The latest DeepStream release introduces 13 AI tools, natural-language pipeline creation, multi-view 3D tracking, and an open-source codebase for vision AI applications.
NVIDIA has released DeepStream 9.1 as an open-source software development kit (SDK) for video analytics, introducing 13 new AI capabilities. The release introduces features such as agentic AI, natural language pipeline builder, MV3DT, AutoMagicCalib, and also support for JetPack 7.2.
Natural-language pipeline creation is one of the major additions that enables building, editing, and managing video analytics pipelines using prompts in natural language instead of manual configuration for each step in the process. According to NVIDIA, tasks can now be completed in hours instead of weeks.
The DeepStream software has been moved to NVIDIA’s unified GitHub repository. It makes it easier for developers to access the latest code, adopt updates more easily, and contribute improvements.
Another addition is Multi-View 3D Tracking (MV3DT), which helps maintain the identity of any object, whether people, vehicles, or other objects, while moving across multiple camera views within a common 3D coordinate system. The other feature of note is AutoMagicCalib, which automates camera calibration for multi-camera deployments.
DeepStream 9.1 forms part of NVIDIA’s broader Metropolis vision AI platform that includes the data center GPUs as well as the Jetson-based edge devices. According to NVIDIA, the new features will be particularly useful in applications such as smart surveillance, industrial monitoring, logistics, retail, and warehousing.
