Dfrnet-H: Dynamic Feature Refinement Network with Heterogeneous Kernels and Weighted Fusion for Highway Monitoring
Xu Liu, Wei Han, Siren Batu, Peng Zhang, Kai Liu and Ming Ma∗
AI summary
Problem
Highway vehicle detection struggles with large scale variations, motion blur, and frequent occlusions, causing standard YOLO-based detectors to lose accuracy on distant small objects and complex scenes.
Approach
The authors redesign the YOLO architecture with three lightweight modules: CSP-MSPA for enhanced small-object representation, RDA for expanded receptive fields via recursive dilated convolutions, and AFPN for adaptive, weighted multi-scale feature fusion.
Key results
- +4.4% mAP improvement over YOLOv11-N on UA-DETRAC
- +2.5% mAP gain on the new HVD dataset
- Superior accuracy for distant small vehicles and large trucks
- Maintained lightweight architecture with real-time inference speed
Why it matters
Provides a robust, efficient detection solution for intelligent transportation systems and autonomous driving applications operating in challenging highway environments.
Abstract
Highway vehicle detection remains challenging due to scale variation, motion blur, and frequent occlusions. While YOLO-based detectors meet real-time demands, their feature extraction and fusion remain limited in complex traffic scenes. To address this, we propose DFRNet-H (Dynamic Fea- ture Refinement Network), which integrates three lightweight modules: CSP-MSPA enhances small-object representation with fractal partial convolution, RDA enlarges receptive fields through recursive dilated aggregation, and AFPN adaptively reweights multi-scale features for efficient fusion. On the UA- DETRAC benchmark dataset, DFRNet-H achieves a +4.4% improvement over YOLOv11-N on mAP50−95, and on our self-constructed Highway Vehicle Detection (HVD) dataset it achieves a further +2.5% gain. These results demonstrate that DFRNet-H effectively balances accuracy and efficiency under complex highway scenarios.