‘Camera wali gaadi lagti hai, crime down ho jaata hai’: How Delhi Police is using face-tech to catch criminals

While unique, facial recognition technology isn’t a new phenomenon. The Delhi Police has been using the technology, an Israeli software, since 2018.

A Facial Recognition System van parked in Northwest district. (Express photo)

Since then, it has been acquired by two police districts — North and Northwest — for more localised scanning. Currently, FRS vans, armed with cameras and computers, are stationed in different parts of the two districts every day, scanning faces and alerting them of potential hits.

In the Northwest district alone, the FRS has enabled the police to nab at least 200 criminals since April 2024. This February, police in the district solved two burglary cases using the software.

How it works

On a warm spring afternoon, the Northwest district’s FRS van, with the Delhi Police logo emblazoned on the sides, stands conspicuously in front of Deep Market in Ashok Vihar.

“Jis din camera wali gaadi lagti hai uss din area mein crime down ho jaata hai (On days the van is stationed somewhere, the crime in the area goes down),” says Constable Mukesh Kaushik, one of the two policemen manning the van that day.

The van has six CCTV cameras — four outside and two inside — along with two Automatic Number Plate Reader (ANPR) cameras which have been configured to scan faces instead of licence plates.

Inside the van, the set-up is tight: A desk sits in the middle with two screens, and a couch is pushed towards the back.

On the screens, feed from the CCTVs plays as the ANPR cameras scan people passing by. People appearing on-camera are projected with green squares around their faces. A small text above the square shows the software’s estimates of the person’s age and height. Age: 30-39, Height: 5-5.5 feet, it reads for one.

The screen is split into two halves, with one half showing possible matches to suspects. “We have set the limit to 60% and above; anything with a match rate higher than 60% will

be flagged to us,” explains Constable Kaushik.

The other screen has the criminal dossier database open.

All of a sudden, a red square appears on the screen and the other half pops up with the picture and dossier number of a criminal from their database. “Oh, looks like we’ve caught someone,” says Constable Kaushik.

He walks over to the van’s window, spots the man in question, and calls him over. As the man enters the van, Kaushik asks, “Have you ever committed a crime before?” He denies it. Kaushik then enters the dossier number provided by the system.

“Is this you? Were you arrested in 2019 under the Excise Act?” Kaushik points at the mugshot on the screen, his voice has a harsh edge to it now. The man wipes his brow and says, “Yes, but the case was dropped in 2020… my boss got into trouble, I was only working in the bar.”

Kaushik asks for the man’s Aadhaar card details and notes them down in a register. “Go to the police station with the court order and get your name removed from the dossier,” he says.

According to DCP (Northwest) Bhisham Singh, the software is sophisticated enough to recognise faces despite age or disguises. It can also recognise faces behind masks or beards.

While the van has been in use in the Northwest district since April last year, the Delhi Police has been using FRS for security for at least six years.

Following a High Court order, Automated Facial Recognition Software (AFRS) was acquired in March 2018 by the Delhi Police to identify lost and found children by matching photos. Since then, it has been used to scan crowds for possible suspects during high-profile events.

The Indian Express earlier reported that Prime Minister Narendra Modi’s Ramlila Maidan rally in December 2019 was the first political rally where police used the software to screen the crowd.

Going big

Come June, the Delhi Police hopes to expand the ambit of the FRS to all of Delhi.

The initiative, part of the ‘Safe City’ project, will see the installation of 10,000 high-resolution cameras across the Capital whose live feed will be beamed directly to a command centre at the Police Headquarters.

Over time, the feeds will be integrated with CCTV footage from “legacy networks” such as the New Delhi Municipal Council, Residents’ Welfare Associations, Market Welfare Associations, Railways, etc. “We will ask them to share their footage but there are storage concerns over replicating such vast amounts of footage,” says B S Jaiswal, who was Joint Commissioner (Tech and PI). He has recently been transferred to his next assignment as Joint CP (Central Range).

Apart from fixed cameras, Prakhar Vans with mobile cameras scan crowds and crime-prone areas. Currently, 88 Prakhar Vans — each equipped with four CCTVs — are in use across the city, said Jaiswal.

The city-wide project, however, will not be using the Israeli software. Instead, the Centre

for Development of Advanced Computing, an R&D organisation of the Ministry of Electronics and Information Technology, has undertaken the massive project.

It is in charge of not only training AI models for more accurate results on facial recognition but will also be responsible for setting up the Integrated Command, Control, Communication & Computer Centre — C4I — where integrated video feeds will be beamed.

C4I will also have a data centre, that will store collected video feeds, and two Emergency Operation Centres which will flag and deal with crimes taking place in real-time.
For this to be possible, district headquarters and police stations will be equipped with command centres where they can view live CCTV footage of their jurisdictions.
All video analytics (VA), including facial recognition and ANPR, however, will be managed centrally at the C4I data centre. “The centre will also be enabled to play footage from 1,000 cameras simultaneously, which will include those from legacy networks,” said Jaiswal.

Alongside this, the Picture Intelligence Unit (PIU) will keep track of VA data — it will not store actual video or incident data but will maintain audit logs to help in investigations.

The PIU will be given access to various government databases such as passport, CCTNS (Crime & Criminal Tracking Network & Systems), prisons, AMBIS (Automated Multi-Modal Biometric Identification System), ZIPNET, criminal dossier, and other sources.

It will then collect and categorise photos from police sources, such as newspapers, raid images, and photos sent by the public, tagging them with features like gender, age, scars, and tattoos, creating a repository for the FRS.

