This paper describes a prototype system that learns to associate driver behavior with traffic controls. The system combines adaptive and hand-crafted components in a way typical of many industrial applications. The adaptive component is a supervised neural network trained to identify controls on individual road segments using measures of driver behavior as input. We preferred accurate knowledge extraction over complete area coverage during system design because applications that affect driver safety may employ the system[1]. The hand-crafted component increased accuracy by using rules about roads and intersections to reject suspicious results.
We intend our system to locate traffic controls and detect anomalous driver behavior. The system can serve to improve many telematic services, such as trip information and safety services[2]. One application is finding the locations of traffic controls and augmenting digital maps with these locations. Another application is detecting traffic disruptions by comparing expected behavior, as defined by controls stored in the augmented map, with dynamically measured behavior.
The adaptive and hand-crafted components of the system are described below in §2. The prototype system described here was designed to identify those controls related to stopping, namely stop signs, traffic lights, and clear intersections. System performance was excellent as discussed in §3. We discuss possible future work and applications in §4. Finally, we summarize our contribution in §5.