A Traffic Sign Discovery Driven System for Traffic Rule Updating

Traffic rules characterize the relationship of connected roads, and are fundamental data in map. They play an important role in route planning, driver navigation, and autonomous driving system. The traffic rule update is a key ability of cartography since they change day by day in real world. Trajectory mining based method has been proposed to update traffic rule. However, it is usually effected by the trajectory outliers and hard to improve the accuracy. In this paper, we consider using data from a different source, the street images captured by driving vehicle recorders (DVR), and propose a traffic sign driven system to update the rules. To collect candidate traffic rule changes, we train an object detection model to detect the traffic signs in street images. To improve the flexibility of the system, we propose a model compression method to reduce the model size, and integrate it into DVR. Finally, we propose a spatio-temporal attention method to cluster the recognized rules. Our system supports the updating of many types of traffic rules, such as no left/right/u turn, no parking, speed limit, etc., and has high extendibility. We validate our image recognition algorithm in a real world dataset, and it achieves a precision of 99.2% and recall rate of 95.1% for image level output. It demonstrates the advantage of our proposed system.