What does Critical Thinking mean to Developers?

TECH

June 27, 2026

Build a Simple AI Object Detection with Qt

If you want a practical “Qt + AI” demo that actually feels like a real app, object detection on Android is a great starting point: open the phone camera, run a small neural network (ONNX), and draw bounding boxes on top of the preview.

This tutorial shows a minimal but solid pipeline for Qt 6.8+ on Android using:

Qt Quick + Qt Multimedia for camera capture
C++ worker thread for inference (so UI stays smooth)
OpenCV DNN to run an ONNX model (YOLO-style)
A QML UI to display the annotated frames

You’ll end up with a working app that prints detection lines (label, confidence, box) and shows boxes on the video.

What you’ll build

Data flow

Android Camera → QMediaCaptureSession + QVideoSink → QVideoFrame → QImage
QImage → OpenCV (cv::Mat) → net.forward() → detections
Draw boxes → annotated QImage
QML shows image://annotated/frame?... (updates every frame)

Why this approach is good on Android

Qt Multimedia handles camera access reliably (and supports modern devices)
Inference runs in a separate thread
No heavy “engine” architecture—easy to integrate into your real project later

Prerequisites (Android-focused)

Required

Qt 6.8 or newer with Android kits installed (Qt Creator recommended)
Android SDK + NDK configured in Qt Creator with Android Qt 6.8.3 Clang arm64_v8a kit
OpenCV 4.x built/available for Android (with DNN enabled) (download here: https://github.com/opencv/opencv/releases/download/4.12.0/opencv-4.12.0-android-sdk.zip)
An ONNX object detection model (YOLO-style). Recommended for mobile: yolo11n for 320x320 image size (https://huggingface.co/giangndm/yolo11-onnx/resolve/main/yolo11n_320.onnx?download=true).

Project structure

QtAIDetectAndroid/
CMakeLists.txt
main.cpp
android/AndroidManifest.xml
qml/Main.qml

FrameGrabber.h/.cpp
DetectorWorker.h/.cpp
AnnotatedImageProvider.h/.cpp

third_party/ OpenCV-android-sdk #Extract downloaded OpenCV for Android here

assests/model.onnx assests/labels.txt #(contains model + labels)

Step 1 — Android permissions (Manifest + runtime)

1) Add camera permission to android/AndroidManifest.xml

In Qt Creator: Projects → Build → Build Steps → Build Android APK → “Android package source directory” point it to your android/ folder.

Minimal manifest additions:

2) Request runtime permission in Qt (important)

Android requires runtime permission on modern versions. Qt provides a clean API.
In main.cpp, before starting the camera, request QCameraPermission. (Code is included later in the tutorial.)

Step 2— CMake for Qt 6.8 + QML + OpenCV

CMakeLists.txt (core idea):

Click here for detail...

cmake_minimum_required(VERSION 3.21)
set(APP_TARGET "QtAIDetectAndroid")
set(QML_URI "QtAIDetectAndroid")

project(${APP_TARGET} LANGUAGES CXX)

set(CMAKE_CXX_STANDARD 17)
set(CMAKE_CXX_STANDARD_REQUIRED ON)

find_package(Qt6 6.8 REQUIRED COMPONENTS Quick Multimedia)

qt_standard_project_setup()

# ---- OpenCV Android SDK ----
if(ANDROID)
set(OPENCV_ANDROID_SDK "${CMAKE_CURRENT_LIST_DIR}/third_party/Opencv-android-sdk")
set(OpenCV_DIR "${OPENCV_ANDROID_SDK}/sdk/native/jni")
endif()

find_package(OpenCV REQUIRED)

# ---- App ----
qt_add_executable(${APP_TARGET}
MANUAL_FINALIZATION
main.cpp
)

qt_add_qml_module(${APP_TARGET}
URI ${QML_URI}
VERSION 1.0
QML_FILES
qml/Main.qml

RESOURCES
NO_RESOURCE_TARGET_PATH
SOURCES
    FrameGrabber.h FrameGrabber.cpp
    DetectorWorker.h DetectorWorker.cpp
    AnnotatedImageProvider.h AnnotatedImageProvider.cpp
    android/AndroidManifest.xml
)

