Stellarview, in partnership with NVIDIA, today announced the deployment of India's first vision language model (VLM) running entirely on edge hardware for highway safety and surveillance. The system represents a new class of intelligent infrastructure where generative AI operates at the point of capture, eliminating the latency, bandwidth constraints, and reliability concerns of cloud-dependent systems.
India's national highways carry over 40% of the country's traffic on just 2% of its road network, with violations contributing to fatality rates among the highest globally. Traditional enforcement struggles to scale: manual footage review is labor-intensive, and cloud analytics introduce delays that compromise real-time intervention. Stellarview's edge-native architecture ensures every frame is processed locally with full AI capability regardless of network conditions.
Technical Architecture
Running on NVIDIA Jetson Orin NX in SUPER mode — a high-performance configuration highlighted at GTC — the system executes nine concurrent AI models: detecting various violations like seatbelt, helmet, triple-riding, HSRP-compliance, vehicle make, model, color, on top of a fine-tuned VLM optimized for Indian road conditions. All inference uses NVIDIA TensorRT, and TensorRT LLM achieves 60 FPS processing.
The VLM component addresses a fundamental limitation of traditional video analytics: the inability to search for what wasn't explicitly programmed. Operators can now query footage using natural language in English, Hindi, or Marathi — with queries like "red truck with horse sticker" or "man wearing orange turban on motorcycle" returning results in under half a second. What previously required hours of manual review becomes instant retrieval.
The system pairs its AI stack with Global Shutter imaging, capturing frames instantaneously to enable accurate detection at speeds up to 230 km/h — critical for evidentiary validity on expressways. Currently deployed across NH-4, Multi-Lane Free Flow tolling infrastructure, and Lucknow Ring Road, the system has processed over 20 million vehicles and identified more than 1 million violations across six categories, fully compliant with NHAI requirements.
"India's highways deserve Indian-built AI solutions. Together with NVIDIA, we're proving that cutting-edge technology for national infrastructure can be developed and deployed locally, at scale, and without compromise."