A fully autonomous drone simulation (Unity) that searches a city for a very small flag, filters noisy detections, and autonomously drops a payload.
A Unity drone streams images to a server running a YOLO model (OpenVINO optimized). False positives are reduced with Evidence Accumulation and positional noise is smoothed with an Exponential Moving Average (EMA). A small Finite State Machine (FSM) controls safe decision-making and the drop behavior.
- Real-time small-object detection (YOLO + OpenVINO)
- Evidence accumulation to reduce false positives
- EMA for stable position estimates from noisy detections
- FSM-based decision logic for robust autonomous actions
- Unity client + server inference pipeline (easy to extend)
- Unity drone streams camera frames to a server.
- Server runs a YOLO model (OpenVINO-optimized) and returns detections.
- Evidence accumulation integrates multiple detection frames to confirm targets.
- EMA smooths the detected position for safe approach.
- FSM decides behaviors: search → approach → drop → return.