Optimized graph extraction and locomotion prediction for redirected walking

Redirected walking with advanced planners such as MPCRed or FORCE requires both knowledge about the virtual environment - mostly in the form of a skeleton graph representing the virtual environment - and a robust prediction of the user's actions. This paper presents methods for both parts and evaluates them with a number of test cases. Since frame rate is crucial for a virtual reality application, the computationally heavy extraction and preprocessing of the skeleton graph is done offline while only parts directly linked to the user's behavior such as the prediction are done online. The prediction is done using a target-based long-term prediction and the targets are determined automatically and combined with targets predefined by the designer of the virtual environment. The methods presented here provide a graph that is well suited for planning redirection and allows prediction techniques previously only demonstrated in studies to be applied to large scale virtual environments.