ROV Detection and Pose Estimation (MIT)
Perception · Machine learning
My rolePerception model training & pose estimation
A YOLOv8 detector and geometric pose estimator ran in real time on a Raspberry Pi with 95% accuracy.
Problem
The final challenge was autonomous light tag between underwater robots, which needs reliable detection and relative pose in changing lighting, turbidity, and occlusion.
My contribution
Trained a YOLOv8 detector on 3,000+ labeled frames from the ROV camera. Developed a geometry-based estimator that computes depth, distance, and bearing from the bounding box and known camera parameters, and integrated both into the ROS control system.
Result
95% classification accuracy with real-time inference on a Raspberry Pi, and real-time relative pose during pool tests. Selected for the program at under 9% acceptance.
Details
- Training images were pulled frame by frame from ROV camera video in varied lighting, turbidity, and occlusion.
- Histogram equalization and Gaussian noise augmentation improved robustness.
- Pose is derived with trigonometry from pixel displacement and known camera parameters.
- Designed additional ROV components in Onshape and presented findings to 500+ faculty and peers.