RK3588 Dual-Light + YOLO: 9 Deployment Case Studies

Key Takeaways

  • The RK3588’s 6 TOPS NPU plus its dual ISP is enough to run visible-light and thermal YOLO inference on-device — nine documented deployments span counter-UAV, security, airborne reconnaissance and open-source frameworks.
  • Decision-level fusion (two YOLO models, results merged) gets a usable prototype in 2–4 weeks. Feature-level fusion (pseudo-siamese backbones) buys accuracy but requires a full retraining and quantization cycle.
  • Published benchmarks: 56.55 ms per frame, 45.84 FPS and 78.5% average detection accuracy for a fused UAV-detection model; 4K capture at 30 FPS on a dual-MIPI design; under 10 W total on an FPV interceptor.
  • The real bottleneck is not the NPU. It is ISP tuning, sensor synchronization, and fusion modules that survive RKNN conversion without accuracy loss.
  • If a fully domestic stack is required, the Starlight Kylin dual-MIPI reference design is the closest existing template.

In one line: the RK3588 has become the default edge platform for dual-light (visible + thermal) YOLO work, and enough projects have shipped on it that you no longer have to start from a blank schematic. Aomway has grouped nine reference cases into three buckets — industry deployments, research validations and open-source engineering — covering counter-UAV, security and airborne reconnaissance, with the numbers and the trade-offs spelled out so you can pick a starting point rather than rebuild one.

Industry Deployments: Security and Counter-UAV

1. Portable Counter-UAV Detection and Jamming System

Approach: a sense–decide–act closed loop with the RK3588 as the AI processing unit, paired with multi-camera panoramic vision and an RF jamming module. Detection, identification, tracking and countermeasure run as one automated flow, aimed at critical infrastructure and large-event security.

  • Core technology: an optimized YOLOv8 deployed on the RK3588 NPU with DCNv2 deformable convolution integrated, plus a re-matching tracker; panoramic image stitching and dehazing preprocessing; vision-guided RF cooperative countermeasures.
  • Expansion path: the native architecture can add a thermal channel directly, using the RK3588’s dual ISP hardware to reach decision-level dual-light fusion quickly and improve detection at night and in haze.
  • Why it is useful: a complete counter-UAV engineering frame covering detection, tracking, alerting and countermeasure — usable as a project prototype reference.

2. Dual-Light Gimbal Payload (Huishi Optoelectronics)

Approach: a mature commercial design from Chengdu Huishi Optoelectronics that pairs a compact dual-light gimbal with an RK3588 processing board as one unit. It can fly on a drone for reconnaissance or act as a fixed ground node for low-altitude defence.

  • Configuration: an HD visible module and a thermal imaging module as two inputs; the RK3588 board carries dual-light capture, ISP tuning, YOLO detection and target tracking end to end. The board is small and light enough for gimbals, vehicles and portable gear.
  • Why it is useful: a proven reference for industrial-grade dual-light hardware adaptation, ISP parameter tuning and integrated algorithm deployment — buy-and-integrate, cutting the low-level R&D cycle substantially.

3. FPV Autonomous Interceptor Onboard Perception (SoluTech)

Approach: an overseas FPV interceptor design with the RK3588 as the onboard compute unit, integrating visible and thermal cameras. Detection, trajectory solving and intercept guidance all run on the edge, with no back-end control loop.

  • Core technology: a lightweight YOLO quantized to INT8; RGB and thermal streams inferred in parallel; total system power held under 10 W to fit onboard battery endurance.
  • Measured behaviour: the 6 TOPS NPU sustains real-time processing of both video streams while also running target detection, model classification and intercept trajectory computation.
  • Why it is useful: a benchmark for the power–performance balance in embedded dual-light YOLO, directly applicable to portable, vehicle-mounted and airborne counter-UAV designs.

4. Border and Coastal Multi-Spectral Surveillance Gimbal

Approach: a fully domestic multi-spectral gimbal from Shantian Optics with an RK3588 core, an HD visible camera and a 640×512 thermal core, for all-weather monitoring of low-altitude and cross-border targets, with counter-UAV use as a secondary role.

  • Core technology: multi-spectral image fusion and enhancement; feature amplification for long-range small targets; integration with gimbal servo control so a detection triggers continuous tracking.
  • Why it is useful: a reference for dual-light algorithm optimization in long-focal-length, long-range small-target scenes, plus hardware stability design for outdoor industrial equipment.

Research Validations: Algorithm Optimization References

5. Pseudo-Siamese Dual-Branch YOLOv5 with Feature-Level Fusion

Approach: an academic team built a pseudo-siamese dual-branch network on YOLOv5s. Visible and thermal images each enter their own backbone to extract shallow features, a lightweight multimodal fusion module concatenates them, and a unified detection head produces the output. It was validated on-device on an OrangePi 5 Ultra (RK3588).

  • Core optimization: feature-level fusion that balances accuracy and model size; the fusion module was adapted to the RK3588 NPU’s operator support to avoid quantization accuracy loss.
  • Why it is useful: a complete “dual-light YOLO design → training → RKNN quantization → RK3588 deployment” pipeline with validation data, directly reusable for network-structure changes and quantization tuning. The right reference if you are chasing accuracy.

6. UAV-Borne Multimodal Cooperative Detection

Approach: aimed at multi-drone cooperative reconnaissance, this work designs a semantics-driven visible–thermal fusion detector deployed on the RK3588 edge platform for ground small-target detection in complex environments.

