Deniz Meral

Synthetic Dataset Generator for Underwater Pose Estimation

A synthetic data generation pipeline that produces photorealistic, automatically annotated training images for the team's autonomous underwater vehicle. Real underwater imagery is expensive to collect and almost impossible to label accurately for six-degree-of-freedom pose, so the scenes, lighting, turbidity and object placement are generated instead and the ground truth is taken directly from the simulator.

The generator outputs pose annotations for DOPE alongside instance and semantic segmentation masks, which lets the vehicle both recognise competition targets and estimate where they are relative to the hull — the input the guidance and control layer needs in order to approach and manipulate them.

My contribution: built the scene generation and annotation export pipeline and the Python tooling that converts simulator output into training-ready datasets.

Autonomous Underwater Vehicle — Software Sub-Team

Ongoing development of the software stack for an autonomous underwater vehicle entered in international robotics competitions. The vehicle must locate, approach and interact with submerged targets with no operator input, no reliable radio link and no satellite positioning, which places the entire burden on onboard sensing, estimation and control.

Work spans perception pipelines for underwater object detection, integration of the vision stack with the navigation nodes under ROS, and simulation-based testing of behaviours before they reach the pool. A large share of the effort is integration: agreeing interfaces with the electrical sub-team for sensor and thruster hardware and with the mechanical sub-team for frame geometry and buoyancy.

Cloud Resource Automation Tool

A monitoring and automation utility written in Python that takes an inventory of running compute instances, compares utilisation and cost against defined budget thresholds, and shuts down resources that have been idle beyond a set limit. Structurally it is a supervisory control problem: measure, compare against a setpoint, apply a corrective action, and report.

It applies a hysteresis-style rule so that instances near the idle threshold are not started and stopped repeatedly, produces a savings report for each run, and is packaged in a container so it runs identically on any host.

Automated Build & Deployment Pipeline

A continuous integration pipeline that compiles a project from source, caches intermediate build artefacts, and deploys the result automatically on every change to the main branch, with no manual intervention. The same discipline applies directly to embedded work: a firmware image that is built reproducibly by a machine, from a known commit, is far easier to trust in the field than one built on somebody's laptop.

Control Systems Coursework & Laboratory Work

Continuing university work in modelling and control: deriving plant models for electromechanical systems, designing compensators against specifications for overshoot, settling time and steady state error, and verifying them in Simulink before implementing the discrete-time equivalent on a microcontroller. Laboratory sessions cover analogue and digital circuit construction, measurement practice, and instrumentation of sensor signals.