In autonomous driving, detecting small and occluded objects remains a substantial challenge due to the complexity of real-world environments. To address this, we propose RSO-YOLO, an enhanced model based on YOLOv12. First, the bidirectional feature pyramid network (BiFPN) and space-to-depth convolution (SPD-Conv) replace the original neck network. This design efficiently integrates multi-scale features while preserving fine-grained information during downsampling, thereby improving both computational efficiency and detection performance. Additionally, a detection head for the shallower feature layer P2 is incorporated, further boosting the model’s capability to detect small objects. Second, we propose the feature enhancement and compensation module (FECM), which strengthens features in visible regions and compensates for missing semantic information in occluded areas. This module improves detection accuracy and robustness under occlusion. Finally, we propose a lightweight global cross-dimensional coordinate detection head (GCCHead), built upon the global cross-dimensional coordinate module (GCCM). By grouping and synergistically enhancing features, this module addresses the challenge of balancing computational efficiency with detection performance. Experimental results demonstrate that on the SODA10M, BDD100K, and FLIR ADAS datasets, RSO-YOLO achieves mAP@0.5 improvements of 8.0%, 10.7%, and 7.2%, respectively, compared to YOLOv12. Meanwhile, the number of parameters is reduced by 15.4%, and model complexity decreases by 20%. In summary, RSO-YOLO attains higher detection accuracy while reducing parameters and computational complexity, highlighting its strong potential for practical autonomous driving applications.

In autonomous driving systems, fog decreases visibility and contrast, blurs the boundary of objects, and increases scale variations, which results in missed detections. To solve these problems, we introduce Fog-YOLO, a lightweight real-time detector based on the YOLOv12 framework. First, we design the A2C2f-FSA module, which enhances the representation of fog-affected areas and low-contrast objects by modeling long-range dependencies in the frequency domain. This module effectively suppresses the interference of fog and background noise while maintaining low computational overhead. Second, we propose a bidirectional feature fusion module (BFFM) that uses decoupled attention paths to fuse deep semantic features and shallow texture details. This approach enhances robustness across multiple scales, ensuring the capture of fine-grained texture information and the preservation of global contextual information in foggy environments. Third, we introduce GSConv, which reduces parameters and computational cost by balancing spatial correlation modeling and computational complexity, optimizing the feature extraction process. Finally, we design the F-WIoU v3 loss function, which optimizes bounding box regression through dynamic focusing and difficulty re-weighting strategies, thereby reducing the influence of low-quality samples while improving the model’s localization robustness in foggy conditions. Experiments on the RTTS real-world fog dataset and the VOC-FOG synthetic dataset show that Fog-YOLO outperforms the baseline by 5.2% and 7.3% in mAP@0.5 with real-time inference speed. Overall, Fog-YOLO outperforms mainstream lightweight detectors, demonstrating its practical usefulness for autonomous driving in foggy environments.

Rain, snow, fog, and dust can degrade road-scene images, blur fine details, and consequently reduce the reliability of perception systems for autonomous driving. To address this problem, this paper proposes PF-ConvNeXt, an adverse weather recognition model built upon the ConvNeXt architecture. First, a lightweight pyramid split attention (PSA) module is introduced to enable multi-scale feature fusion, so that both global degradation patterns and local texture details can be captured simultaneously. Second, a feature enhancement channel and spatial attention module (FECS) is designed. It adaptively recalibrates features along the channel and spatial paths, thereby suppressing interference from complex backgrounds and noise. Third, during training, Focal Loss is adopted to strengthen learning for hard samples and minority weather categories, alleviating recognition bias caused by class imbalance. Experiments are conducted on a dataset of 5000 images constructed by integrating RTTS, DAWN, and a self-collected rainy-weather dataset. The results show that PF-ConvNeXt achieves 90.16% accuracy, 95.24% mean average precision, and a 92.18% F1-score. It outperforms the ConvNeXt baseline by 4.74%, 5.46%, and 5.95%, respectively, and surpasses multiple mainstream classification models. This study provides an effective recognition framework for robust environmental perception under challenging weather conditions and demonstrates promising potential for practical deployment.

