Multi-task Perception Network for Autonomous Driving Based on CSPDarkNet and Attention Mechanism

Abstract

Aiming at the problems of environmental perception accuracy and multi-task collaborative processing in the process of autonomous driving, this paper proposes an Autonomous Driving Multi-Task (ADMT) perception model based on Cross Stage Partial DarkNet (CSPDarkNet) and attention mechanism. This model shares features and combines information by creating a feature decoupling and fusion module and optimizing the loss function design. This reduces the complexity of training and improves the collaborative effect among tasks. Dynamically weighting the spatiotemporal features of different tasks improves the performance of target detection, instance segmentation, and target tracking. The experimental results showed that, compared with YOLOv4 and YOLOv5, ADMT has achieved significant performance improvements on the KITTI and Cityscapes datasets. Among them, on the KITTI dataset, the F1score of the model reached 0.94, and APiou=0.5 was 0.92. The corresponding values of YOLOv4 and YOLOv5 were 0.93 and 0.91, respectively. These results indicate that ADMT effectively enhances the accuracy and efficiency of target recognition for autonomous driving systems in complex environments, providing strong technical support for future intelligent transportation systems.