YOLOP-MVF: A Multi-Task Autonomous Driving Perception Detection Method Based on Multi Scale Feature Weighted Fusion

To address challenges such as large-scale variations, background interference, and occlusions in multi-task autonomous driving perception, this paper proposes YOLOP-MVF, a multi-task detection framework based on multi-scale feature weighting fusion. The model integrates a sub-pixel 3D fusion module and a triple feature encoding module to enhance the representation of multi-scale features. A multi-scale convolutional attention-weighting mechanism is further introduced to adaptively emphasize critical spatial information. To improve feature extraction flexibility, deformable convolutions are incorporated, enabling dynamic sampling based on input characteristics. Additionally, the Powerful-IoU loss is employed to guide anchor box regression with adaptive penalty and gradient regulation, accelerating convergence. Experimental results on the BDD100K dataset demonstrate that YOLOP-MVF outperforms baseline models, achieving improvements of 1.2% in mIoU, 8.8% in accuracy, and 4.7% in mAP50, validating its effectiveness for robust multi-task perception in complex driving scenarios.