HMPACN-YOLO: Traffic Signal Head Detection in Complex Road Scenes via YOLO11 and Hypergraph Computation
Traffic signal head detection in complex road scenes is challenged by variations in target scale, progressive degradation of fine-grained features, interference from visually similar light sources, and limited aggregation of contextual information across spatial locations. To address these challenges, we propose HMPACN-YOLO, a traffic signal head detector built on YOLO11n. The name combines Hyper-MFM (HM), CSP-PMSFA (P), ADown (A), C2TSSA-DyT (C), and NWD-Focaler-WisePIoUv2 (N). Using YOLO11n as the baseline, the method jointly optimizes feature representation, information preservation, bounding-box regression, and model compression. First, Hyper-MFM combines hypergraph propagation with adaptive feature fusion to aggregate contextual information and model higher-order dependencies among spatial feature nodes. Second, multi-scale feature extraction, detail-preserving downsampling, and adaptive attention are integrated to mitigate the loss of fine-grained information during downsampling and suppress spurious background responses under uneven illumination. Third, an NWD-Focaler-WisePIoUv2 composite loss is designed to improve robustness to slight localization deviations of signal-head targets and enhance bounding-box regression across samples with a broad quality distribution. In addition, cross-granularity transfer learning is used for initialization, and Layer-Adaptive Magnitude-based Pruning (LAMP) with subsequent L2 knowledge distillation is applied to compress the enhanced model. Experiments on the self-constructed LTLD dataset use 856 validation images for configuration selection and a held-out test split of 1712 images for final evaluation. On the test split, HMPACN-YOLO achieves 83.2% [email protected] and 41.2% [email protected]:0.95, compared with 81.3% and 39.4% for YOLO11n, corresponding to differences of 1.9 and 1.8 percentage points in a single training run per configuration. At 640 × 640 input resolution, LAMP pruning reduces the enhanced model from 7.8 to 4.8 GFLOPs, compared with 6.3 GFLOPs for YOLO11n; L2 distillation retains the pruned inference architecture. This final configuration represents the combined architecture and compression pipeline. The reported performance gains were observed under the evaluated settings; their consistency across independent training runs remains to be assessed. No wall-clock latency or frame-rate measurement was performed, so the reported complexity figures describe arithmetic cost only.
Authors
- Yutian Wu
- Yaokun Ku (ORCID: https://orcid.org/0009-0002-1150-4354)
- Bailiang Lu
- Xiaoxue Cao
- Chenao Zhang
- Wenyan Han
- Zhi Chen
Institutions
- North China University of Technology (CN)
- University of Science and Technology Beijing (CN)
Publication Details
- Journal
- Applied Sciences
- Published
- 2026-11-18
- DOI
- https://doi.org/10.3390/app162211435
- Primary Topic
- Advanced Neural Network Applications
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- National Natural Science Foundation of China