Work Zone Detection for Autonomous Vehicles

Due to road maintenance and repair, a work zone can appear dynamically, and it may not be included in the map database. For driving safety, an autonomous vehicle needs to detect the work zone on the fly. However, work zone detection is relatively unexplored due to a lack of standard definitions and data. In this paper, we first formulate the work zone detection problem and its evaluation metrics. We then supplement additional work zone annotations in the autonomous driving dataset nuScenes. Next, we propose a detection pipeline for work zones that allows multi-modality sensor configurations. We implement three baseline algorithms using images, a lidar point cloud and a combination of both. Finally, we evaluate and compare their performance both qualitatively and quantitatively in our experiments. Our experimental results show that the fusion-based approach balances the detection accuracy of lidar-based and the range of camera-based approaches.