# Embed model + labels into Qt resources at ":/assets/..."
qt_add_resources(${APP_TARGET} app_assets
PREFIX "/assets"
BASE "assets"
FILES
assets/model.onnx
assets/labels.txt
)

target_link_libraries(${APP_TARGET}
PRIVATE Qt6::Quick Qt6::Multimedia
PRIVATE ${OpenCV_LIBS}
)

target_include_directories(${APP_TARGET}
PRIVATE ${OpenCV_INCLUDE_DIRS}
)

if(ANDROID)
set_property(TARGET ${APP_TARGET} PROPERTY
QT_ANDROID_PACKAGE_SOURCE_DIR "${CMAKE_CURRENT_LIST_DIR}/android"
)

set(_opencv_so "${OPENCV_ANDROID_SDK}/sdk/native/libs/${ANDROID_ABI}/libopencv_java4.so")
if(NOT EXISTS "${_opencv_so}")
set(_opencv_so "${OPENCV_ANDROID_SDK}/sdk/native/libs/${ANDROID_ABI}/libopencv_java3.so")
endif()

if(NOT EXISTS "${_opencv_so}")
message(FATAL_ERROR "OpenCV .so not found for ABI='${ANDROID_ABI}'")
endif()

set_property(TARGET ${APP_TARGET} APPEND PROPERTY
QT_ANDROID_EXTRA_LIBS "${_opencv_so}"
)

target_link_libraries(${APP_TARGET} PRIVATE log)
endif()

if(ANDROID)
    set_target_properties(${APP_TARGET} PROPERTIES
            QT_ANDROID_PACKAGE_SOURCE_DIR "${CMAKE_CURRENT_SOURCE_DIR}/android"
            QT_ANDROID_PACKAGE_NAME "org.qtproject.QtAIDetectAndroid"
            QT_ANDROID_APP_NAME ${APP_TARGET}
            QT_ANDROID_TARGET_SDK_VERSION 31
            QT_ANDROID_MIN_SDK_VERSION 31
            QT_ANDROID_VERSION_NAME "1.0")

qt_finalize_executable(${APP_TARGET})
endif()

install(TARGETS ${APP_TARGET}
    BUNDLE DESTINATION .
    LIBRARY DESTINATION ${CMAKE_INSTALL_LIBDIR}
    RUNTIME DESTINATION ${CMAKE_INSTALL_BINDIR}
)

qt_generate_deploy_qml_app_script(
    TARGET ${APP_TARGET}
    OUTPUT_SCRIPT deploy_script
    MACOS_BUNDLE_POST_BUILD
    NO_UNSUPPORTED_PLATFORM_ERROR
    DEPLOY_USER_QML_MODULES_ON_UNSUPPORTED_PLATFORM
)

install(SCRIPT ${deploy_script})

OpenCV on Android: find_package(OpenCV REQUIRED) works if your OpenCV Android build is correctly provided to CMake (commonly via OpenCV_DIR). In Qt Creator you can set CMake cache variables (e.g., -DOpenCV_DIR=...).

Step 3 — Frame grabbing with QVideoSink

FrameGrabber.h

Click here for detail...

#pragma once
#include <QObject>
#include <QVideoSink>
#include <QImage>

class FrameGrabber : public QObject {
Q_OBJECT
Q_PROPERTY(int frameCounter READ frameCounter NOTIFY frameCounterChanged)

public:
    explicit FrameGrabber(QObject *parent = nullptr);
    QVideoSink* videoSink() { return &m_sink; }
    int frameCounter() const { return m_frameCounter; }

signals:
void frameReady(const QImage &img);
void frameCounterChanged();

public slots:
void onAnnotatedFrameReady(const QImage &img);

private slots:
void onVideoFrameChanged(const QVideoFrame &frame);

private:
QVideoSink m_sink;
int m_frameCounter = 0;

    // Simple throttle (optional): process only every Nth frame
    int m_skip = 1;     // set to 2 or 3 on slower phones
    int m_count = 0;
};

FrameGrabber.cpp

Click here for detail...

#include "FrameGrabber.h"
#include <QVideoFrame>

FrameGrabber::FrameGrabber(QObject *parent) : QObject(parent) {
connect(&m_sink, &QVideoSink::videoFrameChanged,
this, &FrameGrabber::onVideoFrameChanged);
}

void FrameGrabber::onVideoFrameChanged(const QVideoFrame &frame) {
if (!frame.isValid())
return;

if (++m_count % m_skip != 0)
return;

QImage img = frame.toImage();

if (img.isNull())
return;

emit frameReady(img);
}

void FrameGrabber::onAnnotatedFrameReady(const QImage &) {
++m_frameCounter;
emit frameCounterChanged();
}

Step 4 — Object detection worker (OpenCV DNN + ONNX)

This worker:

converts QImage → cv::Mat
runs model
applies threshold + NMS
draws boxes and emits annotated image

DetectorWorker.h

Click here for detail...