Challenges & solutions 

However, there are significant challenges to be surmounted before the lofty heights of such technology can be reached.

“Facial recognition has worked 100% of the time when we have tried them in constrained environments,” says Jaiswal. “Take DigiYatra for instance. There is good illumination, the angle of the person’s face is straight, the camera is fixed and at a certain height and distance… it’s easier for the system to recognise and pinpoint the identity of the person.”

However, CCTVs face a multitude of issues — the illumination of the area, angle of the camera’s placement, the subject’s distance from the camera, weather conditions, motion blur in the footage because of wind shaking the camera, and the camera’s quality — making it a challenge to get high confirmation rates with FRS.

Similarly, changes in the subject — disguises, aging, a beard or a lack of one, masks and sunglasses — can also bring false negatives in the FRS results. Nonetheless, Jaiswal assures the bugs will be ironed out before the rollout in June.

However the real challenge, he says, is training machines to read data in real-time.

According to the project proposal, ideally, the machine should take less than five seconds to pinpoint matches from an input given. “Foundation models of FRS are mammoth systems trained on mammoth datasets… these become the backbone of the models that are used for security apparatus,” says Jaiswal.

He also explains how models exported from other countries are usually trained in recognising faces from those countries, and often get confused or flag high false rejection or false acceptance rates when inputs are from other demographics. “A model used in India needs to be trained specifically on demographics,” he says.

“… we can’t scrape pictorial data from the internet anymore because of privacy concerns. There used to be publicly available datasets that one could use before but those have been withdrawn because of privacy issues,” he adds.

But even training an FRS model with tailor-made datasets presents problems.

“If we put all the pictures available to us and feed it to the system, it already knows what to look out for… it’s like training a student for an exam with a set of questions that he knows will be the only one asked,” Jaiswal says, explaining that fresh inputs, outside the datasets, would again lead to higher rates of false acceptances or rejections.

The sweet spot, the officer says, would be a range between false acceptance and rejection rates — an equal error rate.

Simply put, an Equal Error Rate (EER) is the point where a biometric system makes mistakes at an equal rate, meaning false acceptance and false rejection rates occur equally. A lower EER indicates a more accurate and efficient biometric system, especially in facial recognition technology.

Additionally, what can help improve the EER is enhancing AI and machine learning capabilities.

For this, the Delhi Police has plans as well. For its AI-based video intelligence platform, the police plan on using machine learning for systems to estimate the number of people in a crowd in dense environments such as rallies and festivals, identify jaywalkers, identify people who collapse suddenly, and, most interestingly, identify a person in distress.

“Over time, we can develop the capability of the system in sending alerts to the local police if a person is waving their hands for help or if they seem to be in trouble… machines learn like how beat officers learn — over time, beat officers learn on the job about spotting anomalous situations. It’ll be the same with machine learning…,” says Jaiswal.

While technological hurdles remain, Apar Gupta, co-founder of the Internet Freedom Foundation, argues that the more urgent concerns are legal and ethical. “India currently lacks a comprehensive legal framework governing use of facial recognition,” says Gupta.

“At a minimum, we need a robust, bespoke law that addresses its deployment. This includes a data protection law recognising facial data as sensitive personal information, with strict rules on how it’s collected, stored, and used. It also requires an electronic surveillance law dedicated to regulating technologies like FRT, setting out clear limits and independent oversight for law-enforcement agencies,” he says.

Tech comes at a cost

Experts have warned that the increasing integration of such technology across platforms may come at a cost. “Facial recognition technology is a quantum leap in the state’s ability to identify and monitor individuals. Unlike conventional CCTV, which captures footage but still preserves a degree of anonymity, FRT can instantly put a name to every face in a crowd, eliminating the practical obscurity we once had in public,” says Apar Gupta, co-founder of Internet Freedom Foundation, a digital rights organisation.

“This technology creates a biometric map of one’s face — a unique identifier as precise as a fingerprint — and links it to databases in real-time. Modern FRT systems can be integrated with analytics platforms that draw on multiple databases — social media, call detail records, vehicle registrations, and more — building an instant, 360-degree view of a person’s life. That transforms ordinary cameras into powerful tracking tools, with a chilling effect on civil liberties,” Gupta adds.

He says integrated command centres being developed across India represent a step toward predictive policing: “Some vendors even propose ‘emotion analysis’, claiming to detect intent or agitation from facial expressions — an area of research widely questioned for its scientific reliability.”

Gupta adds that such systems are notorious for reflecting and amplifying existing biases in law-enforcement data. “If the underlying data is skewed — disproportionately targeting certain neighbourhoods or communities — the algorithm’s ‘predictions’ will be equally skewed. This can result in further over-policing of marginalised groups, creating a feedback loop where these communities are constantly singled out for scrutiny.”
False matches, too, can have devastating consequences, he says. “If the police treat a poor match as a positive identification, an innocent person can be pulled into an investigation or even detained. At one point, the Delhi Police considered an 80% match as ‘positive,’ meaning a one-in-five chance of error was deemed acceptable. Even lower matches might not be discarded entirely, putting people who only vaguely resemble suspects on watchlists.”

He cites the Northeast Delhi riots as an example of where the deployment of facial recognition created real danger. “A false match can lead to wrongful arrests, criminal charges, and denial of bail. Socially, once someone is mistakenly labeled a ‘criminal’ or ‘rioter,’ their reputation may suffer permanent damage.”