A New Multi-task Network for Autonomous Driving: Efficientnetv1_Unet

Multi-task networks have found widespread applications in the field of autonomous driving, particularly as perception tasks within multi-task learning continue to gain traction. We propose a novel multi-task network aimed at completing several tasks, including object detection, drivable area detection, lane detection, and height-width restriction detection. On the BDD100K dataset, our model achieves Recall at 94.1% for object detection and IoU at 27.7% for lane detection. Model introduces new task heads into a multi-task network using keypoint detection to address the height-width restriction detection. It demonstrates performance improvements compared to previous networks.