SDG-YOLOv8: Single-domain generalized object detection based on domain diversity in traffic road scenes

Abstract

Object detection is a fundamental task of environment perception in traffic road scenarios, and its accurate detection results are of great significance for improving the reliability of autonomous driving, improving public transportation services, and detecting traffic violations. However, the problem of domain offset between different traffic road scenarios leads to a poor generalization of the target detector. To overcome this problem, we propose a single-domain generalized object detection algorithm SDG-YOLOv8 based on domain diversity. First, we designed a local–global transformation module to transform the source domain into an auxiliary domain with the same annotations, increasing the domain diversity of the training data at the image level. Second, we design a normalization perturbation fusion module to implicitly change the input image’s style and increase the training data’s domain diversity in the feature space. Finally, we design an effective training loss function that further reduces the sensitivity of the detection model to domain offsets and improves the generalization ability of the target detector to access the unknown target domain. We conducted experiments on multiple datasets containing different weather, different cities, and virtual-to-reality, and our method significantly improves the detection accuracy for unknown target domains and outperforms other domain generalized object detection algorithms.