IndiVNet A region adaptive semantic image segmentation for autonomous driving in unstructured environments.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Autonomous navigation faces challenges in developing regions due to heterogeneous traffic and poor road infrastructure.
• Current semantic segmentation models, trained on structured data, lack generalization capabilities for unstructured environments.

Purpose of the Study:

• To propose IndiVNet, a novel semantic segmentation architecture designed for unstructured Indian driving scenarios.
• To enhance the robustness and scalability of autonomous navigation systems in diverse real-world conditions.

Main Methods:

• Introduced a progressive dilation encoder (6→16) within the IndiVNet architecture to capture multi-scale contextual information effectively.
• Trained and evaluated the model on the India Driving Dataset (IDD) and the CAMVID dataset.

Main Results:

• IndiVNet achieved 69.98% mean Intersection over Union (mIoU) on the IDD, surpassing existing CNN and Transformer baselines.
• Demonstrated strong cross-domain generalization by reaching 73.2% mIoU on the CAMVID dataset.
• The model balances contextual understanding with real-time processing efficiency.

Conclusions:

• IndiVNet provides a scalable and region-aware solution for robust autonomous navigation in complex, unstructured environments.
• The proposed architecture addresses the limitations of current models in diverse driving conditions.
• This work contributes to advancing autonomous driving technology in developing regions.