Efficient Edge-Aware Multi-Task Network With Hybrid Attention for Autonomous Driving Perception
Authors:
- 20 March 2026
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 985-989
Abstract
Real-time semantic segmentation of drivable areas and lane markings is a crucial component of autonomous driving perception systems. While existing lightweight models like TwinLiteNet+ achieve remarkable inference speeds, they often struggle with imprecise boundary localization, limited feature discrimination in complex scenarios, and insufficient multi-scale contextual awareness. In this paper, we propose an enhanced multi-task segmentation framework that significantly improves upon the TwinLiteNet+ baseline. First, we introduce an adaptive Hybrid Attention Module (HAM) that dynamically fuses channel and spatial attention using learnable weights, enhancing feature expression in diverse driving scenes. Second, we propose an explicitly differentiable Edge-Aware Loss function with learnable Sobel operators, which automatically emphasizes boundary pixels during training to yield sharper lane and road delineations. Finally, a Lightweight Atrous Spatial Pyramid Pooling (L-ASPP) module built with depthwise separable convolutions is integrated to capture multi-scale context with negligible computational overhead. Extensive experiments on the challenging BDD100K dataset demonstrate that our proposed model achieves state-of-the-art performance-efficiency trade-offs. Compared to the baseline, our model improves Drivable Area mIoU from 92.9% to $\mathbf{9 3. 8 \%}$ and Lane Detection IoU from $\mathbf{3 4. 2 \%}$ to $\mathbf{3 6. 1 \%}$, while requiring only 2.12 M parameters and 18.4 GFLOPs, rendering it highly suitable for resource-constrained edge deployments.