#pragma once
#include <QObject>
#include <QImage>
#include <QMutex>
#include <QStringList>
#include <QVector>
#include <QRectF>
#include <opencv2/dnn.hpp>

class DetectorWorker : public QObject
{
Q_OBJECT
public:
explicit DetectorWorker(QObject *parent = nullptr);

public slots:
void loadModel(const QString &onnxPath);

    // Optional configuration
    void loadLabels(const QString &labelsPath);          // supports ":/..." and "qrc:/..."
    void setInputSize(int w, int h);                     // e.g. 320,320
    void setUseLetterbox(bool enabled);                  // recommended for YOLO exports
    void setThresholds(float confThres, float nmsThres); // tune

    // Call for every camera frame. Internally runs one inference at a time
    // and keeps only the latest pending frame (latest-frame-wins).
    void processFrame(const QImage &img);

signals:
void annotatedFrameReady(const QImage &img);
void detectionsTextReady(const QStringList &lines);

private slots:
void doWork(); // runs in worker thread

private:
    struct Det {
        QRectF box;
        int classId;
        float score;
    };

    struct LetterboxInfo {
        float r = 1.0f; // scale
        int dw = 0;      // padding left
        int dh = 0;      // padding top
    };

static cv::Mat qimageToBgrMat(const QImage &img);
cv::Mat letterbox(const cv::Mat &srcBgr, LetterboxInfo &info) const;

QVector<Det> runYoloFromOutput(const cv::Mat &outRaw, int imgW, int imgH, const LetterboxInfo &lb);
QImage drawDetections(const QImage &src, const QVector<Det> &dets);

private:
cv::dnn::Net m_net;
bool m_ready = false;

    // Model / postprocess settings
    int m_inW = 320;
    int m_inH = 320;
    bool m_useLetterbox = true;

float m_confThres = 0.20f; // lower default for 320
float m_nmsThres = 0.45f;

QStringList m_labels;

    // Backpressure state
    bool m_busy = false;
    QImage m_current;
    QImage m_pending;
    QMutex m_mutex;
};

DetectorWorker.cpp

Click here for detail...

#include "DetectorWorker.h"
#include <QDebug>
#include <QFile>
#include <QPainter>
#include <QMetaObject>
#include <opencv2/imgproc.hpp>
#include <opencv2/core.hpp>

static QString normalizeQrc(const QString &path)
{
    // QML often uses qrc:/..., QFile uses :/...
    if (path.startsWith("qrc:/"))
        return ":" + path.mid(3); // "qrc:/x" -> ":/x"
    return path;
}

static inline float sigmoidf(float x)
{
    if (x >= 0.0f) {
        const float z = std::exp(-x);
        return 1.0f / (1.0f + z);
    } else {
        const float z = std::exp(x);
        return z / (1.0f + z);
    }
}

static inline float scoreToProb(float s)
{
    // If the model outputs logits (<0 or >1), convert via sigmoid.
    if (s < 0.0f || s > 1.0f)
        return sigmoidf(s);
    return s;
}

cv::Mat DetectorWorker::qimageToBgrMat(const QImage &img)
{
QImage tmp = img.convertToFormat(QImage::Format_RGBA8888);

    cv::Mat rgba(tmp.height(), tmp.width(), CV_8UC4,
                 const_cast<uchar*>(tmp.bits()),
                 tmp.bytesPerLine());

cv::Mat bgr;
cv::cvtColor(rgba, bgr, cv::COLOR_RGBA2BGR); // allocates bgr

return bgr;
}

DetectorWorker::DetectorWorker(QObject *parent) : QObject(parent) {}

void DetectorWorker::setInputSize(int w, int h)
{
if (w > 0) m_inW = w;
if (h > 0) m_inH = h;
}

void DetectorWorker::setUseLetterbox(bool enabled)
{
m_useLetterbox = enabled;
}

void DetectorWorker::setThresholds(float confThres, float nmsThres)
{
m_confThres = confThres;
m_nmsThres = nmsThres;
}

