A Multi-Task Autonomous Driving Environment Perception Network Based on CA-YOLOPv8

To enhance environmental perception in autonomous driving and address the shortcomings of single-task methods, this study introduces a multi-task perception network called You Only Look Once for Panoptic Driving Perception version 8 (YOLOPv8). Leveraging Cross Stage Partial Hierarchical Enhancement (CSPHet) and Attention and Convolution Mix (ACmix), the CA-YOLOPv8 model integrates object detection, lane detection, and drivable area segmentation tasks. The underlying architecture of the network is built upon the You Only Look Once version 8 (YOLOv8) framework, further augmented by incorporating the Spatial Pyramid Pooling-Fast (SPPF) component to bolster its feature extraction efficiency. The YOLOv8 framework has integrated the CSPHet module, which has been designed to supersede the Cross Convolution fusion (C2f) component. This enhancement aims to bolster the efficiency of feature extraction and simultaneously curtail computational demands. Furthermore, the ACmix module amalgamates the self-attention mechanism with convolutional processes, adeptly harnessing both global and local feature insights to augment the model's overall performance. Experiments on the BDD100K dataset demonstrate competitive results in the tasks of target detection (mAP50: 81.1%), dividing of drivable area (mIoU: 91.0%) and lane line detection (IoU: 28.8%).