Step-by-step guide for development and deployment of a person detector using convolution neural network (CNN) on MCU-based systems;
The following is a list of all components that are available in the ml_person_detector folder.
|Test scripts in Python for the multiple person detection model running on PC.
|The original CNN model in format of ONNX.
|Test images and quantization calibration images.
|eIQ® Inference with Glow NN.
|The ML-Person-Detector project of the i.MX RT1170EVK and RT1060EVK.
You need to have both Git and West installed, then execute below commands to gather the whole SDK delivery of the ml-person-detector.
west init -m https://github.com/nxp-mcuxpresso/appswpacks-ml-person-detector.git --mr mcux_release_github appswpacks-ml-person-detector
To build and run the application, please refer to the Lab Guide in the doc folder or check the steps in Run a project using MCUXpresso IDE.
To use the verification tool, go to the scripts folder and run below commands.
In this section, eIQ® Inference with Glow NN is applied to enable ahead-of-time compilation and convert the neural networks into object files. To follow the given deployment steps, you need to download the Glow installer from eIQ-Glow and install it into the converter folder.
Glow uses profile guided quantization, running inference to extract statistics regarding possible numeric values of each tensor within the neural network.Images in png format with the same resolution as the input should be prepared in advance. Using command below to generate yml proflie:
image-classifier.exe -input-image-dir=data/Calibration -image-mode=0to1 -image-layout=NCHW -image-channel-order=BGR -model=models/Onnx/dperson_shufflenetv2.onnx -model-input-name=input.1 -dump-profile=models/Glow/dperson_shufflenetv2.yml
Then you will get a dperson_shufflenetv2.yml under converter folder.
Bundle generation represents the model compilation to a binary object file (bundle). Bundle generation is performed using the model-compiler tool.
model-compiler.exe -model=models/Onnx/dperson_shufflenetv2.onnx -model-input=input.1,float,[1,3,192,320] -emit-bundle=models/Glow/int8_bundle -backend=CPU -target=arm -mcpu=cortex-m7 -float-abi=hard -load-profile=models/Glow/dperson_shufflenetv2.yml -quantization-schema=symmetric_with_power2_scale -quantization-precision-bias=Int8
model-compiler.exe -model=models/Onnx/dperson_shufflenetv2.onnx -model-input=input.1,float,[1,3,192,320] -emit-bundle=models/Glow/int8_cmsis_bundle -backend=CPU -target=arm -mcpu=cortex-m7 -float-abi=hard -load-profile=models/Glow/dperson_shufflenetv2.yml -quantization-schema=symmetric_with_power2_scale -quantization-precision-bias=Int8 -use-cmsis
Here are two examples of the accuracy verification for the quantized model. Although there are slight difference on the person coordinations between the outputs of original float model and quantized ones, the overall detection results are relatively reliable with good precision.
The person detection demo project are built on the NXP MCUs i.MX RT1170EVK and i.MX RT1060EVK respectively. It is well known that the ML model inference usually requires huge computation, while there is usually a single core in MCU. It means that the single core needs to handle not only the model inference task, but also the camera and display parts. To get a real-time performance for capturing the image from camera and showing frame with algorithm results on the display screen, we built a Microcontroller based Vision Intelligence Algorithms (uVITA) System based on FreeRTOS. And its structures are give as below
For other rapid-development software bundles please visit the Application Software Packs page.
For SDK examples please go to the MCUXpresso SDK and get the full delivery to be able to build and run examples that are based on other SDK components.
Official Github: https://github.com/nxp-appcodehub/ap-ml-person-detector