Enhancing semantic segmentation for autonomous vehicle scene understanding in indian context using modified CANet...

Area of Science:

• Computer Vision
• Artificial Intelligence
• Autonomous Systems

Background:

• Traditional computer vision methods face limitations in accurately segmenting complex road scenes, crucial for autonomous vehicle navigation.
• The Indian Driving Dataset (IDD) presents unique challenges due to chaotic road conditions, highlighting the need for advanced segmentation techniques.

Purpose of the Study:

• To enhance semantic segmentation accuracy and efficiency for autonomous vehicles in challenging, unstructured driving environments.
• To propose a novel deep learning model, a modified CANet, that addresses the limitations of existing methods.

Main Methods:

• Development of a modified CANet architecture integrating U-Net and LinkNet components.
• Implementation of a Multiscale Context Module (MCM) with three parallel branches to capture diverse contextual information.
• Evaluation using the Indian Driving Dataset (IDD) to assess performance in complex scenarios.

Main Results:

• The proposed modified CANet achieved a mean Intersection over Union (mIoU) of 0.7053.
• The model demonstrated superior efficiency and performance compared to state-of-the-art semantic segmentation models.
• The architecture effectively captures contextual information at multiple scales for improved segmentation.

Conclusions:

• Deep learning-based semantic segmentation offers a promising solution for safe autonomous navigation in complex traffic.
• The modified CANet provides an accurate, efficient, and resilient approach for road scene understanding.
• Further research can leverage this architecture for real-world intelligent transportation systems.