  • Core technology: a dual-light fusion model plus visual-perception enhancement preprocessing, addressing weak features under uneven lighting and occlusion; combined with distributed cooperative detection scheduling.
  • Measured performance: 56.55 ms single-frame inference latency, 45.84 FPS, and 78.5% average detection accuracy in complex scenes.
  • Why it is useful: a performance baseline for dual-light fusion on the RK3588, plus image-enhancement ideas for small targets in low visibility — transferable to long-range counter-UAV detection.

Open-Source Projects You Can Build On

7. Radxa Multi-Stream YOLOv8 Framework

Project: RK3588-Multi-Stream-YOLOv8-Detection

Approach: a multi-stream detection project designed specifically for the RK3588, driving all three NPU cores in parallel with native support for extra camera inputs — the natural base framework for a dual-light detection prototype.

  • Nine or more video streams inferred simultaneously; can be configured directly as one visible plus one thermal input.
  • Pipeline architecture: hardware decode → NPU inference → CPU post-processing → display, for minimum end-to-end latency.
  • Built-in three-core NPU scheduling, letterbox preprocessing and frame-rate statistics.

Why it is useful: add image registration, dual-light fusion and confidence weighting on top, and you can stand up a working dual-light YOLO system in 1–2 weeks.

8. Dual MIPI Cameras on a Domestic OS

Approach: a Starlight Kylin OS adaptation from Dingchang Electronics, using an RK3588 board for two-channel MIPI capture plus full YOLO-series inference, with a complete SDK and demo applications for a domestic operating system environment.

  • Hardware: IMX415 (4K visible) plus OV13855 sensors, connected through two independent MIPI CSI interfaces with the dual ISP processing each stream separately.
  • Measured performance: 30 FPS real-time inference at 4K capture; over 25 FPS for a single 1080p channel.
  • Why it is useful: a complete engineering manual for dual-camera wiring, ISP configuration and domestic-OS adaptation. Swap one sensor for a thermal module and you have a localized dual-light design.

9. Seeed Studio Containerized YOLO

Approach: a one-click Docker deployment from Seeed Studio for the reComputer RK3588 industrial unit, supporting YOLOv5, YOLOv11 and other mainstream models, with camera and stream inputs and a built-in web interface.

  • Core advantage: containerized deployment with zero environment setup; industrial-grade hardware validated for 7×24 operation.
  • Why it is useful: ideal for fast algorithm validation and POC demos. Extend it with dual-channel capture and fusion logic to shorten prototype development considerably.

Choosing Between the Three Paths

Path Fusion level Time to prototype Best for
Open-source framework plus off-the-shelf dual-light modules Decision-level 2–4 weeks Shipping fast; engineering effort matters more than peak accuracy
Pseudo-siamese dual-branch network Feature-level Full retrain and quantization cycle Accuracy-critical work with a dedicated counter-UAV dataset
Starlight Kylin dual-MIPI reference Decision-level Adaptation of an existing SDK Domestic-stack compliance for government or defence procurement

In short: for fast delivery, take the decision-level route on the Radxa framework with mature dual-light modules. For accuracy, follow the feature-level route and train and quantize the pseudo-siamese network against your own dataset. For domestic compliance, start from the Starlight Kylin adaptation, which covers the stack from hardware through operating system.

Have questions about this article? Feel free to contact us at [email protected] — we’re happy to help!

Frequently Asked Questions

What does “dual-light” mean in these deployments?

It means two imaging channels on one platform: a visible-light camera and a thermal infrared camera, processed together so the system sees both reflected light and heat. Fusion can happen at decision level (separate detectors, merged results) or feature level (shared backbone and detection head).

Should I choose decision-level or feature-level fusion?

Decision-level if you need a prototype in weeks: two YOLO models on the dual ISP, results merged, and no retraining. Feature-level if accuracy is the priority and you can afford a full training and RKNN quantization cycle, since the fusion module must be adapted to the NPU’s operator support.

Is the RK3588 NPU fast enough for two video streams?

Yes for typical dual-light workloads. Published figures include 56.55 ms per frame (45.84 FPS) for a fused detection model, 30 FPS inference at 4K capture on a dual-MIPI design, and under 10 W total on an airborne interceptor running RGB and thermal in parallel.

What is the hardest part of porting a dual-light YOLO to RK3588?

Not the model conversion. It is ISP tuning, sensor synchronization and keeping the fusion module quantization-friendly — custom operators that RKNN cannot map will either break the build or quietly cost accuracy. Budget time for ISP parameters and per-channel calibration.

Can this run on a domestic operating system?

Yes. The Starlight Kylin adaptation provides a working reference with two MIPI sensors, dual ISP and a complete SDK. It reports 30 FPS at 4K capture and over 25 FPS on a single 1080p channel, which is enough for most surveillance and counter-UAV nodes.

Source: original Chinese roundup of RK3588 dual-light + YOLO engineering cases (WeChat). Original: mp.weixin.qq.com.

About Aomway

Aomway designs and manufactures FPV video transmitters, antennas, cameras and goggle systems for drone pilots worldwide, with more than fifteen years in the UAV industry. Its hardware serves racing, long-range and professional aerial platforms where low-latency video and reliable RF links matter most.

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top