ObjectDetection

Object Detection Python CLI Application

Table of contents

• Overview
• Requirements
• Preparing model assets
• Running Python app

1. Overview

Object detection application for Windows on Snapdragon® with YOLOv8-Detection using ONNX runtime.

Ultralytics YOLOv8 is a machine learning model that predicts object bounding boxes and classes in images. Optimized for Qualcomm Compute platform, this application enables real-time object detection.

This Python application demonstrates how to use QNN Execution Provider to accelerate the model using the Snapdragon® Neural Processing Unit (NPU).

2. Requirements

Platform

• Windows on Snapdragon® (Qualcomm Compute platform, e.g. X Elite and X Plus)
• Windows 11
• This application is tested on ASUS Vivobook S15 (S5507).

Tools and SDK

• Python

  • This application is tested with Python 3.10.9.
  • Download and install Python 3.10.9 64-bit on your Windows on Snapdragon.
  • Required packages.
    • numpy
    • onnxruntime-qnn
    • opencv-python

• Qualcomm AI Runtime SDK : QNN SDK

  • The required QNN dependency libraries are included in onnxruntime-qnn package.
  • If you plan to use a specific version of QNN libraries, download and install Qualcomm AI Runtime SDK from Qualcomm Software Center.
  • This Python application is tested with default QNN libraries from onnxruntime-qnn and QNN v2.39.0.250926.
  • Find your QNN_SDK_ROOT. For example, QNN_SDK_ROOT = C:\Qualcomm\AIStack\QAIRT\2.39.0.250926.
  • Remember this directory if you plan to use a specific version of QNN libraries.
    • <QNN_SDK_ROOT>\lib\arm64x-windows-msvc

3. Preparing model assets

Compiling and downloading the model from Qualcomm AI Hub

Set up Python environment as described in Qualcomm AI Hub.

Follow the instruction and run the export script to produce a model optimized for on-device deployment.

python -m qai_hub_models.models.yolov8_det.export --device "Snapdragon X Elite CRD" --target-runtime onnx

After the compilation is finished, go to Qualcomm AI Hub Jobs page to check compiled result and download the ONNX model and model.data file to ./assets/ directory.

Video assets

Prepare your video assets and place into ./assets/ directory.

Model input resolution is 640 x 640.

If input video is in different resolution, it will be resized to 640 x 640.

4. Running Python app

Checking the assets directory

Please ensure that you have followed the section above and placed the following assets into the specific directory. You may change the directory if needed.

• Video assets : ./assets/
• ONNX model from Qualcomm AI Hub : ./assets/

Installing required Python packages

Open your terminal and navigate to the project directory.

pip install numpy onnxruntime-qnn opencv-python

Running object detection app via CLI

The default confidence threshold and IoU threshold for NMS are 0.5 and 0.4, respectively.

Run the application with default QNN libraries from onnxruntime-qnn package.

python .\Object_Detection_ONNX.py --onnx_path .\assets\model.onnx --video_path .\assets\people_run.mp4

You can also run the application with a specific version of QNN libraries.

python .\Object_Detection_ONNX.py --onnx_path .\assets\model.onnx --video_path .\assets\people_run.mp4 --qnn_path C:\Qualcomm\AIStack\QAIRT\2.39.0.250926\lib\arm64x-windows-msvc\QnnHtp.dll

Running with the specified confidence and IoU thresholds for NMS.

python .\Object_Detection_ONNX.py --onnx_path .\assets\model.onnx --video_path .\assets\people_run.mp4 --conf_thres 0.6 --iou_thres 0.5

Example Output

Inference is accelerated by Snapdragon® Neural Processing Unit (NPU).