Used AI Models for Drone-Based Infrastructure Inspection

FlyScope integrates cutting-edge computer vision models, optimized for aerial inspection, infrastructure diagnostics, and environmental monitoring in urban environments. Each model is chosen based on its performance, weight, and suitability for edge computing.

YOLOv8 (You Only Look Once, v8)

Use case: fast object detection on embedded hardware (Jetson, Qualcomm RB5)

• Detects poles, cameras, light fixtures, corrosion spots, and dirt patches
• Real-time performance on video streams from drones
• Lightweight enough for on-board inference during flight
• Ideal for rapid scanning and defect flagging in low-latency scenarios

SegFormer

Use case: semantic segmentation of infrastructure surfaces

• High-resolution segmentation of rust, paint damage, dirt accumulation, surface cracks
• Used for cleaning prioritization and surface maintenance reports
• Efficient transformer-based architecture suitable for edge/cloud hybrid setups

SAM (Segment Anything Model, Meta AI)

Use case: general-purpose segmentation in uncertain environments

• Zero-shot and promptable segmentation of new object classes
• Useful for post-processing analysis in complex or irregular shapes
• Currently used in offline processing and research pipelines

CLIP / OpenCLIP (for visual-language tagging)

Use case: multi-modal inspection reports and weak supervision

• Maps visual features to descriptive tags
• Helps classify scenes where labeled data is limited
• Used in conjunction with image-text retrieval for Smart City dashboards

Custom-trained classifiers

Use case: binary or multi-class classification of detected objects

• E.g., corrosion: light / moderate / critical
• Trained on internal datasets from field flights and annotation teams
• Continuously improved with human-in-the-loop labeling and feedback loops

We continue to benchmark and retrain models based on:

• Edge device constraints (latency, power, memory)
• Dataset feedback from new cities and object types
• AI explainability and regulatory compliance (e.g., infrastructure-grade inspection protocols)