Comparative Study of Lightweight Target Detection Methods for Unmanned Aerial Vehicle-Based Road Distress Survey.

Area of Science:

• Computer Science
• Civil Engineering
• Remote Sensing

Background:

• Unmanned aerial vehicles (UAVs) offer potential for road anomaly detection but face challenges with discrete spatial coverage and computational demands.
• Existing road anomaly detection algorithms require improvement for robustness and efficiency in UAV-based applications.

Purpose of the Study:

• To evaluate and compare the performance of various deep learning models for road distress detection using UAV imagery.
• To identify the most effective model for accurate and computationally efficient identification of road anomalies.

Main Methods:

• K-means clustering was employed to determine optimal prior anchor boxes.
• Several deep learning models, including Faster R-CNN, YOLOX-s, YOLOv5-s, YOLOv7-tiny, YOLO-MobileNet, and YOLO-RDD, were developed and trained on UAV-collected road image data.
• Performance was evaluated using metrics such as mean average precision (mAP) and average precision (AP) at an Intersection over Union (IoU) threshold of 0.5.

Main Results:

• The YOLO-RDD model demonstrated the highest performance, achieving an mAP of 0.701 and excelling in detecting all four types of road distress.
• YOLO-RDD showed particular success in pothole detection (AP of 0.790), with better identification of significant distresses compared to minor cracks.
• YOLO-RDD achieved an 85% computational reduction compared to YOLOv7-tiny while maintaining high accuracy.

Conclusions:

• The YOLO-RDD model is a highly effective and computationally efficient solution for UAV-based road distress detection.
• While YOLO-RDD shows promise, further refinement may be needed for optimal detection of minor road defects like small cracks.