In intelligent transportation systems, accurate vehicle target recognition within road scenarios is crucial for achieving intelligent traffic management. Addressing the challenges posed by complex environments and severe vehicle occlusion in such scenarios, this paper proposes a novel vehicle-detection method, YOLO-BOS. First, to bolster the feature-extraction capabilities of the backbone network, we propose a novel Bi-level Routing Spatial Attention (BRSA) mechanism, which selectively filters features based on task requirements and adjusts the importance of spatial locations to more accurately enhance relevant features. Second, we incorporate Omni-directional Dynamic Convolution (ODConv) into the head network, which is capable of simultaneously learning complementary attention across the four dimensions of the kernel space, therefore facilitating the capture of multifaceted features from the input data. Lastly, we introduce Shape-IOU, a new loss function that significantly enhances the accuracy and robustness of detection results for vehicles of varying sizes. Experimental evaluations conducted on the UA-DETRAC dataset demonstrate that our model achieves improvements of 4.7 and 4.4 percentage points in mAP@0.5 and mAP@0.5:0.95, respectively, compared to the baseline model. Furthermore, comparative experiments on the SODA10M dataset corroborate the superiority of our method in terms of precision and accuracy.

With the swift evolution of artificial intelligence in the automotive sector, autonomous driving has ascended as a pivotal research frontier for automotive manufacturers. Environmental perception, as the cornerstone of autonomous driving, necessitates innovative solutions to address the intricate challenges posed by data sensitivity during vehicle operations. To this end, federated learning (FL) emerges as a promising paradigm, offering a balance between data privacy preservation and performance optimization for perception tasks. In this paper, we pioneer the integration of FL into 3D object detection, presenting personalized FedM²former, a novel multi-modal framework tailored for autonomous driving. This framework aims to elevate the accuracy and robustness of 3D object detection while mitigating concerns over data sensitivity. Recognizing the heterogeneity inherent in user data, we introduce a personalization strategy leveraging stochastic gradient descent optimization prior to local training, ensuring the global model’s adaptability and generalization across diverse user vehicles. Furthermore, to address the sparsity of point cloud data, we innovate the attention layer within our detection model. Our balanced window attention mechanism innovatively processes both point cloud and image data in parallel within each window, significantly enhancing model efficiency and performance. Extensive experiments on benchmark datasets, including nuScenes, ONCE, and Waymo, demonstrate the efficacy of our approach. Notably, we achieve state-of-the-art results with test mAP and NDS of 71.2% and 73.6% on nuScenes, 67.14% test mAP on ONCE, and 83.9% test mAP and 81.8% test mAPH on Waymo, respectively. These outcomes underscore the feasibility of our method in enhancing object detection performance and speed while safeguarding privacy and data security, positioning Personalized FedM²former as a significant advancement in the autonomous driving landscape.