Computer Vision Researcher

Highlights in full

• Engineered the perception stack for autonomous driving end to end: radar and lidar capture, timestamp alignment, point-cloud processing, and what survives into a guidance decision.
• Built a replay harness over recorded sensor logs so a perception change could be evaluated against the same drive twice, which is the only way a change is attributable rather than anecdotal.
• Studied HD-map localization: matching observed lane geometry against a prior map to correct GPS drift in urban canyons where satellite fixes degrade exactly when precision matters most.

Sensing

• Solved the extrinsic calibration that puts radar and lidar in one frame of reference, since two sensors disagreeing by a few centimetres produce a phantom object rather than a better one.
• Compared early fusion against late fusion on identical sequences: raw point-level merging preserves detail and inherits both sensors' noise, object-level merging is robust and loses the evidence that would have resolved a disagreement.

Localization

• Examined vehicle-to-vehicle communication as a trust problem rather than a bandwidth one: what a car can publish about its own state, and how a receiver decides whether to act on a claim it cannot verify.
• Traced the map-matching failure that matters: a prior map is only as fresh as its last survey, so a repainted lane makes the localizer confidently wrong rather than uncertain.

What breaks

• Documented the sensor failure modes a fusion layer has to arbitrate: rain attenuating lidar returns, radar multipath off guard rails reading as a stationary obstacle, and low sun blinding the camera lane that both others depend on.