AI Model for Matrice 4 Series - Road Defects

Product Overview

Mirror Mapper’s AI Object Detection Model is a robust, subscription-based solution for drone-based road defect detection, identifying potholes, cracks, debris, and more with high precision. Designed for simplicity and reliability, it’s retrained every 30 days to ensure optimal performance across diverse environments.

This model is designed for enterprise drone operators, municipalities, and infrastructure management teams focused on road maintenance and safety. It enables real-time or post-flight analysis of drone imagery to detect road surface issues, streamlining inspection processes and reducing manual labour. Ideal for large-scale road networks, construction sites, or disaster response, it supports proactive maintenance and regulatory compliance in urban and rural settings.

Objects Detected:

• Potholes: Surface depressions of varying sizes.

• Cracks: Longitudinal, transverse, and alligator cracks.

• Debris: Objects like branches, tires, or litter (min. 15cm length).

• Surface Anomalies: Spalling, raveling, or delamination patches.

• Road Markings (Damaged): Faded or chipped lane lines and symbols.

Model Details

• Architecture: YOLOv8-nano, optimized for lightweight, efficient processing.

• Training Environment: Google Cloud T4 NVIDIA GPU (16GB VRAM).

• Training Time: 12 hours for 100 epochs on a dataset of 12,000 labeled images.

• Dataset: 12,000+ images from global road defect datasets (RDD2022 + SVRDD), augmented with synthetic data for edge cases like low-light or wet roads.

• Accuracy Metrics:

  • Mean Average Precision (mAP@50): 92.8% across all classes.

  • mAP@50:95: 89.4% for fine-grained detection.

  • Precision: 94.5% (potholes), 92.1% (cracks), 91.0% (debris).

  • Recall: 91.2% (potholes), 90.8% (cracks), 89.5% (debris).

• Inference Speed: 22ms per frame on Google T4 NVIDIA GPU.

• Input: 640x640 RGB images from any drone camera.

• Output: Bounding boxes with defect class, confidence scores, and defect type.