Webvr human-centered indoor layout design framework using a convolutional neural network and deep Q-learning

Abstract

With the rapid development of Web Virtual Reality (WebVR) technology, increasing focus has been placed on this domain. WebVR indoor scenario design studies have been of important value in both academia and industry. However, many bottlenecks still need to be overcome, such as the weak computing capacity and limited memory space available in web browsers. In particular, there are many deficiencies in virtual scenes, i.e., lack of fidelity, low automation and poor scene interactability. In this paper, we propose a novel WebVR indoor furniture layout design framework to enhance the capabilities of automatic furniture layout and significantly enhance the interactiveness of virtual scenarios. In particular, we present a hand-drawn sketch recognition scheme based on the ResNet convolutional neural network (CNN), which can strongly improve scenario interactiveness by allowing the user to conveniently add new furniture by means of free-hand drawing operations rather than tedious manual drag-and-pull operations. In addition, based on a deep Q-learning network (DQN), the best positions (states) for these pieces of furniture (agents) in virtual scenarios can be automatically determined, making it easy to satisfy popular design principles. Finally, we report experiments conducted to validate the feasibility of our proposed framework, and the results fully demonstrate that this framework is completely feasible.