void DetectorWorker::loadLabels(const QString &labelsPath)
{
const QString p = normalizeQrc(labelsPath);
QFile f(p);

if (!f.open(QIODevice::ReadOnly | QIODevice::Text)) {
qWarning() << "Failed to open labels file:" << labelsPath << f.errorString();

return;
}

QStringList labels;

    while (!f.atEnd()) {
        const QString line = QString::fromUtf8(f.readLine()).trimmed();
        if (!line.isEmpty())
            labels << line;
    }

f.close();

m_labels = labels;
qDebug() << "Loaded labels:" << m_labels.size() << "from" << labelsPath;
}

void DetectorWorker::loadModel(const QString &onnxPath)
{
    m_ready = false;
    try {
        m_net = cv::dnn::readNetFromONNX(onnxPath.toStdString());
        m_net.setPreferableBackend(cv::dnn::DNN_BACKEND_OPENCV);
        m_net.setPreferableTarget(cv::dnn::DNN_TARGET_CPU);
        m_ready = true;

        qDebug() << "Model loaded:" << onnxPath;
    } catch (const cv::Exception &e) {
        qWarning() << "Model load failed:" << e.what();
        m_ready = false;
    }
}

cv::Mat DetectorWorker::letterbox(const cv::Mat &srcBgr, LetterboxInfo &info) const
{
    const int srcW = srcBgr.cols;
    const int srcH = srcBgr.rows;
    const float r = std::min((float)m_inW / (float)srcW, (float)m_inH / (float)srcH);
    const int newW = (int)std::round(srcW * r);
    const int newH = (int)std::round(srcH * r);
    const int dw = (m_inW - newW) / 2;
    const int dh = (m_inH - newH) / 2;
    info.r = r;
    info.dw = dw;
    info.dh = dh;
    cv::Mat resized;

    cv::resize(srcBgr, resized, cv::Size(newW, newH), 0, 0, cv::INTER_LINEAR);
    cv::Mat out(m_inH, m_inW, CV_8UC3, cv::Scalar(114, 114, 114));
    resized.copyTo(out(cv::Rect(dw, dh, newW, newH)));

return out;
}

void DetectorWorker::processFrame(const QImage &img)
{
if (!m_ready || img.isNull())
return;

    bool shouldStart = false;
    {
        QMutexLocker lock(&m_mutex);
        if (!m_busy) {
            m_busy = true;
            m_current = img; // implicit-share
            shouldStart = true;
        } else {
            // Overwrite pending with newest
            m_pending = img;
        }
    }

    if (shouldStart) {
        // Run in worker thread event loop; yields between frames
        QMetaObject::invokeMethod(this, &DetectorWorker::doWork, Qt::QueuedConnection);
   }
}

void DetectorWorker::doWork()
{
    QImage img;
    {
        QMutexLocker lock(&m_mutex);
        img = m_current;
    }

    if (!m_ready || img.isNull()) {
        QMutexLocker lock(&m_mutex);
        m_busy = false;

return;
}

// Convert & preprocess
const cv::Mat bgrSrc = qimageToBgrMat(img);

    LetterboxInfo lb;
    cv::Mat bgrInput;
    if (m_useLetterbox) {
        bgrInput = letterbox(bgrSrc, lb);
    } else {
        lb = LetterboxInfo{};
        cv::resize(bgrSrc, bgrInput, cv::Size(m_inW, m_inH));
    }

    // Blob: swapRB=true => BGR->RGB
    cv::Mat blob = cv::dnn::blobFromImage(
        bgrInput, 1.0/255.0, cv::Size(m_inW, m_inH),
        cv::Scalar(), true, false
        );

m_net.setInput(blob);
std::vector<cv::Mat> outs;

try {
        m_net.forward(outs, m_net.getUnconnectedOutLayersNames());
    } catch (...) {
        outs.clear();
        outs.push_back(m_net.forward());
    }

    QVector<Det> dets;
    if (!outs.empty())
        dets = runYoloFromOutput(outs[0], img.width(), img.height(), lb);

