With terrorism and COVID-19 growing threat, governments and law enforcement bodies worldwide are looking for next generation face recognition (FR) systems able to address increasing security concerns in crowded places. There is a need to move FR systems to large-scale and more unconstrained environments (e.g. large indoor and outdoor places with changing lighting conditions, views, occlusions…). It is also of the utmost importance to introduce crowd behavior analytics capacities and detect abnormal events, such as mass panic, stampedes or bottlenecks. Particularly, the automatic monitoring of crowd density, and thus of the distance between individuals, becomes vitally important to control the spread of COVID-19.
Current solutions are no longer sufficient to cover these needs; important customer pain points arise:
1. Computational requirements exponentially increase because of the vast amount of video streams and faces to concurrently analyze in real-time.
2. Lack of scalability. When it comes to covering geographically extended scenarios, current FR systems have to be replicated (including databases of enrolled persons). The installation process becomes much harder, prone to errors, vulnerable and time- and cost-consuming.
3. Current FR systems do not provide crowd behavior analytics.
4. There are increasing concerns about privacy, which demand a trade-off between security and privacy, and ensuring compliance with the EU General Data Protection Regulation (GDPR) and other local regulations.
The AWARE project aims at bringing to the market a ground-breaking product that will change the way Face Recognition (FR) systems are deployed. The proposed FR system will be able to cover geographically extended, crowded and unconstrained scenarios, and to massively process faces with high performance; altogether with the detection abnormal crowd behaviour such as mass panic, stampedes or bottlenecks.
HERTA will solve state of the art limitations in the deployment of FR and crowd analysis algorithms with a hybrid edge and distributed computing pipeline, keeping secured databases in the back end. Part of the algorithmic pipeline is executed on-premise, in the front end, on low-power embedded Deep Neural Network (DNN) accelerators. This allows saving costs and increasing the system scalability, efficiency, ease-of-installation and performance. While face detection and crowd analysis are carried out on the edge, identity matching takes place in the back end, so that the database of subjects and their personal data are stored in a secure location and remain protected according to the GDPR.
