YOLOv8 Multi-Stream

Multi-stream Video Object Detection with YOLOv8 on RK3588 NPU

Features

• Multi-stream Processing: Supports 9+ video streams simultaneously
• 3-core NPU Acceleration: Fully utilizes RK3588's 3 NPU cores
• Pipeline Architecture: Decode → NPU Inference → Post-processing → Display
• Real-time Display: Grid display with frame rate statistics (video FPS + inference FPS)
• Letterbox Preprocessing: Maintains aspect ratio to ensure detection accuracy

Installation

Clone Project Repository

Device

git clone https://github.com/ZIFENG278/RK3588-Multi-Stream-YOLOv8-Detection.git

Install Dependencies

tip

Please install dependencies in a Python virtual environment

Device

pip3 install rknn-toolkit-lite2 opencv-python

Usage

Basic Run (9 streams, 3 cores)

tip

--video-dir: Please specify the folder containing video files

Device

python3 main_rknn_pipeline.py --video-dir ./video

No Display Mode (suitable for testing)

Device

python3 main_rknn_pipeline.py --video-dir ./video --no-display --max-frames 100

Custom Parameters

Device

python3 main_rknn_pipeline.py \
    --num-streams 6 \
    --num-cores 3 \
    --model yolov8n-i8-3588.rknn \
    --video-dir video \
    --conf-threshold 0.5 \
    --iou-threshold 0.45

Command Line Parameters

Parameter	Default	Description
--num-streams	9	Number of video streams
--num-cores	3	NPU core count (1-3)
--model	yolov8n-i8-3588.rknn	Model file path
--video-dir	video	Video directory
--conf-threshold	0.4	Confidence threshold
--iou-threshold	0.45	NMS IoU threshold
--no-display	False	Disable display
--max-frames	None	Limit number of frames

Performance Reference

Test Conditions

Device

python main_rknn_pipeline.py \
    --num-streams 9 \
    --num-cores 3 \
    --model yolov8n-i8-3588.rknn \
    --video-dir video \
    --max-frames 1000

Hardware	Memory	Video Count	NPU Cores	Model	Quantization	Frames
ROCK5B+	8GB	9	3	yolov8n	INT8	1000

Test Results

Stage	Time (avg ms)
Decode	2.24
Preprocess	2.00
NPU Inference	25.77
Post-process	15.66
Draw	1.43
Total	47.10