    // Emit text
    QStringList lines;
    lines.reserve(dets.size());
    for (const auto &d : dets) {
        const QString label =
            (d.classId >= 0 && d.classId < m_labels.size()) ? m_labels[d.classId] : QString("class_%1").arg(d.classId);

        lines << QString("%1 score=%2 box=[%3,%4,%5,%6]")
                     .arg(label)
                     .arg(d.score, 0, 'f', 2)
                     .arg(d.box.x(), 0, 'f', 0)
                     .arg(d.box.y(), 0, 'f', 0)
                     .arg(d.box.width(), 0, 'f', 0)
                     .arg(d.box.height(), 0, 'f', 0);
    }

emit detectionsTextReady(lines);
emit annotatedFrameReady(drawDetections(img, dets));

    // Schedule next if pending exists
    bool hasNext = false;
    {
        QMutexLocker lock(&m_mutex);
        if (!m_pending.isNull()) {
            m_current = m_pending;
            m_pending = QImage();
            hasNext = true;
        } else {
            m_busy = false;
        }
    }

    if (hasNext) {
        QMetaObject::invokeMethod(this, &DetectorWorker::doWork, Qt::QueuedConnection);
    }
}

QVector<DetectorWorker::Det> DetectorWorker::runYoloFromOutput(const cv::Mat &outRaw, int imgW, int imgH, const LetterboxInfo &lb)
{
    // Ensure float32 for safe reading
    cv::Mat out;
    if (outRaw.depth() != CV_32F)
        outRaw.convertTo(out, CV_32F);
    else
        out = outRaw;

    // Debug shape (prints a few times)
    static int dbg = 0;
    if (dbg < 5) {
        dbg++;
        QString shape = "dims=" + QString::number(out.dims);

        for (int i = 0; i < out.dims; ++i)
            shape += " " + QString::number(out.size[i]);
        qDebug() << "ONNX output:" << shape << "type=" << out.type();

        cv::Mat flat = out.reshape(1, 1);
        double mn=0, mx=0;
        cv::minMaxLoc(flat, &mn, &mx);
        qDebug() << "output min/max:" << mn << mx;
    }

    QVector<Det> candidates;
    auto clamp01orPixels = [&](float &x, float &y, float &w, float &h, int refW, int refH) {
        if (x <= 1.5f && y <= 1.5f && w <= 1.5f && h <= 1.5f) {
            x *= (float)refW;
            y *= (float)refH;
            w *= (float)refW;
            h *= (float)refH;
        }
    };

    auto mapXYXYFromInputToImage = [&](float &x1, float &y1, float &x2, float &y2) {
        if (m_useLetterbox) {
            x1 = (x1 - (float)lb.dw) / lb.r;
            y1 = (y1 - (float)lb.dh) / lb.r;
            x2 = (x2 - (float)lb.dw) / lb.r;
            y2 = (y2 - (float)lb.dh) / lb.r;
        } else {
            x1 = x1 * imgW / (float)m_inW;
            x2 = x2 * imgW / (float)m_inW;
            y1 = y1 * imgH / (float)m_inH;
            y2 = y2 * imgH / (float)m_inH;
        }

        x1 = std::max(0.0f, std::min(x1, (float)imgW - 1.0f));
        y1 = std::max(0.0f, std::min(y1, (float)imgH - 1.0f));
        x2 = std::max(0.0f, std::min(x2, (float)imgW - 1.0f));
        y2 = std::max(0.0f, std::min(y2, (float)imgH - 1.0f));
    };

    auto addCandidateXYWH = [&](float cx, float cy, float w, float h, float score, int cls) {
        if (score < m_confThres)
            return;

clamp01orPixels(cx, cy, w, h, m_inW, m_inH);

        float x1 = cx - w * 0.5f;
        float y1 = cy - h * 0.5f;
        float x2 = cx + w * 0.5f;
        float y2 = cy + h * 0.5f;

mapXYXYFromInputToImage(x1, y1, x2, y2);

const float ww = std::max(0.0f, x2 - x1);
const float hh = std::max(0.0f, y2 - y1);

if (ww <= 1.0f || hh <= 1.0f)
return;

candidates.push_back({ QRectF(x1, y1, ww, hh), cls, score });
};

    // Your model prints: dims=3 1 84 2100 -> treat as [1, C, N] with C=84 (4 + 80)
    if (out.dims == 3 && out.size[0] == 1 && out.size[1] >= 6 && out.size[2] > 100) {
        const int C = out.size[1];
        const int N = out.size[2];
        const float *p = (const float*)out.data;
        const int labelCount = m_labels.size();
        const int clsStart = 4;
        const int clsEnd = (labelCount > 0) ? std::min(C, clsStart + labelCount) : C;

