Multitask Driving Perception Method Based on Improved YOLOv8

In the field of autonomous driving technology, the role of the environmental perception system is increasingly important. Despite significant advancements in environmental perception technologies in recent years, there are still issues with missed and false detections in certain situations. To address these challenges, we propose an end-to-end multitask network model based on an improved YOLOv8, designed to solve problems related to traffic object detection, lane line segmentation, and drivable area segmentation in autonomous driving scenarios. We replaced the original backbone network with RepGhostNet to reduce the complexity of computations while enhancing feature extraction capabilities. The attention mechanism incorporates an improved SPPFCSPC module to enhance the weighting of important features and suppress less significant ones, thereby improving inference speed and detection accuracy. The detection part of the network uses one detection head and two segmentation heads to perform scene object detection, lane line segmentation, and drivable area segmentation respectively. Experiments on the BDD100K dataset show that the mAP50 for object detection reached 78.6%; lane line detection achieved an mIoU of 27.2%; and drivable area detection an mIoU of 89.8%. These improvements significantly enhance the environmental perception capabilities of our model in autonomous driving scenarios, laying a foundation for achieving a safer and more reliable autonomous driving system.