Privacy-Preserving People Analytics

Industry

R&D / Computer Vision

Region

Poland (EU)

Neural Network Inference Anonymization by Design 3D Point Cloud Analysis Depth Imaging Machine Vision

Home

Recent projects

Privacy-Preserving People Analytics

Client Overview

In collaboration with a computer vision specialist from a Polish university, we designed a machine vision system for people counting and zone monitoring using only depth data. No RGB, no faces — privacy compliance is not a policy configuration, it is baked into the sensor layer.

Challenge

Retailers need flow analytics but RGB cameras create high GDPR exposure. Depth sensors solve privacy but standard detection models are trained on color images. Accurate tracking requires understanding 3D geometry and orientation without conventional visual cues or hardware-heavy lab setups.

Solution

We built a two-module pipeline that separates scene understanding from detection and tracking, ensuring real-time performance and privacy:

3D Scene Understanding — uses RANSAC-based plane detection to identify ground and wall geometry from raw point clouds.
Depth-Encoded Detection — neural network inference adapted for depth-encoded input, enabling person localization without any RGB data.
Multi-Object Tracking — implements a Hungarian algorithm-based tracker for stable path reconstruction and directional entry/exit counting.
ToF Sensor Simulation — a synthetic data generation framework based on Blensor that accelerates R&D without the need for physical hardware.
3D-Aware Zone Analytics — handles arbitrary sensor placement and room geometry to provide accurate orientation and occupancy data.

Key Outcomes

GDPR-Compliant by Design — No identity data is captured at the sensor level, making it safe for healthcare, retail, and public spaces.
Accurate 3D Zone Analytics — Reliable entry/exit counting and directional tracking in real-world environments with varying geometry.
Hardware-Independent R&D — Simulation framework enabled rapid experimentation and model validation without dedicated lab infrastructure.
Expert-Validated Research — Developed with a university specialist and tested against MOT benchmarks and real-world scenarios.