Frequency-oriented adaptive real-time object detector for cluttered traffic scenes

Abstract

Given the growing demand for traffic object detection in autonomous driving, achieving both efficiency and accuracy on in-vehicle platforms remains challenging. To address this issue, we propose a Frequency-Oriented Adaptive Detector for vehicle-mounted intelligent traffic object detection. It integrates high- and low-frequency information to enhance the texture and semantic representations of multi-scale objects in complex traffic scenarios while maintaining low computational overhead. Specifically, we introduce Frequency Dynamic Convolution to construct a lightweight backbone with frequency-domain adaptive dilated receptive fields and balanced effective bandwidth. The adaptive kernel decomposes convolutional weights into high- and low-frequency components, which are selectively recalibrated to balance frequency responses in feature maps. Moreover, we propose an Adaptive Frequency-Oriented Fusion framework to reorganize high- and low-frequency features across scales. The framework balances object details, fine boundaries, and deep semantic features, thereby reducing feature inconsistencies during multi-scale fusion. Extensive experiments demonstrate that our Frequency-Oriented Adaptive Detector outperforms state-of-the-art detectors. With only 8.77 million parameters, it achieves mAP values of 90.9%, 53.7%, 50.2%, and 53.1% on KITTI, BDD100K, Cityscapes, and Waymo datasets respectively. This study presents an efficient AI detector for intelligent vehicles called Frequency-oriented Adaptive Detector. By adaptively using high- and low-frequency visual information, it improves object detection in complex traffic scenes while keeping computation low.