        // Print best-score range once to help tuning threshold
        static bool printedMax = false;
        float globalMax = 0.f;
        int globalIdx = -1;

        for (int i = 0; i < N; ++i) {
            float cx = p[0*N + i];
            float cy = p[1*N + i];
            float w = p[2*N + i];
            float h = p[3*N + i];
            int bestCls = -1;
            float bestScore = 0.f;

for (int c = clsStart; c < clsEnd; ++c) {
                float s = scoreToProb(p[c*N + i]);
                if (s > bestScore) { bestScore = s; bestCls = c - clsStart; }
            }

            if (!printedMax && bestScore > globalMax) {
                globalMax = bestScore;
                globalIdx = bestCls;
            }

addCandidateXYWH(cx, cy, w, h, bestScore, bestCls);
}

        if (!printedMax) {
            printedMax = true;
            qDebug() << "Max class score seen (first frame):" << globalMax
                     << "classIndex=" << globalIdx
                     << "(try confThres ~0.15-0.30)";
        }
    }

    // Fallback: [1, N, 5+classes]
    else if (out.dims == 3 && out.size[0] == 1 && out.size[2] >= 6) {
        const int N = out.size[1];
        const int C = out.size[2];
        const float *p = (const float*)out.data;
        const int labelCount = m_labels.size();
        const int clsStart = 5;
        const int clsEnd = (labelCount > 0) ? std::min(C, clsStart + labelCount) : C;

        for (int i = 0; i < N; ++i) {
            float cx = p[i*C + 0], cy = p[i*C + 1], w = p[i*C + 2], h = p[i*C + 3];
            float obj = scoreToProb(p[i*C + 4]);
            int bestCls = -1;
            float bestClsScore = 0.f;

            for (int c = clsStart; c < clsEnd; ++c) {
                float s = scoreToProb(p[i*C + c]);
                if (s > bestClsScore) { bestClsScore = s; bestCls = c - clsStart; }
            }

            addCandidateXYWH(cx, cy, w, h, obj * bestClsScore, bestCls);
        }
    }
    else {
        qWarning() << "Unsupported output shape. dims=" << out.dims;
        return {};
    }

    // NMS
    std::vector<cv::Rect> boxes;
    std::vector<float> scores;
    boxes.reserve((size_t)candidates.size());
    scores.reserve((size_t)candidates.size());

    for (const auto &d : candidates) {
        boxes.emplace_back((int)d.box.x(), (int)d.box.y(),
                           (int)d.box.width(), (int)d.box.height());
        scores.push_back(d.score);
    }

    std::vector<int> keep;
    cv::dnn::NMSBoxes(boxes, scores, m_confThres, m_nmsThres, keep);
    QVector<Det> finalDets;
    finalDets.reserve((int)keep.size());

for (int idx : keep)
finalDets.push_back(candidates[idx]);

return finalDets;
}

QImage DetectorWorker::drawDetections(const QImage &src, const QVector<Det> &dets)
{
QImage outImg = src.convertToFormat(QImage::Format_RGBA8888);

QPainter p(&outImg);
p.setRenderHint(QPainter::Antialiasing, true);

    QPen pen(Qt::green);
    pen.setWidth(3);
    p.setPen(pen);

    QFont f = p.font();
    f.setPointSize(32);
    f.setBold(true);
    p.setFont(f);

    for (const auto &d : dets) {
        const QString label =
            (d.classId >= 0 && d.classId < m_labels.size()) ? m_labels[d.classId] : QString("class_%1").arg(d.classId);

        p.drawRect(d.box);
        const QString text = QString("%1 %2").arg(label).arg(d.score, 0, 'f', 2);
        QRectF tr(d.box.x(), d.box.y() - 42, 320, 42);
        p.fillRect(tr, QColor(0, 0, 0, 160));
        p.setPen(Qt::white);
        p.drawText(tr.adjusted(6, 0, 0, 0), text);
        p.setPen(pen);
    }

return outImg;
}

Step 5 — QML image provider (to show annotated frames)

Same concept as before (thread-safe setImage() + requestImage()), so as short:

setImage(QImage) stores the latest annotated image (mutex protected)
QML pulls image://annotated/frame?ts=<counter> to refresh

AnnotatedImageProvider.h

Click here for detail...

