Recon at NVIDIA GTC 2026

About the event

GTC is NVIDIA's conference on AI and accelerated computing. GTC 2026 ran from 16 to 19 March 2026 at the San Jose McEnery Convention Center, with workshops on 15 March. Sessions were open in person and online.

Who we met

GTC brings the people who set the direction of AI into one building. At GTC 2026 we met Jensen Huang, founder and CEO of NVIDIA, and Howard Lutnick, the United States Secretary of Commerce.

The poster

“Recon: A Real-Time Agentic AI Platform for Intelligent Traffic Incident Detection” is poster P81265 in the GTC 2026 catalogue. The authors are Niranjan Patil and Mayur Kulkarni.

NVIDIA's listing describes Recon as a real-time agentic AI platform built on NVIDIA Jetson AGX Orin. It combines multi-object detection based on NVIDIA TAO with the VILA vision language model.

The poster starts from a limit of current video incident detection systems. They use fixed rules, so they detect only a few incident types, such as a stalled vehicle or an object on the road. Recon reasons about the scene, so it can report incidents that have no predefined category.

The pipeline has three stages:

• Multi-object detection: TAO-based models detect six vehicle types, pedestrians, animals, fire and smoke, and number plates, at 96.8% mAP@0.5.
• Vision language analysis: NVIDIA VILA models answer natural language queries about the scene.
• Alert generation: a compact language model writes an alert with traffic context, for example “Debris detected: ladder in Lane 1 at km 24.6”.

The results on the poster, from a deployment on national highway infrastructure provided by NHAI:

• Vehicle based incidents (stalled, blocking, incorrectly parked): 89% accuracy, 2.4% false positive rate.
• Behaviour based incidents (unsafe merge, wrong direction, phone usage): 93% accuracy, 2.6% false positive rate.
• Emergency situations (ambulance, military and police movement): 98% accuracy, 0.5% false positive rate.
• Hazards (debris, sudden pothole, fire, smoke): 96% accuracy, 1.5% false positive rate.

Recon generated more than five times the number of alerts of a rule-based video incident detection system on the same roads.

Recon in production

The Recon camera runs on an NVIDIA Jetson Orin NX in SUPER mode. Nine AI models run at the same time: violation checks such as seatbelt, helmet, triple riding and HSRP compliance, vehicle make, model and colour, and a vision language model fine-tuned for Indian roads. Inference uses NVIDIA TensorRT, and TensorRT-LLM reaches 60 FPS.

With the vision language model, an operator can ask for “red truck with horse sticker” in English, Hindi or Marathi and get results in under half a second.

Read more

The architecture is in our press release on the NVIDIA partnership. The products are the Recon Global Shutter Camera, VIDS for incident detection and AI Search. Stellarview is a member of NVIDIA Inception, and our research team won Track 3 of the AI City Challenge 2026.