DBA-Net: Dual-Branch Asymmetric Attention Network for Multitask Perception in Autonomous Driving Under Challenging Environments

Multitask perception has emerged as a critical technology for autonomous driving systems. Despite significant progress, current systems remain vulnerable to environmental variations, particularly challenging lighting conditions (e.g., low-light scenarios, sudden illumination changes) and adverse weather (e.g., rain, fog, snow), which degrade the visibility of crucial road elements such as vehicles, pedestrians, and lane markings. These environmental factors substantially reduce the available visual information, compromising the accuracy and reliability of multitask perception systems. Moreover, the inherent complexity of real-world driving scenarios further exacerbates the difficulty in effective feature extraction for conventional multitask networks. To address these challenges, we present a novel Dual-Branch Asymmetric Attention (DBA) module that introduces two key innovations: (1) a bidirectional spatial weighting mechanism that dynamically allocates attention by modeling structural disparities between horizontal and vertical dimensions in driving scenes, and (2) an advanced progressive compression architecture that systematically expands receptive fields while preserving local feature details. In a controlled comparison against the YOLOP baseline, our single-run evaluation demonstrates improvements in several perception metrics, achieving absolute percentage-point increases of +0.7 mAP50 and +0.4 recall for object detection, +0.2 mIoU for drivable-area segmentation, and +1.1 accuracy for lane detection. These numerical differences are preliminary single-run observations, and their stability remains to be established through repeated independent training. Qualitative day/night examples illustrate the model outputs and motivate further evaluation of DBA as a lightweight feature-refinement design.