Driver distraction recognition is critical for assisted driving and intelligent vehicles, while often fails in cross-domain scenarios. Previous studies only focus on behavior recognition on a single data set, ignoring the case of inconsistent distribution between training and test sets (domains). To eliminate the domain gap caused by different persons or scenarios, we propose a domain adaptation framework for cross-domain distracted driver recognition, which aims to reduce the domain gap by constraining the feature separability of the source and target domains and maximizing the similarity of the feature structures of the two domains. Firstly, we represent the structure of the domain-level distribution by the spatial relationship between classes in the feature space, where the class is represented by class prototypes to avoid the perturbation of noise samples. Secondly, we utilize the structure consistency across domains by defining the spatial relationship between classes in terms of sample-to-class distances and encourage this relationship to be consistent when adapting from source to target domain. Thirdly, we build a new cross-domain dataset for driver distraction behavior, which contains two public datasets and one self-collected dataset. Experiments demonstrate that the proposed method can effectively alleviate the driver-distracted behavior recognition problem across datasets and obtain distinguishable results compared to existing UDA methods. Code is available at: https://github.com/seizetheday00/Structure_consist_UDA.
Structure Consistent Unsupervised Domain Adaptation for Driver Behavior Recognition
2023-09-24
8350295 byte
Conference paper
Electronic Resource
English
DEEP CLASSIFICATION-DRIVEN DOMAIN ADAPTATION FOR CROSS-MODAL DRIVER BEHAVIOR RECOGNITION
British Library Conference Proceedings | 2020
|Toward Human-Robot Cooperation: Unsupervised Domain Adaptation for Egocentric Action Recognition
Springer Verlag | 2023
|Unsupervised Evaluation of Lidar Domain Adaptation
IEEE | 2020
|