. This paper proposes an improved Stacked Hourglass structure for human body gesture recognition, which further reduces the weight and number of hyperparameters of the Stacked Hourglass structure without reducing the recognition rate. In addition, after each convolutional layer in each Hourglass subnet, part of the weighting parameters is discarded according to a certain probability, further reducing the computational cost. Experimental results based on the human body pose data set show that this improved method has certain advantages in accuracy, and the computational cost is reduced due to its structural design.