System for automatic blurring of video streams from video surveillance cameras

Context

• RATP operates  50,000 video surveillance cameras throughout the
   Île de France  network.
• In this context, new technologies such as Edge Computing and Deep Learning represent new tools to exploit this capital and extract information from video streams.
• Video streams include personal biometric data and are subject to regulatory and legal constraints, including the General Data Protection Regulation (GDPR).

Challenge

• To  create a dataset to train AI models for smart video analysis and ensure their effectiveness in dense and complex environments such as public transport environments.
• To comply with the regulatory requirements of the GDPR on the use of videos containing personal data. Indeed, the CNIL prohibits the storage of videos containing biometric data for more than 72 hours, which makes the data impossible to use in the context of experimentation.

Solution

A real-time and irreversible anonymization tool

Deepomatic has developed with RATP a real-time anonymization tool. In order to remove all traces of personal data, video streams from video surveillance cameras are processed by detection algorithms that locate biometric data subject to GDPR, such as head, hair, and face. The identified elements are then anonymized by an irreversible blurring tool.

Results

• Acquisition of a video database that can be used to train artificial intelligence algorithms.
• Reduction of the workload during the CNIL’s Privacy Impact Assessment (PIA), and guarantee provisions for the protection of personal data.

What are the next steps?

1. Evaluate the Deep Learning algorithms on safety and operational use cases.
2. Evaluate the feasibility of embedding video recognition algorithms at the sensor level for localized computing (Edge Computing)