Accelerate Video Analytics Development with NVIDIA Jetson Orin

Transform Data from Trillions of IoT Sensors into Valuable Insights

With over one billion network cameras, one of the most deployed IoT sensors,  in use worldwide, AI-based video analytics at the edge has become crucial for many industries. These systems enhance security, manage traffic, and capture real-time data from cameras, analyzing each frame to extract key events and store them securely. This allows businesses to gain valuable insights from video streams. However, the vast amount of video content is challenging to manage. Teams responsible for analytics often struggle to review all footage due to limited time and resources, and human error is always a risk, leaving much of the video data underutilized.

In this blog, we will introduce how the NVIDIA Jetson Orin platform accelerates the development of video analytics solutions. We’ll explain why Jetson is the ideal choice by examining its hardware design and performance, which support multiple video streams. Notably, NVIDIA’s ecosystem speeds up the engineering time for enterprises to bring products to market. The latest JetPack 6 update introduces Jetson Platform Services, a modular architecture featuring a comprehensive collection of customizable software and reusable microservices for building vision AI applications. It offers foundational services for infrastructural capabilities, AI services for generating insights, and a reference cloud for secure edge-to-cloud connectivity.

Hardware Design and Performance: Supporting Multiple Video Streams

NVIDIA Jetson Orin is the ideal choice for handling multiple cameras and complex AI video analysis tasks, particularly in edge computing scenarios where high performance, low latency, and an efficient, compact design are critical. While servers remain the mainstream choice for large-scale deployments, NVIDIA Jetson Orin offers a balanced solution for edge computing scenarios, providing high real-time performance with efficient, compact hardware. For instance, the Jetson Orin NX 16GB can handle up to 18 streams of 1080p30, making it suitable for mid-sized scenarios like supermarket sections, gas stations, and retail stores.

Edge servers with dedicated GPUs are typically used for video analytics, as they are designed to process large-scale video streams, including high-resolution (4K) or multiple 1080p video streams. However, they are not economical for smaller setups with fewer than 100 streams because edge servers come with higher power consumption and require more complex system integration. On the other hand, low-compute-power processors, such as microcontrollers, are also popular in vision AI solutions but are suited for scenarios that do not require real-time analytics and involve only single-stream detection. These processors typically send image data rather than video streams, making them less suitable for complex, real-time video analysis tasks.

To determine the appropriate frame rates for your project and the maximum number of cameras that can connect to one device, please check out our wiki tutorial on comparing the number of cameras to the model performance benchmark as a reference.

Model	AI Performance (TOPS)	CPU	Video Decode	Power Consumption
Orin Nano 4GB	20 TOPS	6-core Arm Cortex-A78AE v8.2 64-bit CPU 1.5MB L2 + 4MB L3	2x 4K30 Up to 11x 1080p30	5W – 10W
Orin Nano 8GB	40 TOPS	6-core Arm Cortex-A78AE v8.2 64-bit CPU 1.5MB L2 + 4MB L3	2x 4K30 Up to 11x 1080p30	7W – 15W
Orin NX 8GB	70 TOPS	6-core A78 Arm V8 1.5MB L2 + 4MB L3	4x 4K60	Up to 25W
Orin NX 16GB	100 TOPS	8-core A78 Arm V8 2MB L2 + 4MB L3	8x 4K60	Up to 25W
AGX Orin 32GB	200 TOPS	8-core A78 Arm V8 2MB L2 + 4MB L3	16x 4K60 32x 1080p30	Up to 60W
AGX Orin 64GB	275 TOPS	12-core A78 Arm V8 3MB L2 + 6MB L3	32x 4K60 64x 1080p30	Up to 60W

Unleashing Parallel Processing: Running Multiple Deep Learning Modles

The NVIDIA Jetson Orin series is designed to excel in parallel computing, providing robust concurrent processing capabilities. For instance, consider a video analytics solution for traffic flow analysis and block scheduling. Such a complex task cannot be addressed by a single object detection model alone. To achieve a comprehensive understanding of the current traffic situation, multiple AI tasks must be integrated, including congestion prediction through peak and trough traffic flow analysis, vehicle counting, and classification, license plate recognition, illegal turn detection, and pedestrian detection on crosswalks. Each task requires a specific AI model or algorithm. Therefore, the critical question is: how does your edge device perform when running multiple models concurrently?

