Research on multitask framework optimization of A-YOLOM for real-time autonomous driving perception

Multitask learning has the advantage of simultaneously handling multiple tasks, significantly reducing computational costs. However, when faced with real-time autonomous driving perception systems, multitask frameworks still suffer from issues such as insufficient detection accuracy. We propose a multitask network in the article that can simultaneously handle three tasks: object detection, drivable area segmentation, and lane line segmentation. By adopting an end-to-end multitask model with a unified segmentation structure and introducing learnable parameters, the same loss function is used for all segmentation tasks, eliminating the cumbersome process of constantly modifying the loss function and enhancing the model's generalization ability. In addition, the network incorporates downsampling techniques, attention mechanisms, and inverted block structures. Comparative experiments on the BDD100K dataset demonstrate that the proposed method is competitive. Our multitask framework achieves a mAP50 of 77.7% in object detection, a mean intersection over union of 90.9% in drivable area segmentation, and an intersection over union of 27.5% in lane segmentation, demonstrating its application value in the field of multitask learning.