Real-time object detection for aquatic environments
An edge vision system that detects and tracks moving objects on and below the water line, running entirely on-device for a marine sensing product.
- Runs
- Fully on-device
- Latency
- Real-time
- Challenge
- Detect and follow objects in water in real time, on hardware mounted in the field, with no reliable network to lean on.
- Approach
- Computer vision and deep-learning models optimised for an embedded accelerator, with a GStreamer capture-and-inference pipeline tuned for low latency and glare.
- Outcome
- A tracking system that holds targets across changing light and surface conditions, deployed on the product without a cloud dependency.
Stack
- Edge AI
- Deep learning
- GStreamer
- Embedded accelerators
Background
A marine technology team needed to add perception to a field-mounted product. Objects on and just under the water line had to be detected and tracked continuously, and the deployment sites had no dependable connectivity — everything had to run on the device.
What we did
We built the perception stack around an embedded accelerator: detection and tracking models trained for the target objects, then quantised and profiled until they held real-time frame rates within the power budget. A GStreamer pipeline handled capture, pre-processing, inference, and overlay, with specific handling for glare, reflection, and wake turbulence.
Result
The system keeps a lock on targets through the light and surface changes that break naive detectors, and it does so without a network connection. It shipped as part of the product.
Engagement details are adapted and figures are confirmed with the client before publication.