The problem "Not HAL" solves

Combing through tonnes of CCTV footage is cumbersome and resource intensive and public institutions need to employ a lot of people to monitor security cameras, especially in the case of large-scale deployments such as Smart Cities, where the sheer volume of cameras prevents video feeds from being monitored continuously. This turns out to be very costly in terms of time, effort, money and human resource. Moreover, many anomalies go unnoticed and most are realised after the damage has already been done, which then leads to authorities checking the corresponding footage. Also, all existing solutions are based on user-defined rules. Even the few existing smart security devices in the market right now work on predefined irregularities, such as, intruder in house when owner is outside. An Intelligent Surveillance System, with Artificial Intelligence (AI) backed Anomaly Detection technology can thus help solve a myriad of problems by increasing the customers' ability to quickly respond to unforeseen security and safety events. Our product is a robust self-learning technology that instantly alerts users to atypical incidents by leveraging cutting-edge AI research and Deep Learning-powered algorithms. It models regular and irregular behaviour from only a small sample of anomaly data. It allows users to be proactive and control situations in almost real-time; instead of reacting to anomaly's after they have caused damage. It can also be very easily integrated with a user's existing security cameras and use the Software as a Service (SaaS) with heavy computation in the cloud, reducing hardware overhead. We are infusing security surveillance systems with data-driven intelligence. Without pre-definition by the user, "Not HAL" Anomaly Detection can quickly learn the regular movement and traffic in a scene, and alert to irregular incidents, such as a crowd forming or running, a traffic accident or interruption, and more.

Challenges we ran into

There are several datasets related to action recognition which target specific actions or environmental conditions. But to succeed, we needed to build an intelligence engine which would have the ability to differentiate between normal incidents and corresponding deviation from normal - anomalous events. And since most of the datasets available online are more aligned to western scenarios, tweaking the models to allow for customizability as well as scalability turned out to be very challenging.
Also, to achieve practical usability, we needed to get out of our academic comfort zones and work towards a robust model while also maintaining presentability of the end product. Scheduling the tasks properly and coordinating with the team in order to give proper justice to every stage of product development was not only very challenging but also an important part of why we were able to complete the project in time.