View Proposal

---

Proposer

Zi Hau Chin

Title

Privacy-Preserving Classroom Occupancy Estimation Without Facial Recognition Project Purpose

Goal

Estimate classroom occupancy without identifying or tracking individuals.

Description

This project involves developing a privacy-preserving system that estimates the number of people in a classroom without recognising, identifying, or tracking individuals. The student will investigate suitable people-counting approaches, such as object detection, crowd counting, background subtraction, low-resolution vision, silhouettes, depth images, thermal sensing, or other anonymous sensing methods. The minimum prototype may use a camera, but it should process data locally and avoid permanently storing identifiable images wherever possible. A key part of the project is investigating the trade-off between occupancy accuracy and privacy. For example, reducing image resolution, applying irreversible visual transformations, using depth information, or immediately deleting frames may provide stronger privacy but reduce counting accuracy. The student should implement and compare selected approaches under different classroom conditions, including low and high occupancy, different lighting levels, partial occlusion, seated and moving participants, and people entering or leaving the room. The completed system should display the estimated occupancy, confidence level, processing status, and historical aggregate counts through a dashboard. It may also provide configurable room-capacity alerts, but it must not record attendance, identify individuals, or perform facial recognition. Evaluation should consider counting error, processing latency, frame rate, memory and CPU usage, robustness under different conditions, and the amount of potentially identifiable information exposed or retained. The final system should demonstrate whether useful classroom occupancy information can be obtained while minimising the collection and storage of personal data. -What accuracy is lost when privacy-preserving transformations are applied? -Can edge processing achieve acceptable counting accuracy and latency? -How do RGB, low-resolution, depth, silhouette, or Wi-Fi approaches compare? -Does the system remain reliable under different lighting, movement, and occlusion conditions?

Resources

ESP32? Raspberry Pi 4b? People counting dataset, CounfFi, WiVi32 People COunting toolkit, OpenVINO person counting notebook...

Background

Python, computer vision and image processig, object detection, people-counting, machine learning

Url

Difficulty Level

Challenging

Ethical Approval

Full

Number Of Students

1

Supervisor

Zi Hau Chin

Keywords

privacy-preserving computer vision, people counting, occupancy estimation, edge ai, anonymous sensing, wi-fi csi, object detection

Degrees

Bachelor of Science in Computing Science