Outdoors, a quick look at a smartphone is enough to find your way around; indoors, this convenience disappears. Pedestrian dead reckoning based on inertial sensors fails precisely where pedestrians need it most — at directional changes — and progress is slowed down as there is too little training data in the real world to train complex models. In our indoor navigation system, URWalking (described in (Ludwig et al. 2023)), we provide routing instructions to many users per day for routes of 500 metres or more. An analysis of the implemented tracking module in (Jackermeier and Ludwig 2018) revealed a 50% drop in PDR accuracy at turns in tight corridors. To improve performance, we collected new multi-modal training data and applied a standard optical flow algorithm to improve turn detection. Our experiments demonstrate that we can predict turntaking with 90% accuracy. We consider this result to be a significant step forward in improving the long-distance tracking capabilities of indoor navigation systems.