IndiRTS: Real Time Segmentation for Autonomous Vehicles for Indian Conditions

IndiRTS: Real Time Segmentation for Autonomous Vehicles for Indian Conditions

연구 분야: Artificial Intelligence

학회: SN Computer Science

In this paper, we present a lightweight semantic segmentation model IndiRTS, specially designed for real-time inference on autonomous vehicles. Real-time and high-performance segmentation is essential for autonomous vehicles, particularly for navigation and safety in complex urban environments like those in India. Traditional convolutional neural network (CNN)-based models for semantic segmentation often face significant computational challenges, making real-time processing difficult on resource-constrained hardware typical in autonomous vehicles. Our model IndiRTS incorporates an efficient downsampling module that preserves spatial information while reducing computational load and a novel dense link module that captures high-level semantic details while maintaining spatial accuracy. Experimental results show that IndiRTS surpasses state-of-the-art networks on the IDD dataset, achieving a mean IoU of 0.6064. The model contains only 0.45 million parameters allowing it to operate on typical autonomous vehicle hardware with downsampled frames.