Application of YOLOv8 in Monocular Downward Multiple Car Target Detection

Shangke Lyu

TL;DR

This paper enhances the YOLOv8 object detection model for autonomous driving by incorporating structural reparameterization, a bidirectional pyramid structure, and a novel detection pipeline. The improved model achieves a detection accuracy of 65.2% for multi-scale, small, and distant objects, showing significant promise for real-world applications and autonomous driving competitions.

What Problem Does This Paper Address?

Autonomous driving technology promises increased safety and efficiency, but a critical component is accurate object detection. Current systems, relying on radar and cameras, face challenges like high costs, susceptibility to weather and lighting, and limited resolution, especially for small, distant, or occluded objects. This makes robust detection of road targets, such as pedestrians and other vehicles, a significant hurdle. The paper aims to overcome these limitations by improving the performance of YOLOv8 for detecting a wide range of object sizes and distances, crucial for both general autonomous driving and competitive scenarios like the Formula Student Autonomous China (FSAC).

Methods

The authors propose an enhanced YOLOv8 framework with three key improvements:

  • Structural Reparameterization: This technique optimizes the network's structure to better handle multi-scale and small targets without sacrificing accuracy. It's integrated into the C2f-DBB module to boost detection speed and precision.
  • Bidirectional Pyramid Structure Network Model: The original YOLOv8's neck structure (PAFPN) is unidirectional, potentially leading to incomplete feature extraction. This paper introduces a bidirectional pyramid structure to enhance feature integration and flexibility, allowing for more comprehensive capture of multi-scale and semantic information.
  • Novel Detection Pipeline: A new pipeline structure is introduced to optimize information flow, improving inspection efficiency and accuracy through enhanced feature fusion and information transfer.

Datasets

The paper uses two datasets for evaluation:

  • SODA-D (Small Object Detection in Aerial Images - Drone): This dataset focuses on detecting small objects in UAV aerial images, providing data relevant to challenges like occlusions and partial visibility.
  • VisDrone: A large-scale dataset for UAV video analysis, offering diverse scenarios and environmental conditions, including small-sized targets.

Both datasets are chosen for their suitability in addressing small target detection needs relevant to autonomous driving competitions.

Key Insights and Ideas

  • The Power of Bidirectionality: Moving from a unidirectional to a bidirectional pyramid structure in the neck of the network significantly improves the integration and utilization of multi-scale and semantic features. This is crucial for accurately detecting objects of varying sizes and distances.
  • Structural Reparameterization for Efficiency: Applying structural reparameterization techniques, particularly within modules like C2f-DBB, can enhance detection speed and accuracy simultaneously, a critical trade-off in real-time autonomous systems.
  • YOLOv8 as a Strong Foundation: The paper demonstrates that by strategically enhancing the existing YOLOv8 architecture, substantial gains in performance can be achieved, making it a robust base for advanced object detection tasks.

Key Results

The improved YOLOv8 model demonstrates significant performance gains:

  • Detection Accuracy: The enhanced model achieves a detection accuracy of 65.2% for both larger and smaller objects on the SODA-D dataset.
  • mAP Improvement: The overall [email protected] is approximately 7% higher than the original YOLOv8 model, with a maximum mAP value of 0.716.
  • Qualitative Improvements: Visualizations show the improved model achieving higher detection accuracy and more precise object positioning compared to the baseline method.

Limitations and Future Work

  • Limitations: The paper does not explicitly detail limitations beyond the inherent challenges of object detection in adverse conditions. The reported 65.2% accuracy, while an improvement, suggests there is still room for further enhancement.
  • Future Work: The authors highlight the model's potential for real-world applications and autonomous driving competitions like FSAC, particularly for single-target and small-object detection scenarios. Further research could explore deploying this model in more complex, dynamic environments and testing its robustness against a wider range of weather and lighting conditions.

Summary updated

Abstract

Autonomous driving technology is progressively transforming traditional car driving methods, marking a significant milestone in modern transportation. Object detection serves as a cornerstone of autonomous systems, playing a vital role in enhancing driving safety, enabling autonomous functionality, improving traffic efficiency, and facilitating effective emergency responses. However, current technologies such as radar for environmental perception, cameras for road perception, and vehicle sensor networks face notable challenges, including high costs, vulnerability to weather and lighting conditions, and limited resolution.To address these limitations, this paper presents an improved autonomous target detection network based on YOLOv8. By integrating structural reparameterization technology, a bidirectional pyramid structure network model, and a novel detection pipeline into the YOLOv8 framework, the proposed approach achieves highly efficient and precise detection of multi-scale, small, and remote objects. Experimental results demonstrate that the enhanced model can effectively detect both large and small objects with a detection accuracy of 65%, showcasing significant advancements over traditional methods.This improved model holds substantial potential for real-world applications and is well-suited for autonomous driving competitions, such as the Formula Student Autonomous China (FSAC), particularly excelling in scenarios involving single-target and small-object detection.
BibTeX
@inproceedings{lyu2025application,
  author = {Shangke Lyu},
  title = {{Application of YOLOv8 in Monocular Downward Multiple Car Target Detection}},
  booktitle = {Applied and Computational Engineering},
  year = {2025},
  doi = {10.54254/2755-2721/2025.20045},
}
Figure