Mask-based anomaly segmentation in complex driving scenes

Abstract

Road anomaly segmentation plays a pivotal role in advancing the safety of autonomous driving by facilitating the detection of unknown objects in complex traffic environments. Nevertheless, traditional semantic segmentation models, limited by predefined categories, often struggle to accurately identify anomalous objects. In this study, we propose AnomaskDrive, a mask-based anomaly segmentation approach that integrates a comprehensive mask-based attention mechanism and a mask refinement strategy within an RbA framework, enhancing the detection of anomalous objects in complex scenes. The proposed mask-based attention mechanism effectively distinguishes between foreground and background regions, thereby enhancing the segmentation of anomalies in cluttered road environments. Additionally, the inclusion of a mask refinement strategy minimizes false positives and elevates overall segmentation accuracy, demonstrating the robustness and effectiveness of our method. Benchmark evaluations on Road Anomaly and Fishyscapes Lost &Found datasets demonstrate that AnomaskDrive outperforms existing methods, achieving AUC/AP/FPR@95 scores of 98.56%/90.87%/4.68% and 97.75%/74.54%/4.25%, respectively, underscoring its competitive advantage in anomaly segmentation.