Abstract:Object detection in complex open-world traffic scenarios is a core challenge for autonomous driving and related fields. Current mainstream methods, primarily based on convolutional neural networks and Transformer architectures, address these challenges to some extent by employing complex feature pyramid structures or incorporating knowledge distillation strategies. However, these methods still face significant limitations in systematically handling geometric deformations, scale variations, and continual learning in dynamic environments. To address these limitations, an end-to-end detection network named the dynamic orthogonal projection network (DOPNet) is proposed. This network comprises three core modules: (1) the adaptive deformation-aware module (ADM), which integrates multi-scale deformable convolutions and an attention mechanism to robustly model local non-rigid deformations, thus providing a stable input for feature extraction; (2) the orientation-aware feature enhancement (OAFE) module, in which a scene-aware mechanism is introduced based on traditional orthogonal projection to optimize the feature space through dynamic adjustment of projection bases; (3) the dual-branch decoupled feature pyramid network (D-DFPN) module, which employs a parallel architecture to explicitly separate discriminative features and detailed features, and achieves complementary integration through an adaptive gated fusion mechanism. Experimental results on the BDD100K, COCO-OW and self-built bad weather datasets demonstrate that the proposed method outperforms existing mainstream approaches across several key metrics. Specifically, for known categories, a 3.8 percentage point improvement in mAP over the sub-optimal method (OrthogonalDet) is achieved. For unknown categories, a 5.4 percentage point increase in U-Recall compared to the ablation baseline is achieved. Only a 6.4 percentage point decrease in anti-forgetting performance (ΔOld-mAP) is observed. Additional experiments under adverse conditions, such as dust, rain, snow, and nighttime, further demonstrate that the proposed method maintains stable detection performance while preserving accuracy.