The Jetson Orin series delivers up to 275 TOPS of AI performance(see Jetson MLPerf Inference Benchmarks), boasting superior GPU capabilities that outperform most other embedded AI systems. This substantial AI capability allows it to handle not only computer vision models but also run large models such as Vision Language Models (VLMs) enable semantic understanding of images and videos by combining vision modalities to LLMs. Learn more about the future of generative AI in video analytics using NVIDIA Jetson Orin on this blog. For practical development software solutions compatible with NVIDIA Jetson devices, please visit our video analytics solution page to get started.

reServer Jetson Orin Series: Local Inference Center Design for AI NVR 

For handling complex models and vast data inputs, we recommend the reServer Jetson Orin series, the AI inferencing center for the best privacy protection with minimal maintenance due to its features:

• Sufficient Local storage: Equipped with 2x drive bays for local SSD/HDD, it can cache video footage temporarily while running entirely locally. For long-term storage, it can help manage cloud transmission bandwidth by storing only critical events in the cloud.
• Hybrid connectivity: Includes 5x RJ45 GbE for multiple real-time processing streams. Four of these ports are 802.3af PSE, allowing you to integrate power supply and data transmission ports for better layout design.

Accelerate Development with a Powerful NVIDIA Ecosystem

Edge computing integrates various technologies and components, including hardware, operating systems, networking, and data processing. This complexity often leads to extended development cycles and increased project risks, especially when building sophisticated AI applications. Developers need cross-disciplinary skills and deep expertise in AI technologies to manage these challenges, such as system complexity, hardware compatibility, and software development cycles.

Introducing Jetson Platform Services: A New Modular and Flexible Architecture

To address developers’ anxieties about application development, NVIDIA has introduced Jetson Platform Services (JPS). Part of the NVIDIA JetPack SDK, JPS is designed specifically for NVIDIA Jetson modules and provides a comprehensive solution for building end-to-end accelerated AI applications. The JetPack 6 version further enhances the flexibility and scalability of the Jetson platform by introducing microservices and a range of new features, making it the most popular JetPack version in 2024.

Jetson Platform Services acts like an all-in-one toolbox, offering robust support for building vision AI applications. JPS adopts a modular architecture, which includes many customizable software components and reusable microservices. These microservices are like building blocks that developers can freely combine to create various functionalities for vision AI applications.

In this toolbox, developers can find various practical microservices. For instance, the Video Storage Toolkit (VST) helps manage video streams from cameras with ease. The AI Perception Service, based on NVIDIA DeepStream, uses advanced deep learning techniques to enable machines to “understand” the world similarly to humans. Additionally, there are generative AI inference services and analytics services, providing powerful insights and analytical capabilities for AI applications.

Success Use Case Deployed in the Field

• Security management: Deploy as an AI NVR to complete intrusion detection at home/entire workspace. Check out the security solution that we’ve implemented for the automotive dealer warehouses across 20 sites in Italy.
• Retail operation management and customer analysis: Enhance retail operations with solutions for people counting, customer preference analysis, heatmap tracking, queue management, personnel allocation, out-of-stock alerts, etc. With our ecosystem ISV partner Isarsoft, find more details about how to gain actionable insights and marketing strategy based on customer choices.
• Traffic analysis: Implement multi-angle vehicle detection, license plate detection, traffic flow monitoring, parking lot occupation detection, illegal turns, line crossing, etc. Discover our solution about how to build your own intelligent transportation solution with the no-code Lumeo AI platform.

These are just some of the latest popular vision AI solutions. Discover more case studies showcasing successful deployments worldwide!

---

Seeed NVIDIA Jetson Ecosystem

Seeed is an Elite partner for edge AI in the NVIDIA Partner Network. Explore more carrier boards, full system devices, customization services, use cases, and developer tools on Seeed’s NVIDIA Jetson ecosystem page.

Stands at the forefront of global hardware innovation, we commit to making technology accessible to all. Specializing in IoT and AI, Seeed offers a spectrum of open hardware, design, and agile manufacturing aimed at lowering the threshold of hardware innovation.

About Author