Enhancing road surface recognition via optimal transport and metric learning in task-agnostic intelligent driving environments

A comprehensive understanding of the environment is essential for ensuring the driving safety of intelligent vehicles. Road surface conditions, e.g., uneven and damaged roads, icy, or slippery surfaces can severely impact driving safety, requiring the vehicle to perceive road surface attributes accurately. Concurrent works have delved into recognizing dynamic or static obstacles with onboard sensors, e.g., Lidar or Camera. However, how to accurately understand road surface conditions remains elusive. Furthermore, collecting and annotating the data from different real-world terrains can be a burden. The recognition model should be able to extract knowledge from different road surfaces with limited data. For robust road surface recognition across varied data distributions in complex environments, we propose a novel multi-task learning (MTL) framework to leverage the information across different road surfaces regardless of the wetness and illumination conditions, especially in scenarios with limited labeled data. By treating road surface type recognition under different levels of road wetness and illumination as separate tasks, we leverage shared knowledge through optimal transport theory. Based on this, we implement adversarial training using Wasserstein-1 Distance to derive task-invariant features. Additionally, previous research has indicated that such methods may lead to ambiguous classification boundaries in the feature space. We thus incorporate a metric learning objective for enhancing the prediction performance by leveraging the feature similarities. To verify the effectiveness of the proposed method, we conduct detailed experiments on the open-source RSCD dataset and our own newly collected MRSD dataset. The results demonstrate the efficacy and superiority of our method, laying a robust foundation for road surface knowledge sharing and boundary discrimination across varied wetness and illumination conditions in intelligent driving applications.