Privacy & Data Protection
Objective
This document explains how KSI VISION's video analytics is designed to operate without processing personal data, the technical guarantees behind that design, and exactly what data the system generates. It is intended for legal teams, data protection officers, and compliance auditors evaluating the deployment of KSI VISION.
Background
Video analytics is a common, globally established practice that emerged with the mass adoption of security camera systems. Today virtually every commercial camera ships with embedded analytics (motion detection, people detection, counting, heatmaps). Since 2012, AI-based models have made it possible to extract far more operational value from video streams, and since 2018 GPUs have made this viable at scale.
KSI VISION has operated since December 2019, applying this technology across Europe, Latin America, and North America — passing numerous commercial and legal validations by the companies where it has been deployed.
Legal Framework
The relevant legal analysis concerns personal data protection laws. The global reference is the GDPR, under which personal data is any information relating to an identified or identifiable living person — including combinations of data that together could identify someone, and de-identified or pseudonymized data that could be re-identified.
KSI is designed not to process personal data at all, which is the foundation of its compliance approach across jurisdictions.
Sector-specific regulations on security camera systems (for example, Spain's Law 5/2014, which regulates who may view security recordings and when they may be shared with authorities) do not affect KSI either: no external agent ever views the video, and no recording of the analyzed spaces is ever generated.
How KSI Avoids Generating Personal Data
KSI avoids personal data through four design criteria:
- No biometric or facial recognition — ever. Nothing is computed that could directly or indirectly identify a person.
- Only aggregated and anonymized data, generated in a non-reversible way, keeping solely the information relevant for commercial use.
- No storage of any recording or image. The analytics runs in real time directly on the encoded video streams (H.264/H.265); every analyzed frame is discarded in less than 1 millisecond, making it impossible for any external agent to access images or video.
- Local, parallel processing. All image processing happens on a server installed inside the client's premises, in parallel with and independent from the CCTV security system — fully under the client's control, similar to a DVR with analytics. Raw images are never transmitted to external servers.
Technical Anonymization Flow (Inside the GPU)
The entire analytics pipeline runs inside the GPU, with anonymization guaranteed from the point of entry:
- Decoding wired to the anonymization filter. The compressed stream (H.264/H.265) is decoded by the GPU's integrated decoder chip, which sends the decoded data straight to the anonymization filter — the frames are never exposed to any external component.
- Immediate anonymization. The first layer of the convolutional neural network (CNN) irreversibly transforms the visual information, discarding any biometric detail. Only abstract features needed for the analysis are retained: contours, trajectories, patterns.
- Irreversible transformation. Decoded images can never be reconstructed or reversed at any later stage; originals are discarded in under 1 millisecond.
- Metadata processing. Subsequent network layers turn the abstract features into metadata — counts, occupancy densities, movement vectors — containing no visual or biometric information.
- Real-time operation. The full flow, from decoding to metadata, processes each frame in under 1 millisecond, implemented in optimized GPU routines (CUDA).
- Compliance by design. There is no storage, transfer, or access to original data at any point.
Data Generated by the Analytics
The system produces only anonymous, aggregated, non-reversible structured metadata — never images, video, or biometric information:
| Data type | What it measures | Example (format) |
|---|---|---|
| People counting | People crossing virtual lines or entering/exiting defined zones | {"timestamp": "2025-04-02T10:15:00Z", "zone": "Entrance A", "count_in": 12, "count_out": 8} |
| Occupancy & density | How many people are present in an area and how they are distributed | {"timestamp": "...", "zone": "Food Court", "occupancy": 47, "density": "medium"} |
| Heatmaps | Aggregated movement patterns and concentration over time | Grid matrix of presence intensity per interval |
| Dwell time | How long people stay in a zone | {"timestamp": "...", "zone": "Display B", "avg_dwell_time": 78} |
| Trajectories (identity-free) | Anonymous paths through a space, as non-identifiable time vectors | {"path_id": "xyz123", "path": [{"x": 10, "y": 14, "t": 0}, ...]} |
| Group detection | Whether people walk alone or in groups (spatial/temporal proximity) | {"timestamp": "...", "zone": "Main Hall", "groups": 3, "individuals": 5} |
Key characteristics of all generated data:
- Non-personal — no names, IDs, faces, or any identifiers
- Anonymous — no link to persons or devices
- Aggregated — grouped by time interval and zone
- Ephemeral — temporary identifiers (such as trajectories) are deleted immediately after use
- Real-time — generated and transmitted with sub-second latency
Conclusion
KSI VISION's architecture — processing entirely inside the GPU in real time, discarding frames immediately after anonymization, and never transmitting raw images — provides strong technical guarantees against misuse or unauthorized access. By avoiding the capture, storage, or processing of any personally identifiable information, and generating only anonymized, aggregated metadata, KSI ensures compliance with data protection laws such as the GDPR and its regional equivalents, and simplifies legal assessments for clients across jurisdictions.
Support
For the full legal analysis document, data processing questionnaires, or any compliance questions: