Skip to main content

YOLOv4 Object Detection

This document explains in detail how to utilize NPU hardware acceleration on the Sirider S1 to infer the YOLOv4 model.

The document is divided into two parts: Quick Start and Detailed Tutorial.

Quick Start​

Radxa provides a ready-to-use YOLOv4 object detection example, allowing users to run the yolov4_tiny model on the Sirider S1 using AIPU for inference. This eliminates the need for complex model compilation and execution code setup, making it an ideal choice for those who want to quickly use AIPU without building models from scratch. For those interested in the full workflow, refer to the Detailed Tutorial section.

  • Clone the repository:

    git clone https://github.com/zifeng-radxa/siriders1_NPU_yolov4_tiny_demo.git

  • Install dependencies:

    tip

    It is recommended to use virtualenv.

    cd siriders1_NPU_yolov4_tiny_demo/demo
    pip3 install -r requirements.txt

  • Run the YOLOv4 demo program:

    python3 yolov4_aipu.py -m [mode] -i [your_input_path] -r
    # Example: python3 yolov4_aipu.py -m camera -r

    Parameter details:

    -h, --help: Print parameter information.

    -m, --mode: Select input mode, supports ['camera', 'video', 'image'].

    -i, --input: Input file path, required when mode is ['video', 'image'].

    -r, --real_time: Real-time preview.

    -s, --save: Save the output to the output folder.

    input.webp

Detailed Tutorial​

To deploy the target model on the Zhouyi Z2, there are three main steps: model conversion, compiling the inference file, and application-level programming.

Model Conversion​

tip

This step is completed on an x86 host. Before starting model conversion, install the Zhouyi SDK according to the Zhouyi AIPU SDK Installation Guide and complete the Configure the nn-compiler Environment.

  • Clone the repository:

    git clone https://github.com/zifeng-radxa/siriders1_NPU_yolov4_tiny_demo.git

  • Generate quantization data:

    cd siriders1_NPU_yolov4_tiny_demo/convert
    python3 preprocess.py

  • Generate the AIPU model:

    aipubuild tflite_yolo_v4_tinybuild.cfg
    tip

    If the aipubuild command is not found, try adding it to your path:

    export PATH=$PATH:/root/.local/bin

    The target model is generated at ./aipu_yolov4_tiny.bin.

Compile AIPU Executable Inference File​

Compile an executable file for inferring the Zhouyi Z2 AIPU model.

  • Copy the compile folder from the repository to the Zhouyi SDK.

    Copy the compile folder from the siriders1_NPU_yolov4_tiny_demo repository to YOUR_SDK_PATH/siengine. Make sure to replace YOUR_SDK_PATH with your actual path.

    cp -r compile YOUR_SDK_PATH/siengine

  • Cross-compile:

    tip

    Modify the Linux_Tool_ROOT in CMakeLists.txt to match your cross-compilation toolchain path. The default path is /opt/gcc-linaro-7.5.0-2019.12-x86_64_aarch64-linux-gnu/bin.

    Download the cross-compilation toolchain: gcc-linaro-7.5.0-2019.12-x86_64_aarch64-linux-gnu.

    cd YOUR_SDK_PATH/siengine/compile
    mkdir build && cd build
    cmake -DCMAKE_BUILD_TYPE=Release .. && make
    cd ..

    The output files are located in the out folder.

Testing on the Board​

Transfer the generated aipu_yolov4_tiny.bin model file and the files under out/linux to the Sirider S1.

Use yolov4_aipu.py to test the results:

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:linux/libs
python3 yolov4_aipu.py -m image -i YOUR_IMAGE_PATH -r