Privacy-Preserving Group Coordination Sensing

This capstone project with The Ohio State University ECE department explored how to measure and interpret group coordination, neutral activity, and disruption in shared spaces without relying on cameras. The goal was a privacy-preserving approach that could still capture meaningful motion patterns and support local processing.

The team began with an mmWave radar concept because radar produces anonymous motion signals instead of identifiable imagery. During spring testing, the team found that the radar data were not dense enough for reliable multi-person detection or downstream classification. As a result, the project pivoted to the Unitree 4D L2 LiDAR, which provided richer spatial data for the full system.

The final prototype used LiDAR-based sensing, a session capture and conversion pipeline, encrypted local storage, two machine learning layers, and a post-session interface. Raw LiDAR sessions were converted into structured PKL files for model processing. PointNet and LSTM were used to classify individual actions such as standing, sitting, walking, and waving, while a Temporal Graph Transformer classified group states such as coordination, neutral behavior, and disruption.

By the end of the project, the team had built an end-to-end prototype that could capture spatial motion data, convert sessions into usable files, run both classification layers, store data securely, and present results through an interface. The biggest remaining limitation was neutral behavior classification, which was less distinct than coordination or disruption and was sometimes confused with coordination. Future work would focus on expanding the dataset, improving group-level features, reducing occlusion with broader LiDAR coverage, and moving closer to real-time processing.

‍