What happens when a dash cam becomes more than just a recorder? At IIIT Hyderabad’s Applied AI Research Centre (INAI), low-cost cameras mounted on police and towing vehicles are being paired with artificial intelligence and computer vision to turn road footage into actionable traffic-enforcement data. The system can identify violations such as wrong-side driving, helmetless riding and triple riding, read licence plates and feed verified evidence into the e-challan system. CE spoke to Govind Krishnan, head of engineering, Mobility, INAI, to understand how the technology works, the challenges of deploying AI on the road and what lies ahead.
Excerpts
What inspired INAI to develop an intelligent traffic enforcement system?
INAI’s objective is to take research in areas such as artificial intelligence and computer vision and translate it into solutions for real-world problems. We call this translational work — taking a proof of concept, deploying it in the real world and demonstrating that it can benefit society.
One of our flagship areas is road safety through iRASTE. Road accidents have multiple causes, including infrastructure issues and driving behaviour such as wrong-side driving, overspeeding, driving under the influence and not wearing helmets or seatbelts.
While road safety has several dimensions, we are particularly interested in using technology to address driving behaviour and enforcement. How does the system identify traffic violations?
There are broadly two ways of improving driving behaviour — awareness and enforcement. The challenge is that traffic police cannot be everywhere at all times, and manpower is limited.
That is where our dash-cam-based system comes in. We have installed cameras in two towing trucks and two police patrol vehicles and have been running the pilot for around six months in partnership with the police.
The vehicles continue their regular duties while the cameras capture road footage. Our computer-vision models analyse the footage to identify violations such as wrong-side driving, riding without a helmet, triple riding and certain licence-plate violations. The system can also read licence plates.
We have integrated the system with the Telangana Police’s e-challan platform. Potential violations are first verified by our team before being sent to the police, where they undergo another level of verification.
The underlying research originated at IIIT Hyderabad, where professor Ravi Kiran developed the dash-cam-based model as a proof of concept. We enhanced and deployed it as a real-world system.
What safeguards are in place to protect citizens’ privacy?
Footage is retained for seven days and is used only for the purpose for which the system has been deployed. After that, it is deleted unless the police specifically request that particular footage be retained.
The data is stored on a secure server with restricted access.
What have been the biggest challenges in moving from the pilot towards wider deployment?
One of the biggest challenges is ensuring there is sufficient evidence before a challan is issued. If, for example, a licence plate is unclear, the violation could be contested.
That is why we have a human-in-the-loop process. Our team first reviews the evidence and rejects potential violations where it is insufficient. The police then conduct a second verification before issuing a challan.
Reducing false positives is another technical challenge. We continue to improve the models so that genuine violations are identified accurately.
The pilot has so far been demonstrated with four vehicles. The next challenge is scaling the system while maintaining the same level of reliability.
What is the roadmap ahead for INAI?
Traffic enforcement is one of several mobility applications we are working on. Our dash-cam-based solution is called Viola Dash Cam.
We also have a separate system that analyses CCTV footage and flags traffic violations, which has been piloted with the Medak police.
Dash cams offer an advantage because they are relatively inexpensive and mobile. Unlike fixed CCTV infrastructure, they do not require dedicated installation at every location and can be moved between vehicles.
We are in discussions with other police departments, including in Punjab and Haryana, although nothing has been finalised yet.
The model is unconventional because patrol and towing vehicles are already moving through different areas, allowing the cameras to capture violations as they happen. In that sense, it acts like an ‘invisible cop’, while still relying on human verification before enforcement action.
But our objective is not to punish people. We want to reduce road accidents.
If there are better ways of changing driving behaviour without issuing challans, we would be happy to explore them. For now, technology can make enforcement more consistent and scalable.
