Quickrun

Quickrun is a software designed for the efficient and high-concurrency deployment of multiple models on the RK3588 platform with RKNN.

---

Software Framework

1. Session-Based Design:

   • Implements a session concept, allowing the definition of multiple sessions for different task requirements.
   • Example: Supports scenarios like charging pile detection, garbage classification, and cliff detection where multiple models share the same camera, e.g., using YOLOv5.

2. Message Queue for Data Management:

   • Uses a message queue to store photo data, preventing frame loss and ensuring efficient concurrency.
   • Photo data is collected at 25fps, with a total processing time (pre-processing, inference, post-processing) of 40ms per frame.

3. Input Processing:

   • Handles model input of 640x640 while the camera input is 640x480.
   • Decodes the input using cv::imdecode, converts it to RGB format, and leverages RGA for accelerated proportional scaling.

4. Threaded Model Execution:

   • Three independent threads for three sessions to execute models concurrently without interference.

---

Model Output Customization

• For RK3588 YOLOv5 model:
  • When converting to ONNX, remove the cat operation in the forward layer.
  • Configure the model to output three feature maps: 20x20, 40x40, and 80x80.
  • Modify the necessary files: yolo.py and export.py.

---

Performance Overview

• Resource Usage:

  • One model uses:
    • 1.2T of the NPU.
    • 40% of the CPU for pre-processing, inference, and post-processing (including frame drawing).

• Inference Time:

  • Achieves an inference time of 20ms.

• CPU Monitoring:

  • Use the perf top -p command to view the CPU usage rate, down to specific functions.

---

Quick Start

Compilation

Run the following command to build the project:

bash build_rk3588_yolov5.sh

Testing

Run the following command to test the deployment:

bash test_rk3588_yolov5.sh

Quickrun Deployment

Modify the following parameters in the code as per your project’s requirements:

#define OBJ_NAME_MAX_SIZE 16  
#define OBJ_NUMB_MAX_SIZE 64  
#define OBJ_CLASS_NUM 1      
#define NMS_THRESH 0.25       
#define BOX_THRESH 0.5 

Robot Demo Video

View Robot Video on GitHub

机器人视频观看 You can also watch the video below:

Your browser does not support the video tag.

License

Quickrun is licensed under the MIT License by Jim.