Overwatch

AI research ML engineer

YOLOv8 computer vision model trained and evaluated for aerial human detection — precision/recall tradeoffs for disaster and search scenarios.

Problem

Aerial search workflows need reliable human detection across uneven terrain and canopy cover. Overwatch trains and evaluates a YOLOv8 detector on drone-style imagery so response teams can prioritize frames with likely survivors or crowd concentrations.

What it covers

YOLOv8 training pipeline with aerial human class labels
Validation harness with mAP and per-altitude slices
Export to ONNX for edge inference experiments
Python tooling for batch scoring of frame folders

Training and inference

def evaluate_altitude_bucket(model, dataset, min_m: int, max_m: int) -> Metrics:
    subset = [s for s in dataset if min_m <= s.altitude_m < max_m]
    return model.val(data=subset, plots=False)

Visualization placeholder — swap for bounding boxes on a drone frame.