UAV Mission Planning App

Grid-based flight planner for autonomous drone survey missions.

Manual drone flight planning for large survey areas is slow, inconsistent, and produces gaps in coverage. Operators need automated grid generation that accounts for camera FOV, overlap requirements, and terrain.

A mission planning application that generates optimized flight grids from boundary inputs. The planner calculates waypoints based on camera parameters, desired overlap percentage, and altitude, then exports MAVLink-compatible mission files.

User defines a survey boundary → system calculates grid spacing from camera FOV and overlap settings → generates waypoint sequence with altitude and heading → exports as mission file for autopilot upload.

Overlap-based spacing required careful geometry to avoid coverage gaps at boundary edges. Handling non-rectangular survey areas meant clipping grid lines to arbitrary polygons. Balancing flight time against coverage density.

Generates consistent, coverage-verified grids for survey areas. Reduced mission planning time from manual waypoint entry. Next step: terrain-following altitude adjustment.

Silt Fence Detection System

Computer vision model for aerial compliance inspection of erosion control.

Construction site compliance inspections for silt fences are done manually by walking the site—slow, expensive, and inconsistent. Aerial imagery could automate this, but silt fences are thin, low-contrast, and partially occluded by vegetation.

An object detection model trained on aerial drone imagery to identify and locate silt fences in construction site orthomosaics. Includes a data pipeline from raw drone captures to annotated training sets.

Drone captures aerial imagery → images processed and tiled → model runs inference on each tile → detections aggregated and mapped to GPS coordinates → compliance report generated.

Limited training data for an uncommon object class. Silt fences vary widely in appearance—new vs weathered, buried vs exposed. High-resolution tiles needed for detection but increased compute time significantly. Balancing precision vs recall for compliance use case.

Model detects silt fences in aerial imagery with usable accuracy for flagging areas that need manual review. Significantly reduces inspection time compared to walking the full site. Improving with more diverse training data.

Drone Distance Estimation

Monocular distance estimation from drone-mounted cameras for field measurement.

Measuring distances and dimensions from drone imagery typically requires expensive LiDAR or photogrammetry post-processing. A faster, lighter approach using single-camera geometry could enable real-time field estimates.

A distance estimation system that uses known camera parameters, drone altitude (from telemetry), and detected object dimensions to calculate real-world distances from single aerial frames.

Camera intrinsics + drone altitude provide ground sample distance → object detection identifies targets in frame → pixel measurements converted to real-world dimensions using GSD → distances reported in real units.

Accuracy degrades with lens distortion at frame edges. Altitude telemetry has noise that propagates into distance calculations. Non-flat terrain breaks the planar assumption. Needed calibration workflow accessible to field operators.

Produces usable distance estimates for flat-terrain surveys within acceptable error margins. Works in real-time with live telemetry feed. Next: distortion correction and terrain compensation.

n8n AI Automation Workflows

Self-hosted automation pipelines integrating AI models into operational workflows.

Repetitive operational tasks—data processing, notifications, report generation—consume time that should go to engineering work. Off-the-shelf automation tools don't integrate well with custom AI models or self-hosted infrastructure.

A set of automation workflows on self-hosted n8n that chain AI model calls, data transformations, and external service integrations into end-to-end pipelines.

Trigger events (webhooks, schedules, file changes) → n8n orchestrates data flow through processing nodes → AI model inference where needed → results routed to outputs (notifications, databases, reports).

Running on resource-constrained hardware (Raspberry Pi) limited model size and concurrency. Debugging async multi-step workflows required careful logging. Handling API rate limits and failure modes gracefully.

Runs reliably on self-hosted infrastructure. Eliminated several hours of manual work per week. Demonstrates practical AI integration beyond model training.