Small-object detection in unmanned aerial vehicle (UAV) imagery remains challenging because target objects often occupy only a few pixels, exhibit weak feature responses, and are easily obscured by complex backgrounds. These aspects significantly limit the effectiveness of end-to-end detection systems. To overcome these limitations and enhance the detection accuracy in challenging UAV settings, this paper proposes MPC-DETR, a Multi-scale Patch Context Transformer that is based on RT-DETR. To begin with, a Local-Global Attention Fusion Module (LGAF) is proposed to capture fine-grained local features and long-range semantic relations of small objects. LGAF enhances feature representation through a lightweight multi-branch synergistic attention mechanism while introducing limited computational overhead. Second, a Dilated Context-Aware Feature Interaction Module (DCFI) is proposed to enhance the discriminative capability of high-level features in cluttered backgrounds and densely populated small-object scenes. DCFI allows more efficient feature aggregation and contextual comprehension through multi-scale contextual modeling and scale-adaptive feature interaction. Third, a Patch-Guided Multi-scale Feature Fusion Module (PGMFF) is developed to create a patch-guided contextual fusion approach that combines shallow, high-resolution features with deeper semantic information. This process improves the maintenance and representation of fine object information and minimizes information loss in feature propagation. The experimental results on the VisDrone2019 dataset show that MPC-DETR has an mAP50 and mAP50–95 of 52.5% and 33.3%, respectively, which are 4.6 and 4.0 percentage points higher than the baseline model. Further analyses of the UAVDT and HIT-UAV datasets also support the high generalization potential of the suggested method to various UAV-based small-object detection problems. In general, the findings suggest that MPC-DETR provides precise, strong, and efficient small-object detection in complicated UAV images.

Snowflake occlusion, reflections from accumulated snow, and low-contrast imaging in snowy environments can degrade target features and impair road-object detection performance. To address these challenges, this paper proposes SFE-Mamba, a snowy-weather object detection network built on Mamba-YOLO. First, a Snow Spatial–Frequency Feature Extraction Stem (SSF Stem) is designed to jointly model target features through dual spatial- and frequency-domain branches. The frequency-domain information is decomposed into low-, mid-, and high-frequency components and adaptively weighted, thereby preserving target structural information while suppressing frequency disturbances caused by snow particles, snow cover reflections, and illumination variations. Second, a Snow-Aware Mamba Attention Block (SMA Block) is developed by integrating local structural modeling, two-dimensional selective scanning, and Multi-Dimensional Perception Self-Attention (MPSA). This design strengthens the interaction between local details of weak targets and long-range contextual information, improving feature representation under occlusion and low-contrast conditions. Finally, Focus Convolution (FConv) is introduced to reduce the loss of target edges, textures, and positional information during scale transformation. Experiments were conducted on the real-world snowy dataset Snow-ACDC and the simulated snowy dataset Snow-VOC. SFE-Mamba achieved mAP50 values of 74.2% and 71.8% on the two datasets, respectively, outperforming the Mamba-YOLO baseline by 3.6 and 3.9 percentage points. Compared with CF-YOLO, a detector specifically designed for snowy scenes, SFE-Mamba further improved the mAP50 by 1.6 and 1.8 percentage points on the two datasets, respectively. The model contains 6.8 M parameters and reaches an inference speed of 58 FPS. The experimental results demonstrate that the proposed method effectively improves object detection accuracy in complex snowy road scenes while maintaining a favorable balance between detection performance and computational efficiency.

Small objects in UAV imagery often exhibit weak visual representations, suffer from background clutter, and receive insufficient cross-scale feature interaction. To address these challenges, we develop MCI-DETR, a multi-level contextual interaction detector built upon RT-DETR. First, a Local-Context Modulated Interaction Module (LCMI) is designed to selectively modulate contextual features using local structural responses, thereby suppressing irrelevant background responses while expanding the effective receptive field. Second, CATM is incorporated into the high-level feature interaction stage to form a Context Interaction Module (CCIM), strengthening target-related information through spatial- and channel-domain context mapping and additive Query–Key interaction and reducing the interference of complex backgrounds with the high-level semantic representation of weak small objects. Finally, a Multi-Scale Adaptive Gated Fusion Module (MSAGF) is designed to dynamically adjust the fusion contributions of information at different scales according to the input features, while improving the quality of cross-scale feature fusion through gated coordination between local details and global context. On VisDrone2019, MCI-DETR surpasses RT-DETR by 5.2%, 4.8%, and 3.2% in mAP50, mAP50-95, and AP_S, respectively, demonstrating enhanced robustness for small-object recognition in cluttered aerial imagery. MCI-DETR also achieves competitive performance on UAVDT and HIT-UAV against representative detectors, demonstrating the effectiveness of the proposed design under different UAV scenes and imaging conditions. Overall, MCI-DETR strengthens the representation of small objects in aerial scenes and maintains reliable detection performance in the presence of cluttered backgrounds.