Rockchip RK3588: The Ultimate Edge AI Processor for Smart Vision Applications

RK3588 is an octa-core 64-bit processor, blending high-performance Cortex-A76 cores with efficiency-optimized Cortex-A55 cores. This powerhouse is complemented by an integrated ARM Mali-G610 MP4 GPU, ensuring smooth 4K interfaces without lag, and a built-in NPU capable of delivering up to 6 TOPS for advanced neural network processing across various AI applications.

A10-3588

As a pioneer in industrial AI solutions, NAMTSO's A10-3588 leverages the Rockchip RK3588's 6TOPS NPU to deliver unparalleled performance in vision-centric applications, demonstrating robust real-time processing capabilities in edge computing. Below is an in-depth analysis of its representative application scenarios:

1.Facial Recognition

In smart security systems, intelligent service terminals, and access control systems, the A10-3588 achieves millisecond-level face detection and feature extraction through NPU-accelerated deep learning models. Leveraging the IMX415 camera's high-definition imaging capabilities, the system supports parallel recognition of multiple faces.

2.License Plate Recognition & Traffic Management

Utilizing YOLOv8n-pose+VGG16+DenseNet CTC algorithm optimization for license plate recognition workflows, the A10-3588 enables precise character extraction and identification in high-speed toll terminals and parking facilities. Its NPU-supported TensorFlow/PyTorch framework compatibility significantly enhances model conversion efficiency and inference performance.

3.Multi-Object Tracking (MOT) & Behavior Analysis

Combining YOLOv8n model with Kalman Filter algorithms, the A10-3588 achieves continuous tracking of humans and objects in video streams (technical term: Multi-Object Tracking). Compared to YOLOv5, YOLOv8n incorporates a Decoupled Head architecture and Dynamic Positive-Negative Sample Allocation Strategy (TAL), substantially improving small object detection accuracy. The Kalman Filter enhances trajectory continuity through prediction-correction mechanisms, effectively reducing tracking loss caused by occlusion or rapid movement.

Multi-Object Tracking (MOT) Demonstration

Multi-Object Tracking (MOT)

Hardware Configuration:

A10-3588 : RK3588 chip, 6TOPS NPU

ACC-E1080A Camera : 8MP wide-angle lens with autofocus, Sony IMX415 sensor

Software Stack:

Model Training & Optimization:

The YOLOv8n model was trained using the COCO dataset, with redundant layers pruned specifically for the human body detection task, reducing the model size to under 3MB. The model was then converted to the RK3588-specific format (.rknn) using RKNN-Toolkit2.

AIgorithm Workflow

1. Detection Phase:

   The YOLOv8n outputs human bounding boxes and confidence scores per frame, with NPU inference latency of approximately 20ms

2. Tracking Phase:

   The Kalman filter predicts target positions in subsequent frames, while the Hungarian algorithm addresses the association problem between detection bounding boxes and trackers to achieve persistent ID matching.

   
   

Performance Metrics

1. 1080P video stream processing at 30fps

2. CPU utilization ≤5% in high-density scenarios(>50 targets)

What's Next?

The A10-3588, powered by RK3588's 6TOPS NPU and multimodal interface design, has become a cornerstone of edge intelligence solutions. This human tracking implementation represents merely a fraction of its application ecosystem. Subsequent articles will explore advanced topics including model quantization and multi-algorithm fusion techniques.

Best regards,

Namtso Team