Evaluating Jetson-Inference on NVIDIA Jetson Orin NX with Sony LVDS Block Cameras, FCB EV-9520L & FCB EV-9500L

---

Real-Time Edge AI Vision Using LVDS-to-MIPI CSI-2 Camera Integration

A complete evaluation of NVIDIA Jetson Orin NX running Jetson-Inference with Sony FCB-EV9520L and FCB-EV9500L LVDS block cameras. Learn how LVDS Y/Pb/Pr 4:2:2 video is converted to MIPI CSI-2 using a custom adapter for real-time Edge AI vision, object detection, and TensorRT inference

---

Preview:

This blog presents a deep technical exploration of integrating Sony LVDS block cameras (FCB-EV9520L & FCB-EV9500L) with the NVIDIA Jetson Orin NX using a custom LVDS-to-MIPI CSI-2 interface, and evaluates system performance using Jetson-Inference.

---

We cover:

1. LVDS Y/Pb/Pr 4:2:2 → MIPI CSI-2 conversion
2. Custom camera adapter architecture
3. Real-time inference benchmarks
4. TensorRT-optimized object detection

---

Introduction

With the rapid adoption of Edge AI across surveillance, intelligent traffic systems, robotics, and industrial automation, the demand for high-quality imaging paired with high-performance AI compute is higher than ever.

The NVIDIA Jetson Orin NX, with up to 32 TOPS of AI performance, is ideal for real-time video analytics. Meanwhile, Sony FCB-EV9520L and FCB-EV9500L remain the gold standard for long-range industrial imaging—offering:

1. 30× optical zoom
2. Full HD imaging
3. Stable LVDS Y/Pb/Pr 4:2:2 output
4. High clarity optimized for distances up to 10 km

However, the challenge:

Sony’s LVDS output is not directly compatible with Jetson’s MIPI CSI-2 input.

To solve this, we developed a custom LVDS-to-MIPI CSI-2 adapter, enabling seamless integration with Jetson Orin NX. This blog evaluates the end-to-end AI inference performance using Jetson-Inference and TensorRT.

---

Objective

To benchmark real-time AI inference on Jetson Orin NX using LVDS block cameras by measuring:

1. Object detection throughput (FPS)
2. Frame-level & end-to-end latency
3. TensorRT acceleration performance
4. Overall system stability for continuous edge deployments

---

Hardware Architecture

Sony FCB-EV9520L / EV9500L Industrial Cameras

Key Features:

1. 30× optical zoom
2. Full HD 1920×1080 @ 60 FPS
3. LVDS Y/Pb/Pr 4:2:2 digital output
4. Excellent for long-range detection (up to 10 km with additional optics)

Custom LVDS-to-MIPI CSI-2 Adapter Board

A precision-engineered interface enabling direct connectivity to Jetson.

Core Features

1. LVDS → MIPI CSI-2 conversion with low jitter
2. 4-lane CSI-2 for high-bandwidth 1080p60 streaming
3. Optimized differential pair routing
4. Noise-immune 30-pin coaxial connector for Sony blocks
5. Compact low-profile form factor suitable for embedded systems

Jetson Orin NX Dev Kit

1. Ampere GPU with 1024 CUDA cores
2. Up to 32 TOPS AI compute
3. JetPack 5.1.2 (L4T 35.4.1)
4. High-speed CSI-2 capture pipeline
5. Ideal for TensorRT inference, Jetson-Inference, and real-time vision workloads

---

System Diagram

Block Diagram:

Connection Diagram:

---

Software Stack

To optimize performance, we used:

1. JetPack 5.1.2 (L4T 35.4.1)
2. Custom V4L2 driver for MIPI CSI-2 camera input
3. Jetson-Inference framework for live detection
4. SSD-MobileNet-V2 (TensorRT optimized)
5. Jetson-Utils for rendering overlays
6. OpenCV & GStreamer for preprocessing
7. Python API for rapid prototyping

Inference Pipeline Overview

1. MIPI CSI-2 Camera Input (1080p60)
2. Frame Scaling & Normalization
3. TensorRT Inference using SSD-MobileNet-V 2
4. Bounding Box Visualization
5. Real-time display or data streaming

---

Performance Evaluation

Tested using a 1920×1080 live camera feed.

Metric	Value
Inference Speed (1080p)	30+ FPS
TensorRT Latency	10–15 ms per frame
End-to-End Latency	< 50 ms
Video Stability	Excellent at 60 FPS
Thermal Behavior	Stable with compact cooling

Video Demo

▶ https://youtu.be/EWiAnwhChQ8

Shows real-time detection using LVDS-to-MIPI adapter + Jetson-Inference.

---

Key Observations

1. SSD-MobileNet-V2 provided an ideal balance of speed & accuracy
2. LVDS-to-MIPI conversion introduced near-zero latency
3. Jetson Orin NX handled prolonged operation without throttling
4. Long-range zoom capabilities allowed detection over multi-kilometer distances
5. System ran stable at 1080p60, ideal for advanced surveillance and industrial AI

---

Conclusion

By combining Sony FCB-EV9500L / EV9520L with NVIDIA Jetson Orin NX and a custom LVDS-to-MIPI interface, we achieved a production-ready real-time AI vision pipeline with low latency, high stability, and strong inference throughput.

This approach can be deployed across:

1. Smart surveillance
2. Automated inspection
3. Traffic and highway monitoring
4. Border & perimeter security
5. Industrial Edge AI analytics

The integration proves that industrial block cameras can pair seamlessly with modern AI SoCs to build scalable, long-range intelligent vision systems.

---

FAQs