yolo-x: Drone-Based High-Speed Road Congestion Detection Algorithm

Chuanhui Cheng, Guanjun Wang, Xu Huang, Jie Zhang2025 · DOI

Abstract

Addressing the challenges of traditional traffic congestion detection methods in deployment, cost, and road surface integrity, this study proposes yolo-x an innovative rotating object detection solution based on unmanned aerial vehicle (UAV) platforms. This framework achieves three core technological breakthroughs: The Coordinate Attention Augmentation (CAA) module significantly enhances small-scale vehicle detection performance by establishing an explicit mapping between position coordinates and channel attention, improving the AP_S metric by 9.2 percentage points. The Dynamic Upsampler (DySample) adapts to learn optimal upsampling kernels based on local similarity principles, effectively resolving feature blurring caused by traditional interpolation methods; The improved Exponential Moving Average Attention (iEMA) mechanism integrates dilated convolutions with residual connections, enabling precise extraction of multi-scale contextual features while preserving high-resolution feature maps. Building upon this detection framework, the study innovatively proposes a rotation-based traffic congestion quantification algorithm. By calculating the proportion of vehicle rotation box area (TOR) within lane regions, it establishes a four-tier congestion classification system (free-flow, light congestion, moderate congestion, severe congestion). Experimental results demonstrate that compared to traditional axial detection methods, this approach improves congestion judgment accuracy by 10.2 percentage points and reduces false alarm rates by 3.8 percentage points, providing effective technical support for intelligent traffic management.
BibTeX
@misc{cheng2025yolo,
  author = {Chuanhui Cheng and Guanjun Wang and Xu Huang and Jie Zhang},
  title = {{yolo-x: Drone-Based High-Speed Road Congestion Detection Algorithm}},
  year = {2025},
  doi = {10.1109/aiahpc66801.2025.11290623},
}

Loading related content...

Figure