Head Pose Prediction While Tracking Lost in a Head-Mounted Display

Abstract

Head-mounted displays are being used nowadays in various industries like defence, training pilots and many more. With these advancements, there is always a need to increase the area of tracking based on cameras or sensors. However, there will still be a limit on the number of cameras on the device for inside-out tracking and number of sensors for outside-in tracking. The proposed deep neural network is a cheaper and easy-to-adopt way for this pose estimation. The major problem of tracking loss occurs in outside-in tracking and is a common issue. In this paper, the proposed approach handles the tracker loss and efficiently predicts the position faster in less than 1 ms. The method exploits the users’ behaviour and game environment to follow a particular trace when playing, users can relate it with their gaming patterns. The approach has been tested with NVIDIA’s VRWorks SDK for real-time behaviour and it works very well within the timing constraints. Our network achieves an accuracy of nearly 96% with 30% data loss and is better than SHNM and BPNN discussed in [2].