https://github.com/crab2rab/RKNN-YOLOV5-BatchInference-MultiThreading

RKNN-YOLOV5-BatchInference-MultiThreadingYOLOV5多张图片多线程C++推理
https://github.com/crab2rab/RKNN-YOLOV5-BatchInference-MultiThreading

Last synced: 2 months ago
JSON representation

RKNN-YOLOV5-BatchInference-MultiThreadingYOLOV5多张图片多线程C++推理

• Host: GitHub
• URL: https://github.com/crab2rab/RKNN-YOLOV5-BatchInference-MultiThreading
• Owner: crab2rab
• License: apache-2.0
• Created: 2023-11-01T08:27:08.000Z (over 1 year ago)
• Default Branch: main
• Last Pushed: 2023-11-06T05:45:18.000Z (over 1 year ago)
• Last Synced: 2024-08-23T20:47:42.479Z (6 months ago)
• Language: C++
• Homepage:
• Size: 21.5 MB
• Stars: 11
• Watchers: 1
• Forks: 1
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

• awesome-yolo-object-detection - crab2rab/RKNN-YOLOV5-BatchInference-MultiThreading - YOLOV5-BatchInference-MultiThreading?style=social"/> : RKNN-YOLOV5-BatchInference-MultiThreadingYOLOV5多张图片多线程C++推理。 (Lighter and Deployment Frameworks)
• awesome-yolo-object-detection - crab2rab/RKNN-YOLOV5-BatchInference-MultiThreading - YOLOV5-BatchInference-MultiThreading?style=social"/> : RKNN-YOLOV5-BatchInference-MultiThreadingYOLOV5多张图片多线程C++推理。 (Lighter and Deployment Frameworks)

README

        
# 简介

* 此仓库为c++实现yolo5的batch多线程推理, 大体改自https://github.com/leafqycc/rknn-cpp-Multithreading

* 实现4路摄像头的batch=4的推理

# 更新说明

# 使用说明

### 模型转换

  * yolov5工程下export.py转换onnx，直接转，切忌修改batch_size

  * 参考网上方法，rknn-toolkit2进行onnx转rknn，需要修改./rknn-toolkit2/examples/onnx/test.py为本工程test.py

### 演示

  * 系统需安装有**OpenCV**，如果需要读取AHD摄像头还需要安装**gstreamer**

  * 运行build-linux_RK3588.sh

  * 可切换至root用户运行performance.sh定频提高性能和稳定性

  * 编译完成后进入install运行命令./rknn_yolov5_demo **模型所在路径**

### 部署应用

  * 修改include/rknnPool.hpp中的rknn_lite类

  * 修改inclue/rknnPool.hpp中的rknnPool类的构造函数

# 多线程模型帧率测试

* 使用performance.sh进行CPU/NPU定频尽量减少误差

* 设置线程数为8，batch_size=4，读取USB摄像头视频流测试，平均处理速度15.8FPS*4，八核CPU占用率约500%，三核NPU平均占用率75%

![示例](./view.png)

# 补充

* 异常处理尚未完善, 目前仅支持rk3588/rk3588s下的运行

# Acknowledgements

* https://github.com/rockchip-linux/rknpu2

* https://github.com/senlinzhan/dpool

* https://github.com/ultralytics/yolov5

* https://github.com/airockchip/rknn_model_zoo

* https://github.com/rockchip-linux/rknn-toolkit2

* https://github.com/leafqycc/rknn-cpp-Multithreading