Evaluation of YOLOv8 and SSD Object Detection Models for Pothole Detection Systems

Abstract

Introduction:
Pothole detection has become an increasingly vital area of research due to its implications for road safety, vehicle maintenance, and urban infrastructure management. Poor road conditions, notably potholes, contribute significantly to road traffic accidents and increased vehicle repair costs. Traditional detection methods, such as manual inspections and sensor-based systems, are limited by labor intensity, accuracy issues, and high deployment costs across large networks.

Objectives:
This study aims to evaluate and compare the effectiveness of two deep learning-based object detection models—YOLOv8 and SSD—for real-time pothole detection. The objective is to assess which model offers superior detection precision, speed, and feasibility for deployment in varied road and environmental conditions.

Methods:
YOLOv8 and SSD models were trained using annotated datasets featuring diverse road surfaces under varying lighting and weather conditions. Image pre-processing using OpenCV was applied, and detection performance was evaluated based on mean Average Precision (mAP), Intersection over Union (IoU), and processing speed (FPS). Datasets were formatted in Pascal VOC and COCO standards, and the models were tested on mid-range GPUs to simulate real-world hardware constraints.

Results:
The YOLOv8 model achieved a higher mAP of 82%, demonstrating superior accuracy in detecting potholes, especially in complex road textures and poor lighting. However, it required more computational resources. The SSD model, while achieving a slightly lower mAP of 75%, delivered faster inference times and performed efficiently on resource-constrained devices. Both models reliably identified potholes, but YOLOv8 showed better robustness in varied conditions.

Conclusions:
YOLOv8 is more suitable for high-precision, infrastructure-grade applications due to its superior accuracy, whereas SSD is better suited for mobile or real-time deployment scenarios requiring faster processing. The study demonstrates that both models offer viable, scalable solutions for enhancing road maintenance systems and mitigating transportation-related risks.