Sithija Ranaraja

All projectsPerception, AI

LiDAR perception

Detecting and tracking people directly from laser scans

Predicting people and obstacles the LiDAR cannot currently see, in a crowded simulated space.Simulation

Cameras have trouble with lighting and raise privacy questions, so this perception layer works on the laser scan alone. I did the data collection and labelling, the model training, the multi-frame tracker and the deployment, and tuned it on site until it ran reliably.

Type
Hardware
Period
2026
Role
End-to-end AI system development
Stack
Python, PyTorch, U-Net segmentation, 2D LiDAR, ROS 2, multi-object tracking
Links
Paper: occlusion-aware obstacle prediction (arXiv)

Challenge

A 2D scan is sparse. In a single frame a person, a chair leg and a wall corner can look the same.

Approach

A U-Net segments the scan into classes. A zero-shot path lets it take on new object types without retraining from scratch. A tracker follows people and moving objects across frames and keeps their identities. I validated the tracker in simulation against moving and static targets before the hardware.

Result

Deployed and running on the target hardware.

What I built

  • Data collection, labelling, training, tracking and deployment
  • U-Net semantic segmentation on LiDAR data
  • Zero-shot recognition of new objects
  • Real-time human and dynamic-object tracking