Cross-Domain Adaptive Object Detection Based on Refined Knowledge Transfer and Mined Guidance in Autonomous Vehicles

Obeject detection as a fundamental task of environmental perception systems in autonomous vehicles (AVs), is significant for intelligent driving safety that precise detection results improve control accuracy. Facing the cross-domain adaptation scenarios, the generalization capability of the detector is poor due to the lack of target supervised information and the influence of covariate shift among domains especially. To takle such an issue, this article proposes a reliable domain adaptation method based on refined knowledge transferring and mined guidance for cross-domain detection in AVs. Two domain adaptation modules are integrated on baseline detection framework from the perspective of global rough semantic information and local specific content. One module combines the hierarchical domain feature information and transfer network by domain feature refinement to avoid the interference of background information in coarse domain features. The other promotes the alignment and learning of the foreground features in the target domain through the semantic guidance obtained from the potential information mining component and the designed global-local semantic consistency measure. Benefiting from these adaptation methods, the foreground knowledge of source doamin can be effectively transferred to target domain. We conduct extensive experiments on typical cross-domain detection scenarios and the proposed method is proved to outperform previous approaches.