The Goal

A big Bulgarian chain store needed a smart surveillance system to help spot shoplifters, pickpockets and other known troublemakers who usually operate in different locations to avoid detection. The management wanted to provide higher security environment to their customers and staff as well as to prevent inventory shrinkage and other damages. Therefore, they decided to use computer vision technology to facilitate the identification of suspicious characters.

The Challenges

There were various difficulties in achieving this goal. First of all, for such a smart surveillance system to work optimally, it was critical to find suitable locations to install the cameras that would provide the best field of view, optimal lighting conditions, the necessary size of human faces, etc.

Other important challenges were:

• face orientation – detecting faces when people were not looking straight into the camera;
• visual biasing – taking into account hairstyles, glasses, beards, etc.;
• quick service – functioning in real-time or with a minimum fixed delay;
• video clips creation – the ability to reproduce short video clips showing the actions of the person of interest (exact time of appearance, interactions with others, etc.);
• retrospective search – the ability to review the video material and locate persons unknown to the surveillance system.

The Solution: A Smart Surveillance System for Efficient Face Recognition

To account for the place-specific requirements such as camera positioning, FOV, backlight issues, minimal face size, etc., Ontotext installed cameras that provided the necessary optics and embedded functionality.

At the back-end, the smart surveillance solution consisted of several modules. The most important and computationally intensive is the stream analyzer, which performs human detection and tracking as well as face detection and characteristics calculation. The innovative models (trained with machine learning) for detection and characteristics calculations efficiently handle different face orientations. The models also have good resistance on occlusions and visual biasing such as facial hair styles, and are relatively rigid on sunglasses and haircuts.

Using a human head detector and an advanced tracking algorithm, the smart surveillance solution can determine the period of the presence of a specific person. The collected metadata is sent to server processes and is stored in a database of known offenders and other persons of interest. These processes also generate additional data and upon face recognition, notify the user.

The web UI notifications are sent in real-time, which enables users to watch specific incoming streams. There is also a video-recording capability for producing short video clips. This allows users to review the appearance and actions of a particular person upon request. They can also access a wide range of reports such as a log of when a specific person comes into sight. Last but not least, the system allows a retrospective search and, when provided with a photo, will display the locations where this person shows up and what his actions are.

Why Choose Ontotext?

The smart surveillance solution for face recognition features all functionalities of a high-end surveillance system. On top of that, it enables the chain store to:

• easily check videos of a suspicious person;
• efficiently discover shoplifters working together;
• eliminate the need for security staff to remember the faces of wanted persons;
• maintain a centralized database with known offenders and apply it in all chain store locations.

Do you think this case resembles your particular needs?