Safety-aware transformer-enhanced unified multi-task perception for road anomaly detection, drivable area...

The developing countries with unstructured road conditions like in India present considerable challenges to autonomous driving systems. The challenges arise from the presence of road anomalies (potholes, waterlogging, cracks) along with unclear lane markings and mixed traffic participants(pedestrians, cars, bikes, autorickshaws), which demand an efficient perception system. To address these challenges, a Safety Aware Multitask Perception model for Indian roads has been proposed that jointly performs object detection, drivable area segmentation and lane line segmentation. The model adapts a lightweight YOLOv8 backbone integrated with a C3TR Transformer module in the shared encoder which enhances the capture of long-range dependencies for better understanding of unstructured road environments. The network utilizes three task-specific decoders to process the shared feature representations, thereby creating an efficient multitask learning while reducing the computational overhead associated with deploying separate single-task models. To explicitly incorporate real-world safety considerations, a deterministic safety-aware refinement logic is introduced that differentiates between critical hazards (such as potholes and waterlogging) that must be excluded from the drivable area and navigable hazards (such as speed bumps) that are retained within the drivable area as alert regions. The network was evaluated on the Custom Multitask Indian Road Dataset (CMIRD) that is collected from two South Indian cities at Thiruvananthapuram and Nagercoil, as well as on the public BDD100K dataset. The proposed model achieves a mAP@50 of 0.877 for anomaly detection providing 1.9% improvement over the A-YOLOM(s) baseline. The drivable area mIoU is 0.962, where a numerically modest 0.3% improvement is observed reflecting the practical significance of the effective removal of unsafe hazards from the drivable region. The proposed method maintains comparable lane detection performance, achieving a lane line IoU of 0.447. Extensive ablation studies, sensitivity studies, comparison with a learnable fusion baseline and the cross-region zero-shot validation on unseen geographic region provide deeper insight into model behavior. Condition-wise evaluation further shows that the framework is robust across structured and unstructured environments, while performance degradation primarily occurs in scenarios involving low-contrast or small-scale anomalies and severe road degradation.