A multitask perception network for autonomous driving based on task-adaptive feature allocation and multiscale attention

Aiming at the defects of feature sharing mechanism and insufficient spatial context modeling in the existing multi-task perception networks for autonomous driving, this paper proposes a multi-task joint perception network (DFA-Net). Through the collaborative design of cross-stage fast connection and multi-scale dilated convolution, the backbone reduces the computational complexity and expands the receptive field feature extraction. The BSAM attention module is designed in the neck, which combines local and global channel information, and fuses multi-scale spatial features to significantly enhance the feature representation ability. In terms of loss function, a weighted combination of boundary-aware loss and Dice loss is used to effectively improve the quality of edge segmentation through boundary distance field constraint and class balance optimization. Experiments on the BDD100K dataset show that compared with the YOLOP model, DFA-Net improves the mIoU of drivable area segmentation by 6.36%and the Accuracy of lane detection by 8.85%while maintaining real-time performance. Compared with other mainstream single-task and multi-task models, DFA-Net shows better accuracy and real-time performance