How Smooth is my Ride? Detecting Bikeway Conditions from Smartphone Video Streams

In many countries, bicycling has emerged as a viable alternative to motorized means of transport. Citizens rely on bicycles to commute to their workplaces, transport goods, and use them for sports and leisure activities. Available maps are, however, often scarce of information with relevance for cyclists. Besides the presence of tracks, their intersections, and approximations of their inclinations (through contour lines), little further annotations are available. In particular, the surface type of a track (e.g., asphalt, cobbled paving, or soil) is rarely provided, despite the fact that it determines how easily the track can be passed in diverse weather conditions. Cyclists will often only discover the exact track conditions by the time they pass it (or are unable to pass due to it being washed out or flooded by rain). In this work, we present SURF, a pervasive computing application which allows to detect a track's surface type using an opportunistic bicycle-centric sensing system. SURF relies on the processing of images (collected using a handlebar-mounted smartphone) by means of machine learning tools. We evaluate SURF using more than 67,000 training images collected during actual bicycle rides, and show how the system can determine five major surface types of bikeways at an accuracy of 99.51 %.