Multi-cam AI video analytics using NVIDIA DeepStream SDK 7.1
Abhiram Anil, Agilan Vellalore Saminathadurairaj, Dheeraj Swaroop Saligrama Mahesh, Divya Ramesh Teli, Farzad Mehrdad, Prajwal Mangaluru, Shilpa Gopakumar, Thomas Schumann · IET conference proceedings. · 2026
This paper implements a real-time, multi-camera anomaly detection system using the NVIDIA DeepStream SDK 7.1 on the Jetson Orin Nano platform. The pipeline was extended to support live streams from Intel RealSense cameras via the GStreamer plugin v4l2src, enabling AI-driven video analytics at 1280×720 and 30 FPS. Inference is handled by nvinfer or nvinferserver (plugins from NVIDIA DeepStream SDK), with motion detection using dsdirection. A custom shell script was developed to serve as a wrapper layer, automating multi-camera setup, GStreamer previews, and parallel processing, while also managing device control and synchronization. A dual-view display shows both raw and processed outputs. GPU benchmarking on Jetson Orin Nano with two cameras showed 98.4% usage at high resolution and lower load at reduced resolutions, demonstrating efficiency for low-cost Edge AI.