IIT-Bombay develops AI platform for real-time video surveillance
A state-of-the-art video surveillance platform developed by researchers at the Indian Institute of Technology-Bombay (IIT-B) has found application in military surveillance as well as remote monitoring of social distancing norms violations amid Covid-19 pandemic.
The platform Surakshavyuh, initially designed in 2017, has now evolved into an enterprise grade video analytics solution based on machine-learning-enabled technology that can detect physical intrusion and loitering, monitor perimeters and track objects, count crowds and recognise faces, among other features. It was designed under an industry collaboration between the National Center of Excellence in Technology (NCETIS) at IIT-B and SrivisifAI Technologies Pvt Ltd, a Pune-based company working on making artificial intelligence (AI) and video analytics technology affordable for mass consumption.
While the team at IIT-B worked on the science behind the tools, SrivisifAI took the product to the market.
“For any CCTV surveillance system, the footage needs analysis – either real-time or for retrospective diagnosis. We have developed these solutions with thorough ground studies, research iterations and by developing required algorithms that can alert end-users of products to take requisite actions. This is even as we continue researching to push the frontiers in detection of human-object interaction, unusual events, etc,” said Ganesh Ramakrishnan, Institute Chair Professor, Department of Computer Science and Engineering, who led the team in the project.
Unusual human activity includes loitering, sneaking in, intrusion, someone entering and exiting the premises.
Surveillance currently primarily depends on CCTV camera footage. “CCTV footage videos are typically subjected to post-mortem analysis for events. What our tool does is give us real-time analysis,” said Ramakrishnan.
Apart from Surakshavyuh, the team also offers Jigayasa, an offline analysis solution that works as a video repository and search platform with features such as text search and face search. The two can be used together, too.
As an example use case, the models can be used to issue a trigger warning if more than five people gather at one place or if a person is not wearing a mask, features that can be used in the Covid-19 pandemic to detect adherence to rules in public spaces. A pilot project has been set up at the institute campus.
Subhasis Chaudhuri, director of IIT-B, said, “Visual surveillance has become commonplace, be it in individual houses or at public places. It is also very important to have proper surveillance at vital installations. We are very happy that IIT-Bombay has come up with a very user-friendly and robust surveillance system, thanks to the effort of our scientists. I hope that various commercial users will also benefit from this system.”
The algorithmic and data benchmarking suite called Visiocity provisions for condensing hours’ worth of video into a couple of minutes in a domain specific manner, by preserving key events and vignettes from the original video and removing repetitive visual information. Ramakrishanan said, “There are three solutions to look at, but as all are interdependent or complementing each other, the industry partner SrivisifAI Technologies can offer products and services which are combinations too — such as 3rdAI which is a combination of multiple solutions mentioned above. This is exactly the uniqueness that NCETIS has facilitated — the coming together of academia and industry as very strong partners.”
Conventionally, deep machine learning models are trained on large data sets and require lots of computer resources such as multiple graphics processing units, both of which can be expensive. The design of Surakshavyuh is based on Data Efficient Machine Learning — learning with frugal amounts of data and learning efficiently, an attempt by the researchers at IIT-B to train state of the art models in resource constrained environments while minimally affecting the accuracy.
“Humans, based on conscious thinking, can be fairer (if they decide to) and based on complex decision making, can be more precise than machines. Machines can be good at recalling, thanks to their memory and consistency in decision making. The data efficient machine learning paradigm that underlies our democratised AI effort in our products is critical to making this collaboration between humans and AI engine,” Ramakrishnan said.