Frequency-oriented Adaptive Real-Time Object Detector for Cluttered Traffic Scenes

Li, Ziqi; Gao, Tao; Li, Shutao; Chen, Ting; An, Yisheng; Wen, Yuanbo; Lei, Tao

Frequency-oriented Adaptive Real-Time Object Detector for Cluttered Traffic Scenes

Ziqi Li¹, Tao Gao^2,*, Shutao Li^1,*, Ting Chen¹, Yisheng An¹, Yuanbo Wen¹, Tao Lei³

¹School of Information Engineering, Chang'an University, Xi'an 710064, China
²School of Data Science and Artificial Intelligence, Chang'an University, Xi'an 710064, China
³School of Electronic Information and Artificial Intelligence, Shaanxi University of Science and Technology, Xi'an 710021, China
^*Corresponding authors: Tao Gao and Shutao Li

Paper Code Data BibTeX

Traffic objects intra-category similarity, inter-category similarity, and similarity margin.

Abstract

Given the pressing demand for traffic object detection in autonomous driving vehicles, attaining efficiency and accuracy in perception on in-vehicle platforms remains a formidable challenge. To this end, we propose a frequency-oriented adaptive real-time object detector for cluttered traffic scenes. The proposed method explicitly enhances frequency-aware feature extraction and multi-scale adaptive feature fusion, strengthening both detailed texture information and high-level semantic representations. By improving feature representation under complex traffic conditions while maintaining efficient inference, the detector provides a practical solution for robust vehicle-mounted intelligent traffic perception.

Method

FAD is designed to improve traffic object detection in cluttered scenes by modeling frequency-domain characteristics. It integrates frequency-oriented feature extraction with adaptive cross-scale fusion, enabling the detector to capture fine object details, preserve boundary information, and enhance semantic consistency across different object scales.

Results

Representative zoom in view results of FAD on the KITTI, BDD100K and Cityscapes datasets.

Representative detection results of FAD on the Waymo 2D dataset.

Representative real-time traffic scene detection results of FAD on the RK3588 in-vehicle computing platform.

Data Availability

The KITTI data used in this study are available from the KITTI Vision Benchmark Suite. The BDD100K data used in this study are available from the BDD100K database. The Cityscapes data used in this study are available from the Cityscapes database. The Waymo Open Dataset data used in this study are available from the Waymo Open Dataset database. The numerical data generated in this study are provided in the Source Data file.

BibTeX

@article{li2026frequency,
  title={Frequency-oriented adaptive real-time object detector for cluttered traffic scenes},
  author={Li, Ziqi and Gao, Tao and Li, Shutao and Chen, Ting and An, Yisheng and Wen, Yuanbo and Lei, Tao},
  journal={Nature Communications},
  volume={17},
  number={1},
  pages={9787},
  year={2026},
  publisher={Nature Publishing Group UK London}
}