IndiSegNet: Real-time semantic segmentation for unstructured road scenes in intelligent transportation systems

Autonomous driving in developing regions demands perception systems that can operate reliably in unstructured road environments marked by heterogeneous traffic, weak or missing lane geometry, frequent occlusions, and strong appearance variability. Existing semantic segmentation models, although successful in structured Western datasets, exhibit poor generalization to such chaotic conditions and are often too computationally heavy for real-time deployment on low-power edge hardware. To address these gaps, this paper focuses on the challenge of achieving fast, accurate, and resource-efficient segmentation tailored to complex Indian road scenes. We propose IndiSegNet, a lightweight architecture designed explicitly for this setting. The model introduces two novel components—Multi-Scale Contextual Features (MSCF) for capturing irregular object scales and Encoded Features Refining (EFR) for enhancing thin-structure and boundary detail, resulting in a more stable representation for unstructured environments. IndiSegNet achieves 67.2% mIoU on IDD, 78.9% on Cityscapes, and 74.6% on CamVid, while sustaining 112 FPS on Jetson Nano, outperforming standard baselines by 12%–18% IoU on safety-critical classes such as pedestrians, riders, and vehicles. Real-world evaluation across urban, monsoonal, rural, and mountainous regions shows less than 2.5% variance in mIoU with consistent inference speeds above 108 FPS. These results demonstrate that IndiSegNet offers a practical and hardware-efficient solution for high-speed autonomous navigation in the challenging traffic conditions of developing regions. • A novel hardware-aware framework, IndiSegNet, achieves high accuracy on benchmark datasets (78.9% mIoU on Cityscapes) while maintaining real-time performance (112 FPS on Jetson Nano). • The model is specifically optimized for complex driving scenarios, setting a new state-of-the-art performance on the Indian Driving Dataset (IDD) with 67.2. • Extensive real-world deployment across diverse geographical environments, including mountainous, rural, and monsoon-affected urban roads, confirms the model’s robustness with less than 2.5. • The architecture prioritizes safety, improving the Intersection over Union (IoU) for critical classes like pedestrians and vehicles by over 12.