STM32MP2 + Qt6 + X-LINUX-AI: need help for custom AI object detection application

[jumman_JHINGA]

Hi ST Community,

I'm developing a custom AI application on the STM32MP257F-DK using the following software stack:

Software Environment

• Board: STM32MP257F-DK
• Starter Package: OpenSTLinux 6.6
• X-LINUX-AI Expansion Package: v6.0
• X-LINUX-QT: v2.1
• SDK: STM32MP2 SDK (cross-compilation on Ubuntu)
• Qt: Qt 6.6 (cross compiled)
• Language: C++
• IDE: Qt Creator

 

Application Goal

I want to develop my own Qt Quick application (not modify the Python demo) with the following architecture:

USB Camera
      │
      ▼
Camera Capture
      │
      ▼
AI Inference (ST AI Runtime / NPU)
      │
      ▼
Bounding Boxes
      │
      ▼
Qt Quick GUI

The objective is to build a modular C++ application that can later be extended with multiple AI models.

 

What I have already verified

The Logitech C270 webcam is correctly detected.

/dev/video7

v4l2-ctl --all reports:

• Resolution: 640x480
• Format: YUYV
• 30 FPS

I can also run:

gst-launch-1.0 v4l2src device=/dev/video7 ! videoconvert ! autovideosink

and the pipeline reaches the PLAYING state.

 

What I discovered

Initially I implemented camera capture using:

cv::VideoCapture("/dev/video7")

and also

cv::VideoCapture(..., cv::CAP_GSTREAMER)

However, read() always fails and the camera cannot be used reliably.

While inspecting the official X-LINUX-AI demo, I noticed that ST does not use cv::VideoCapture.

The official Python implementation builds a native GStreamer pipeline:

v4l2src
    →
capsfilter
    →
videoconvert
    →
videoscale
    →
appsink

Frames are then extracted from appsink and converted into a NumPy array before being passed to the AI runtime.

This suggests that ST recommends using native GStreamer instead of OpenCV VideoCapture on STM32MP2.

 

My Questions

1. For a new C++ Qt application, what is the recommended camera acquisition method?

Should we:

• use OpenCV VideoCapture, or
• use a native GStreamer pipeline with appsink, similar to the official Python demo?

 

2. Is there an official C++ example that demonstrates:

 

• GStreamer camera capture
• ST AI Runtime integration
• Qt GUI

instead of the Python/GTK implementation?

 

3. What is the recommended architecture for a production application?

 

For example:

Camera Thread
        │
        ▼
Frame Queue
        │
        ▼
AI Thread
        │
        ▼
Qt GUI Thread

or is there another architecture recommended by ST?

 

4. Is there any ST-supported C++ wrapper around the X-LINUX-AI runtime that can replace the Python implementation?

 

 

5. Would ST recommend reusing the GStreamer pipeline from the Python demo inside a Qt application, or is there another preferred multimedia framework for custom applications?

 

 

End Goal

The final application should provide:

• Live USB camera preview
• NPU-accelerated object detection
• Bounding box overlay
• Multiple AI models
• Fully custom Qt Quick interface
• C++ implementation only

I would appreciate any recommendations, architecture diagrams, official examples, or best practices for building such an application on the STM32MP257 platform.

Thank you.