#pragma once
#include <QQuickImageProvider>
#include <QImage>
#include <QMutex>

class AnnotatedImageProvider : public QQuickImageProvider
{
public:
AnnotatedImageProvider();

// Called from your detector/worker when a new annotated frame is ready
void setImage(const QImage &img);

QImage requestImage(const QString &id, QSize *size, const QSize &requestedSize) override;

private:
QImage m_img;
QMutex m_mutex;
};

AnnotatedImageProvider.cpp

Click here for detail...

#include "AnnotatedImageProvider.h"
#include <QPainter>

AnnotatedImageProvider::AnnotatedImageProvider()
: QQuickImageProvider(QQuickImageProvider::Image)
{
}

void AnnotatedImageProvider::setImage(const QImage &img)
{
QMutexLocker lock(&m_mutex);
m_img = img;
}

QImage AnnotatedImageProvider::requestImage(const QString &id, QSize *size, const QSize &requestedSize)
{
    // IMPORTANT: id may include query string, e.g. "frame?ts=123"
    QString baseId = id;
    const int q = baseId.indexOf('?');

if (q >= 0)
baseId.truncate(q);

    QImage out;
    {
        QMutexLocker lock(&m_mutex);
        out = m_img;
    }

    // If no image yet, return a placeholder instead of failing
    if (out.isNull()) {
        const int w = requestedSize.width() > 0 ? requestedSize.width() : 640;
        const int h = requestedSize.height() > 0 ? requestedSize.height() : 480;
        out = QImage(w, h, QImage::Format_RGBA8888);
        out.fill(QColor(20, 20, 20));

        QPainter p(&out);
        p.setPen(Qt::white);
        p.drawText(out.rect(), Qt::AlignCenter, "Waiting for frames...");
    }

if (size)
*size = out.size();

// If you ever want to support multiple ids, check baseId here:
// if (baseId != "frame") { ... }

return out;
}

Step 6 — main.cpp (Android runtime permission + camera setup)

This is the Android-specific part that matters: request permission, then start camera.

Click here for detail...

#include <QGuiApplication>
#include <QQmlApplicationEngine>
#include <QQmlContext>
#include <QCoreApplication>
#include <QDebug>
#include <QFile>
#include <QDir>
#include <QStandardPaths>
#include <QThread>
#include <QPermission>
#include <QCamera>
#include <QCameraDevice>
#include <QMediaCaptureSession>
#include <QMediaDevices>
#include "FrameGrabber.h"
#include "DetectorWorker.h"
#include "AnnotatedImageProvider.h"

extern int qInitResources_app_assets();
extern int qCleanupResources_app_assets();

static QString extractResourceToFile(const QString &qrcPath, const QString &targetName)
{
    QString dir = QStandardPaths::writableLocation(QStandardPaths::CacheLocation);
    if (dir.isEmpty())
        dir = QStandardPaths::writableLocation(QStandardPaths::AppDataLocation);

if (dir.isEmpty()) {
qCritical() << "No writable location for extracted assets.";

return {};
}

QDir().mkpath(dir);
const QString outPath = QDir(dir).filePath(targetName);

    QFile in(qrcPath);
    if (!in.exists()) {
        qCritical() << "Resource does not exist:" << qrcPath;

return {};
}

if (!in.open(QIODevice::ReadOnly)) {
qCritical() << "Failed to open resource:" << qrcPath << in.errorString();

return {};
}

    QFile out(outPath);
    if (!out.open(QIODevice::WriteOnly | QIODevice::Truncate)) {
        qCritical() << "Failed to open output:" << outPath << out.errorString();

return {};
}

    const QByteArray data = in.readAll();
    out.write(data);
    out.close();
    in.close();

qDebug() << "Extracted" << qrcPath << "->" << outPath << "bytes=" << data.size();

return outPath;
}

int main(int argc, char *argv[])
{
    QGuiApplication app(argc, argv);
    qInitResources_app_assets();
    QQmlApplicationEngine engine;

auto *provider = new AnnotatedImageProvider();
engine.addImageProvider("annotated", provider);

FrameGrabber grabber;
engine.rootContext()->setContextProperty("frameGrabber", &grabber);

    // Worker thread
    QThread workerThread;
    auto *worker = new DetectorWorker();
    worker->moveToThread(&workerThread);

