To address common issues in intelligent driving, such as small object missed detection, false detection, and edge segmentation errors, this paper optimizes the YOLOP (You Only Look Once for Panoptic Driving Perception) network and proposes a multi-task perception algorithm based on a MKHA (Multi-Kernel Hybrid Attention) mechanism, named MKHA-YOLOP.