FIF-YOLO: Feature-Interactive Detection Framework for Traffic Signs in Complex Road Environments
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Accurate Traffic Sign Detection and Recognition (TSDR) serves as the cornerstone for safe decision-making in autonomous driving perception systems. However, in real unstructured road scenarios, traffic signs are frequently challenged by severe scale variation and partial occlusion. Existing one-stage detectors based on Convolutional Neural Networks (CNNs) are constrained by the inductive bias of local receptive fields, making it difficult to reconstruct the global semantic topology of occluded targets in the feature space, which leads to missed detection of long-range small targets and false detection of heavily occluded targets. To address these issues, this paper proposes a novel detection architecture named FIFYOLO (Occlusion-aware Intra-scale Feature Interaction YOLO). Built upon the YOLOv8 baseline, this architecture innovatively integrates an Attention-guided Intra-scale Feature Interaction (AIFI) module and a SEmantic Attention-enhanced Detection Head (SEAM-Head). First, drawing on the design paradigm of Real-Time DETR, the AIFI module is embedded into the highlevel semantic feature extraction stage. Leveraging the MultiHead Self-Attention mechanism, the AIFI module can capture global long-range dependencies in the S5 feature layer, effectively suppress background noise and enhance semantic consistency, thus resolving the problem of detailed feature loss during downsampling in the traditional SPPF module. Second, to tackle feature deficiency caused by occlusion, we design a decoupled detection head integrated with SEAM. This module introduces an exponential attention activation function, which can adaptively amplify the feature response weights of occluded regions based on the contextual information of the visible parts of targets, achieving implicit feature completion. Experiments conducted on our self-constructed large-scale occlusion dataset demonstrate that FIF-YOLO achieves an mAP50 of 80.19%, representing an 8.06% improvement over the baseline model. Meanwhile, while maintaining real-time inference speed, it significantly improves the recall rate for extremely small targets and heavily occluded targets.
Fields of Study & Subjects
Advanced Neural Network Applications
Autonomous Vehicle Technology and Safety
Domain Adaptation and Few-Shot Learning
Publication details
- Journal / source
- International Conference on Smart City and Information System (ICSCIS)
- Pages
- 43–47
- Publication date
- 30 January 2026
- Language
- EN