Architecture and Parameter Analysis to Convolutional Neural Network for Hand Tracking

Abstract

Currently, the hand tracking based on deep learning has made good progress, but these literatures have less influence on the tracking accuracy of Convolutional Neural Network (CNN) architecture and parameters. In this paper, we proposed a new method to analyze the influence factors of gesture tracking. Firstly, we establish the gesture image and corresponding gesture parameter database based on virtual 3D human hand, on which the convolutional neural network models are constructed, after that we research some related factors, such as network structure, iteration times, data augmentation and Dropout, etc., that affect the performance of hand tracking. Finally we evaluate the objective parameters of the virtual hand, and make the subjective evaluation of the real hand extracted in the video. The results show that, on the premise of the fixed training amount of the hand, the effect of increasing the number of convolutional cores or convolution layers on the accuracy of the real gesture is not obvious, the data augmentation is obvious. For the real gesture, when the number of iterations and the Dropout ratio is about 20%–30%, good results can be obtained. This work provides the foundation for future application research on hand tracking.