An annotated water-filled, and dry potholes dataset for deep learning applications

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Potholes pose significant risks to autonomous systems, particularly Electric-Powered Wheelchairs (EPWs), due to their varied shapes and water-filled surfaces.
• Existing deep learning methods show promise for pothole detection, but current datasets lack comprehensive data on challenging pothole conditions.
• The safety and reliability of autonomous assistive technologies are hindered by the inability to accurately detect diverse pothole types.

Purpose of the Study:

• To address the limitations of existing datasets by creating a novel dataset for pothole detection.
• To provide a comprehensive collection of pothole images encompassing various shapes, locations, colors, and conditions, including water-filled and rubble-filled potholes.
• To enhance the performance and robustness of deep learning models for autonomous navigation systems.

Main Methods:

• Manual collection of 713 high-quality photographs of potholes using mobile phones across different UK locations.
• Manual annotation of 1152 potholes within the collected images, detailing their characteristics.
• Inclusion of two supplementary benchmarking videos captured via dashcam to represent real-world driving scenarios.

Main Results:

• The developed dataset features a wide array of pothole conditions, including water-filled, rubble-filled, and irregularly colored potholes.
• The dataset provides detailed manual annotations, crucial for training and validating deep learning models.
• The inclusion of diverse visual data enhances the potential for improved pothole detection accuracy.

Conclusions:

• The new dataset significantly expands the scope of pothole variations available for AI training, particularly for autonomous systems.
• This resource is expected to advance the development of safer and more reliable Electric-Powered Wheelchairs and other autonomous assistive technologies.
• Improved pothole detection capabilities will contribute to enhanced user safety and reduced risk of accidents in autonomous navigation.