MTFENet: A Multi-Task Autonomous Driving Network for Real-Time Target Perception

Abstract

Effective autonomous driving systems require a delicate balance of high precision, efficient design, and immediate response capabilities. This study presents MTFENet, a cutting-edge multi-task deep learning model that optimizes network architecture to harmonize speed and accuracy for critical tasks such as object detection, drivable area segmentation, and lane line segmentation. Our end-to-end, streamlined multi-task model incorporates an Adaptive Feature Fusion Module (AF$^{2}$M) to manage the diverse feature demands of different tasks. We also introduced a fusion transform module (FTM) to strengthen global feature extraction and a novel detection head to address target loss and confusion. To enhance computational efficiency, we refined the segmentation head design. Experiments on the BDD100k dataset reveal that MTFENet delivers exceptional performance, achieving an mAP50 of 81.5% in object detection, an mIoU of 93.8% in drivable area segmentation, and an IoU of 33.7% in lane line segmentation. Real-world scenario evaluations demonstrate that MTFENet substantially outperforms current state-of-the-art models across multiple tasks, highlighting its superior adaptability and swift response. These results underscore that MTFENet not only leads in precision and speed but also bolsters the reliability and adaptability of autonomous driving systems in navigating complex road conditions.