Experimental Technology and Management

[Objective] This study aims to address the critical challenges associated with recognizing small traffic signs in autonomous driving scenarios. These challenges are particularly pronounced in highly dynamic and visually cluttered environments and adverse meteorological conditions, as well as when the signs are far away. [Methods] To achieve enhanced recognition performance, this study adopted an improved systematic detection method based on the YOLO v11n architecture. In particular, an optimized framework, referred to as YOLO-BSNM, was constructed for complex scenarios by integrating multichannel attention(MCA) with a channel–height–width three-branch dynamic fusion mechanism, a normalized weighted loss(NWDLoss), a 160×160-pixel high-resolution small-object detection head, and a Bi FPN feature fusion module. To enhance the discriminability of features of small objects, MCA employed a channel–height–width three-branch dynamic fusion mechanism. NWDLoss addressed the localization sensitivity issues of traditional IoU-based losses via probabilistic distribution matching. The small-object detection head incorporated a sampling–fusion–extraction layer to mitigate detail loss resulting from resolution decay in deep features. Finally, the BiFPN feature fusion module performed the weighted fusion of multiresolution feature maps to reduce the critical feature loss. [Results] The experimental results demonstrated that the improved YOLO-BSNM algorithm achieved a precision(P) of 81.7%, a recall(R) of 75.4%, and an mAP₅₀ of 83.4% when using the custom dataset, indicating substantial advancements over the baseline algorithm. Concurrently, the model size was smaller by 0.62 million parameters through lightweight optimization, achieving higher detection accuracy with fewer computational resources. The results also indicated that the framework showed enhanced adaptability to challenging scenarios, such as environments with blurring and occlusions. The edge-device deployment capability of the algorithm offers a robust foundation for fulfilling the real-time detection needs of intelligent transportation systems. [Conclusions] This study successfully developed an efficient and accurate traffic sign identification technology based on the YOLO-BSNM model. This technology overcame the limitations of traditional recognition methods and showed improved efficiency and precision in recognizing traffic signs. It provides substantial support for advancing autonomous driving. Deployment and performance verification on an embedded Raspberry Pi 4B platform demonstrated that the optimized model achieved a good balance between architecture lightweightness and detection accuracy, effectively reducing the false alarm and missed detection rates. This advancement meets the objective of enhancing performance while minimizing parameter overhead. YOLO-BSNM accomplished simultaneous enhancements in micro-object feature discrimination and multiscale information integration through the optimized combination of an attention mechanism and a feature fusion strategy. The probabilistic distribution matching–based loss function enhanced the stability of bounding box regression. These technical innovations led to an efficient detection framework, providing a novel approach for detecting micro-objects in complex environments. The lightweight detection system represents a promising solution for deployment in resource-constrained scenarios, such as drone-based remote sensing and satellite image analysis. The integration of multimodal data from infrared sensors, LiDAR, and dynamic environment adaptation algorithms enables the model to overcome detection challenges under variable lighting and extreme weather conditions. This research provides a technical framework for autonomous driving applications and elicits new avenues for investigating universal micro-object detection solutions in computer vision.