We trained a supervised neural network to identify controls on individual road segments. The inputs to the neural network were statistics about driver behavior on a given segment. In order to compute these statistics, we collected geographic position data from drivers as described below in §2.1.1. We computed road segment data from geographic position data using a digital road map[3]. We used both geographic and road segment data to compute measures of driver behavior related to stopping. As described in §2.1.2, we applied a number of filters to the data at various stages and computed statistics on the data from all traversals, or passes of a vehicle over a segment. Direction matters, so traversals of a given two-way road segment in one direction are kept separate from traversals in the other direction.
The types of controls we chose to identify -- stop signs, traffic lights, clears -- govern stops at intersections. We designed our input representation to reflect stopping behavior. We chose inputs that showed how often and how long drivers stopped. We also designed our representation to support discrimination between multiple short stops and one or two long stops so as to capture the qualitative differences between stops at stop signs and lights.
The neural network had eleven input values. These included the average and the standard deviation of each of the following: number of times stopped, total duration of all stops, the durations of the three stops closest to the end of the segment. We also computed the percentage of traversals that included at least one stop for each segment. Though we did not formally analyze the utility of the inputs, we observed that the most meaningful predictor was the percentage of traversals that included a stop. We trained the neural network to recognize clear road segments (those with no restrictions), segments with stop signs, and segments with traffic lights. We determined ground truth, i.e. the control locations, from direct observation and from maps obtained from the Transportation Department of the City of Palo Alto.