Autonomous Fruit Counting Drone

An autonomous quadcopter built on the PX4 flight stack with a Raspberry Pi companion computer running a YOLO-based fruit detector. The drone executes a pre-planned survey grid over an orchard, tracks fruit across frames to avoid double-counting, and aggregates counts into a geo-referenced yield map. ROS 2 bridges the flight controller, camera pipeline, and mapping layer.

Manual fruit counting for yield estimation is slow, labour-intensive, and statistically noisy. Growers need accurate pre-harvest yield forecasts to plan labour, storage, and sales, but ground surveys only sample a fraction of the orchard.

A PX4-controlled drone flies an autonomous survey grid while a Raspberry Pi runs a YOLO fruit detector with multi-frame tracking to deduplicate counts. Detections are geo-tagged and aggregated into a per-tree yield map. ROS 2 coordinates the flight controller, perception pipeline, and mapping nodes.

Planned — to be evaluated against manual ground-truth counts: per-tree counting accuracy, area coverage per flight, and yield-estimate error versus actual harvest.

1. Phase 1

   Airframe & autopilot

   Assemble the quadcopter and configure PX4 with a Raspberry Pi companion over MAVLink.

2. Phase 2

   Fruit detector

   Train a YOLO model on orchard imagery and add multi-frame tracking to prevent double-counting.

3. Phase 3

   Survey & mapping

   Autonomous grid survey with geo-referenced count aggregation into a yield map.

4. Phase 4

   Validation

   Compare drone yield estimates against manual counts and actual harvest data.