Real-time Wall Outline Extraction for Redirected Walking Christian Hirt, Markus Zank, Andreas Kunz ETH Zurich 1 Introduction Redirected Walking (RDW) is used in immersive virtual reality applications to allow a user the exploration of virtual environments which are larger than the available physical space while freely walking. In order to achieve this virtual extension of the space, so called redirection techniques are applied which scale the mapping between the user’s physical and virtual movement. When a user approaches the physical boundaries, e.g. walls, safety measures are activated in order to avoid a collision. However, each safety measure itself is an immense intrusion into the virtual experience, which breaks the immersion for the user. Therefore, Nescher et al. [1] designed a Model Predictive Controller serving as a path planner. Based on the physical layout of the environment and the current user position and orientation the planner determines which redirection technique need to be applied such that potential collisions can be avoided. So far, for such a path planner to be used effectively, it was necessary that the physical environment was known beforehand. This required information was entered into the system manually by measuring the dedicated space defined by the tracking system. Accordingly, this dependency on the tracking space renders the system inflexible and immobile. Thus, a new approach is aiming to eliminate the environmental dependency by adding a Simultaneous Localization and Mapping (SLAM) device, which allows the system to be used ad hoc as it tracks the user’s movement while creating a map of the tracking space at the same time. 2 Methodology As proposed by Nescher et al. [2], the device newly added to the system in this approach is the Google Tango Tablet which uses SLAM technology to track its own position in a continuously recorded environment by fusing various different sensor measurements. The tablet is attached to the front side of an HMD such that the field of view of the user and the tablet are aligned. However, even though the tablet continuously records its surroundings during the exploration, the resulting map cannot be utilized in the path planning or the safety algorithms due to its special localization properties. In order to generate a suitable representation, we access the recorded data of the tablet’s depth sensor. This data consists of a cloud of 3D points found using the reflection of infrared rays and a time-of-flight analysis. In this first approach, we show a simple way of building a representation of the surroundings in real-time by indicating the physical limitations of the room applying a sequence of algorithms on the tablet. 1. Since human locomotion in redirected walking is usually constrained to a single plane, the 3D recordings are projected onto a 1x1cm discretized grid which is parallel to the floor. During this projection, all measured data points belonging to the floor (categorized by the previously identified floor height using a floor matching algorithm provided by Google’s SDK) are eliminated such that only obstacles and walls remain in the resulting 2D point cloud. 3. In a next step, a line extraction algorithm is applied to the data set, which reduces the discretized filtered 2D point cloud to line segments, each representing a part of a vertical structure. Due to its trade-off between accuracy and computational effort, the algorithm chosen here is a Random Sample Consensus (RANSAC). Applying the RANSAC algorithm yields 40 line segments on average for our evaluation space (~25m2). However, the resulting line segments tend to be very short and overlap at certain locations. 2. A confidence filter is added, which is based on the assumption that projecting mainly vertical structures, e.g. walls, along the gravitational vector will result in regions with a higher point density. Accordingly, the confidence is an integer assigned to each grid cell, which is increased with each projection onto a particular cell. Using a threshold value, the cells can further be divided into low and high confidence cells. By eliminating all low confidence cells, a data set which is mostly free of noise even obstacles is obtained. 4. In order to finalize the wall outline extraction, a simple alignment algorithm is used to combine close or overlapping lines, if the segments are considered to be part of the same physical structure. This evaluation is based on the smallest perpendicular distance between the lines and their orientation. Whereas an ideal representation describes a rectangular room with only 4 line segments, the described algorithms manage to achieve 7 line segments on average. 3 Discussion References In this poster, an algorithm was presented which is able to reliably extract the outline of vertical structures (e.g. walls or closets) of the environment in real-time. The algorithm excludes small visible surfaces and horizontal objects, which could be mapped using a supplementing algorithm such as an occupancy map. The resulting map, consisting of a set of 2D coordinates, would be further developed into a suitable input for the safety measures and the path planner algorithm. NESCHER, T., HUANG, Y., KUNZ, A., 2014, Planning redirection techniques for optimal free walking experience using model predictive control. In 3D User Interfaces (3DUI), 2014 IEEE Symposium on. IEEE, p.111-118. NESCHER, T., ZANK, M., KUNZ, A., 2016, Simultaneous mapping and redirected walking for ad hoc free walking in virtual environments. In IEEE Virtual Reality Conference, 2016 IEEE, p.239-240. Contact hirtc@ethz.ch