Privacy-Preserving Localization: On-Sensor AI Filtering for Ethically and Legally Safe Camera-Based Mapping
Introduction
Cameras are among the richest and most cost-effective sensors for robot localization and mapping, but their continuous capture of the surrounding environment raises serious ethical, legal, and cybersecurity concerns, since the imagery can contain identifiable faces, license plates, screens, and other sensitive information that must be protected under data-protection regulations. The risk is amplified when raw images leave the sensor and travel across a robot's network or storage, creating an attack surface for unauthorized access to private data. This project proposes shifting privacy protection to the very edge of the pipeline by exploiting smart cameras with onboard AI processing—such as recent OAK / DepthAI devices—so that AI filtering and redaction models run directly on the sensor and only already-sanitized images are ever transmitted. The central challenge is to remove or obscure sensitive content while preserving the visual features (edges, corners, textures, structures) that localization and mapping algorithms depend on, all under tight real-time and accuracy constraints, with the precise list of "must-filter" categories to be defined as part of the work.

Project Activities
• Literature Review: Survey privacy-preserving visual localization/SLAM, on-sensor (edge) AI processing, image redaction/inpainting techniques, and the relevant legal and cybersecurity frameworks for camera data.
• Defining the Filtering Requirements: Work with ethical, legal, and technical stakeholders to establish the concrete list of sensitive categories to be filtered (e.g., faces, license plates, screens, documents) and the acceptable redaction method for each.
• On-Sensor AI Filtering Deployment: Select and deploy optimized detection-and-redaction models onto the smart camera's onboard processor (e.g., OAK / DepthAI) so that only filtered frames leave the device.
• Localization and Mapping Integrity Evaluation: Quantify how the filtering affects downstream visual localization/SLAM accuracy to ensure that features essential for mapping are preserved.
• Real-Time Performance Optimization: Profile and optimize the on-sensor pipeline (model quantization, resolution, frame rate) to meet the latency and throughput required for live robot operation.
• Cybersecurity and Privacy-by-Design Validation: Verify that no raw or unfiltered sensitive data is ever exposed beyond the sensor, validating the end-to-end pipeline against a defined threat model.
Qualifications / Prerequisites
• Programming: Proficiency in Python and/or C++, with comfort working on embedded/edge platforms.
• Computer Vision & Deep Learning: Experience with object detection, segmentation, and image redaction/inpainting models.
• Localization Knowledge: Understanding of visual SLAM / feature-based localization and the image features they rely on.
• Tools & Environment: Linux, Git.
• Desirable: Familiarity with deploying and optimizing neural networks on edge devices (e.g., DepthAI/OAK, OpenVINO, quantization and pruning). Familiarity with ROS for on-robot deployment.
Contact
Alessandro Chiumento – Pervasive Systems, EEMCS, University of Twente (a.chiumento@utwente.nl)
Hojat Mirtajadini Goki – SMART Mechatronics and Robotics Research Group, Saxion University of Applied Sciences (s.h.mirtajadinigoki@saxion.nl)