Privacy-preserving AI video surveillance with edge intelligence

With the increasing deployment of AI-powered video surveillance systems, questions of privacy and regulatory compliance, especially under the General Data Protection Regulation (GDPR), have come to the forefront. Public spaces, transportation hubs, and private facilities rely on smart cameras to identify threats, analyze behavior, and improve safety. But this creates a significant technical and ethical challenge: how can these systems operate effectively in real-time while complying with strict data protection laws?

The answer lies in combining real-time AI inference at the edge with an architecture designed for privacy-by-design, minimizing data retention and avoiding cloud dependency altogether.

GDPR Meets Real-Time AI

Under GDPR, any system processing personally identifiable information (PII)—such as facial features, gait, or body patterns—must ensure lawful processing, data minimization, and transparency. However, AI systems, especially those running in the cloud, often buffer or transmit sensitive data for analysis. This creates privacy risks and network latency, which are unacceptable for real-time applications like crowd monitoring or automated incident response.

To meet GDPR standards while maintaining AI performance, the computation must shift to the edge—processing video frames locally, within milliseconds, and without persistent data storage. This is easier said than done. AI models—especially deep neural networks—are computationally intensive, making real-time inference on low-power edge devices a serious engineering challenge.

Fluendo’s Edge AI Stack: Built for Compliance and Speed

That’s where Fluendo AI Plugins come in. Our plugins are purpose-built for low-latency, high-accuracy inference within GStreamer multimedia pipelines. Designed to run directly on edge devices, they eliminate the need for cloud processing—enabling fast, private, and customizable AI workflows.

Thanks to this approach, our Fluendo AI Plugins, such as our FluAnonymizer enable GDPR-compliant use cases such as:

• Real-time face redaction in public CCTV streams
• Blur or mask application for non-consenting individuals
• Object detection without data retention, processed entirely at the edge
• People counting and motion analytics with anonymized metadata output

This ensures compliance through design, by eliminating the exposure of raw video to external servers or prolonged storage.

Edge Performance Without Compromise

The key component behind our Fluendo AI Plugins is that they have been, built on top of our in-premise accelerated AI engine Raven Engine which has been designed for multiplatform deployment, zero-copy processing, and hardware-accelerated inference. This foundation ensures maximum performance while maintaining full control over the execution environment.

Raven Engine ensures high accuracy real-time performance thanks to:

• Zero-copy memory policy (100% GPU-driven pipeline)
• Optimized rendering capabilities for both pre and post-processing
• Multi-providers AI execution: NVIDIA, AMD, Intel, etc.
• Asynchronous parallel processing pipeline
• Optimized AI models integration

This design ensures sub-30ms latency with mid-tier commercial GPUs, even when performing tasks like face detection, tracking, or background anonymization.

Performance Comparison: Real-Time Anonymization at Full HD

To showcase the efficiency of our FluAnonymizer GStreamer plugin, we compared its performance versus a standard anonymization Python-based application in a NVIDIA GeForce RTX 4060 Laptop GPU using an inputFull HD video feed from a webcam at 60 FPS.

This test evaluates and compares the processed output framerate, the number of frame drops, and the visual quality of the anonymization. This allows us to measure not just raw performance but also the consistency and reliability of the output under real-time conditions.

It is important to highlight that for this study, no pre-processing or input resizing was applied, meaning the AI engine received a full-resolution float32 tensor of shape 1×3×1920×1080 for inference. While a common strategy for handling heavy AI models is to downscale the input tensor to improve speed, this often comes at the cost of significantly reduced output accuracy. Our added value lies in the ability to run models at high resolutions, maintaining maximum accuracy while still achieving real-time performance.

Thanks to their low-level optimizations and native GStreamer integration, four plugins process Full HD streams at real-time speed (60 FPS)—ideal for live anonymization tasks under GDPR. A conventional Python + ONNX Runtime approach achieves just 8 FPS under the same conditions, despite using the same model architecture.

This performance gap highlights the critical value of highly optimized inference runtimes and media-aware AI integration when targeting real-time applications at the edge.

• Standard Python + ONNX Runtime

• Fluendo AI Plugin with Raven Engine

Conclusion: Privacy and Performance Can Coexist

Fluendo’s AI Plugins prove that GDPR compliance does not mean sacrificing real-time performance or system accuracy. By designing for the edge from the ground up, Fluendo enables secure, fast, and privacy-preserving AI deployments in surveillance, smart cities, and beyond.

If you’re looking to deploy intelligent video analytics without legal headaches, our team can help you design a tailored, compliant AI solution. Whether your goal is real-time privacy masking, analytics, or automated alerts, Fluendo’s edge-first architecture is the key.

Contact us to learn how to bring intelligent, compliant video AI into your business operations.