In this case study, we will look at a computer vision edge pipeline built to improve security audits and produce real-time alerts on suspicious activity. The client had a very straightforward problem – minor thefts and violation of factory rules required thousands of hours of CCTV video to be monitored by their security workforce. This process was tiresome, expensive, and more importantly prone to human error. They wanted a system that could automate the detection of common security-related scenarios and help enforce factory rules while they’re at it. And so we helped build a computer vision based security automation tool that ran simultaneously across dozens of security cameras at each zone and in real-time.
This internal product was for one of India’s largest factory groups. This meant we had to deal with a lot of data and across multiple different unstructured zones. A portion of the application was built to produce alerts in real-time whereas custom scenarios were also developed to be reported on demand. The security protocols of the factory demanded an on-premises deployment ruling out any cloud-based optimization. This was a challenging problem both in terms of deployment as well as testing – as any testing of the AI system had to happen on an on-premises CCTV system.
The business value was centered around a robust system capable of detecting suspicious activity through CCTV footage and help search for specific people/actions across all feeds on the enormous factory floor. So we started our investigation by searching for the right data, technology, and deployment method. Multiple public datasets containing a total of two million+ samples were identified to be used as the base computer vision dataset for people and action detection. A specific subset of the CCTV footage from the factory was chosen to optimize, finetune & validate. The data vetting & gathering process was carried in its entirety in under 2 months. Once the business requirements around alerts & suspicious activity reporting were fully fleshed out, a computer vision and web application-based architecture was created to produce repeatable & scalable outputs that meet the business needs. This CCTV AI architecture was approved and subsequently converted into a five-month development plan.
Development was put on the fast track owing to constraints in the client's deployment timelines. The development happened across eleven two-week sprints that iterated over the AI model's features and accuracy. A frontend application was developed simultaneously about two months into the project, as part of an extended scope to further ease the user experience aka UX. Development was completed seamlessly and on time, including additional scope that was added post-investigation as they were deemed necessary for smooth functioning of the application. The entire project was deployed by the third month with functional AI components that ran on the factory floor's CCTV network and produced real-time alerts. The application also produced stats on factory rules and monitored them for regular audit reports. The entire system was tested and wrapped up by the end of the fifth month.
The models were built into scalable APIs that could be deployed on the organization's private cloud (on premises) and integrated into their existing security dashboard through the inclusion of additional UI screens. The deployment was done using a virtual containerization service that was set up to be monitored remotely. Continuous operational & server health monitoring was provided to the client as well to ensure that our product worked out on the field as well as it did in our lab.