Research On Deep Learning-based Unsupervised Domain Adaptive Visual Object Detection Method

In intelligent transportation system,visual object detection,as the most inexpensive environment perception technology,has been extensively studied in recent years.To maintain their excellent detection performance,existing deep learning-based visual object detection models often resort to large-scale labeled data for training.However,due to the complexity and variability of the real-world environments,such as changes in terms of weather conditions,lighting conditions and camera configurations,domain shift often exists between the training and practical application scenarios,thereby visual object detection models may suffer from poor generalization and low detection performance when deployed across domains.To this end,this paper leverages Faster R-CNN as the base detector and delves into the solutions for the above problems with the help of unsupervised domain adaptation methods,under the circumstances where there is no need to add extra annotation costs for the datasets of practical application scenarios.As for the problem that existing cross-domain detection models do not take into account the significant intra-category variation and insufficient inter-category discrepancy of diversified category features in the training batch,this paper proposes a domain adaptive detection method based on category structure-guided alignment.To improve the ability of the model to learn category semantic knowledge between two domains and the degree of alignment among category structure features,this paper leads the model to learn deep features with more inter-category separability and intracategory compactness.In particular,the designed inter-category separation regularization module uses a strategy named hard sample mining sampling to maximize inter-category discrepancy in the distance interaction between two domains to maintain the separability of different category features.The designed intra-category clustering regularization module uses a constraint strategy about distance metric between the category features of two domains and the shared category feature centers to minimize intra-category variation to maintain the compactness of the same category features,and allows the network to update the parameters of category feature centers selectively.Additionally,a jointly supervised optimization objective,consisting of base detector loss,inter-category separation regularization loss and intra-category clustering regularization loss,is also proposed.Finally,through cross-domain experiments and analysis,the effectiveness of the proposed domain adaptive detection method based on category structure-guided alignment is demonstrated in cross-domain object detection scenarios.Furthermore,with regard to the issue that existing cross-domain detection models fail to distinguish features between the foreground and background regions effectively,the paper proposes a domain adaptive detection method based on progressively foreground-aware alignment.In order to raise the attention of the model to the foreground regions of interest and the category structure information within the foreground regions,this paper leads the model to achieve the alignment of category structures along the direction of forward propagation of the network gradually.In particular,the devised foreground-aware alignment module,which focuses on foreground regions,uses the region and channel attention mechanisms to transform the alignment focus of the network from the global regions to the foreground one.The devised category-aware alignment module,which concentrates on the category structures,uses a constraint strategy about similarity metric among category structures to transition the alignment focus to the category structures within the foreground region further.In addition,a jointly supervised optimization objective,comprising base detector loss,foreground-aware alignment loss and category-aware alignment loss,is also proposed.Finally,through cross-domain experiments and analysis,the effectiveness of the proposed domain adaptive detection method based on progressively foreground-aware alignment is demonstrated in cross-domain object detection scenarios.