NVIDIA DeepStream 9.1: Agentic AI for Vision Analytics

NVIDIA has released DeepStream 9.1, an update addressing a long-standing issue in video analytics. Traditionally, tracking an object across multiple cameras required manual calibration and complex calculations. DeepStream 9.1 introduces two key features: Multi-View 3D Tracking (MV3DT) and AutoMagicCalib (AMC). Both are agentic skills for coding agents, designed to help developers build pipelines faster. DeepStream is NVIDIA's toolkit for AI-based video and image understanding, using a GStreamer-based framework. The new version adds 13 agentic skills, with MV3DT handling cross-camera tracking and AMC providing automatic calibration. It also supports NVIDIA JetPack 7.2 for Jetson Orin and Thor edge devices. MV3DT works by projecting detections from calibrated cameras into a shared 3D coordinate system, associating observations of the same object, and assigning a consistent global ID. It supports three object detection models, including PeopleNetTransformer. AMC calibrates cameras automatically by analyzing tracked objects in video. This update streamlines complex video tracking and calibration, potentially making AI-powered video analytics more accessible and efficient.