A feature cascade and recursive fusion architecture for traffic sign detection in vehicle perception

In driving scenarios, traffic sign detection technology frequently suffers from reduced detection accuracy of models due to practical challenges such as tiny object, scale variation, and fluctuating lighting conditions. Specifically, the representational information of tiny traffic signs becomes progressively blurred or even entirely lost as the receptive field expands. Furthermore, existing feature fusion networks place greater emphasis on information integration for conventionally sized objects, thereby exacerbating the insufficiency of fused feature information for small objects. To address the aforementioned issues, we propose an end-to-end traffic sign detection algorithm, termed EMB-RFNet, designed to improve the detection accuracy of small traffic signs. It aggregates multi-scale fine-grained object information by designing a recursive path fusion network. Concurrently, it employs an efficient multi-branch cross-stage partial network to enhance multi-scale feature capture, while utilizing a shallow feature supplementation mechanism to compensate for semantic information loss in small objects. We validate the effectiveness of EMB-RFNet through experiments on the TT100K, GTSDB, and CCTSDB public datasets. Its mAP@0.5 metrics achieve 84.9%, 91.5%, and 84.1% respectively across the three datasets, significantly improving traffic sign detection performance for small objects. This demonstrates the superior capability of EMB-RFNet in traffic sign detection, particularly for detecting small objects.