    QObject::connect(&app, &QCoreApplication::aboutToQuit, [&](){
        if (workerThread.isRunning()) {
            QMetaObject::invokeMethod(worker, &QObject::deleteLater, Qt::QueuedConnection);
            workerThread.quit();
            workerThread.wait();
        } else {
            delete worker;
        }

qCleanupResources_app_assets();
});

    // Pipeline
    QObject::connect(&grabber, &FrameGrabber::frameReady,
                     worker, &DetectorWorker::processFrame,
                     Qt::QueuedConnection);

    QObject::connect(worker, &DetectorWorker::annotatedFrameReady,
                     &app, [provider](const QImage &img) {
                         provider->setImage(img);
                     },
                     Qt::QueuedConnection);

    QObject::connect(worker, &DetectorWorker::annotatedFrameReady,
                     &grabber, &FrameGrabber::onAnnotatedFrameReady,
                     Qt::QueuedConnection);

    // Print detections to logcat
    QObject::connect(worker, &DetectorWorker::detectionsTextReady,
                     &app, [](const QStringList &lines){
                         if (lines.isEmpty())
                             return;

                         qDebug().noquote() << "Detections:\n - " + lines.join("\n - ");
                     },
                     Qt::QueuedConnection);

    // Camera
    QMediaCaptureSession session;
    const QCameraDevice camDev = QMediaDevices::defaultVideoInput();

if (camDev.isNull())
qWarning() << "No default video input device found.";

    QCamera camera(camDev);
    session.setCamera(&camera);
    session.setVideoSink(grabber.videoSink());

const QUrl url(QStringLiteral("qrc:/qml/Main.qml"));

    QObject::connect(&engine, &QQmlApplicationEngine::objectCreated,
                     &app, [&](QObject *obj, const QUrl &objUrl){
                         if (!obj && objUrl == url) {
                             qCritical() << "QML root object creation failed. Exiting.";
                             QCoreApplication::exit(-1);

return;
}

if (!obj)
return;

const QString modelPath = extractResourceToFile(":/assets/model.onnx", "model.onnx");

                         if (modelPath.isEmpty()) {
                             qCritical() << "Model extraction failed. Exiting.";
                             QCoreApplication::exit(-1);

return;
}

if (!workerThread.isRunning())
workerThread.start();

QMetaObject::invokeMethod(worker, [worker](){
                             worker->setInputSize(320, 320);
                             worker->setUseLetterbox(true);
                             worker->setThresholds(0.20f, 0.45f); // lower for 320
                             worker->loadLabels(":/assets/labels.txt");
                         }, Qt::QueuedConnection);

                         QMetaObject::invokeMethod(worker, [worker, modelPath](){
                             worker->loadModel(modelPath);
                         }, Qt::QueuedConnection);

                         auto startCamera = [&](){
                             qDebug() << "Starting camera...";
                             camera.start();
                         };

QCameraPermission camPerm;
const auto status = app.checkPermission(camPerm);

                         if (status == Qt::PermissionStatus::Granted) {
                             startCamera();
                         } else {
                             app.requestPermission(camPerm, [&](const QPermission &perm){
                                 if (perm.status() == Qt::PermissionStatus::Granted)
                                     startCamera();
                                 else
                                     qWarning() << "Camera permission denied.";
                             });
                         }
                     },
                     Qt::QueuedConnection);

engine.load(url);
return app.exec();
}

Step 7 — QML UI (Android-friendly)

qml/Main.qml:

Click here for detail...

import QtQuick
import QtQuick.Controls
import QtQuick.Layouts
import QtMultimedia

ApplicationWindow {
    id: mainWnd
    visible: true
    width: 960
    height: 540
    title: "Qt Android AI Object Detection"

    Rectangle {
        anchors.fill: parent
        color: "#111"

        ColumnLayout {
            id: rootLayout
            anchors.fill: parent
            anchors.margins: 12
            spacing: 10

            Text {
                text: "Camera + ONNX Object Detection"
                color: "white"
                font.pixelSize: 18
               Layout.alignment: Qt.AlignHCenter
            }

            Item {
                Layout.fillWidth: true
                Layout.fillHeight: true
                clip: true

                Image {
                    anchors.fill: parent
                    cache: false
                    asynchronous: true
                    fillMode: Image.PreserveAspectFit
                    source: "image://annotated/frame?ts=" + frameGrabber.frameCounter
                }
            }

            Text {
                text: "Detections are printed to logcat (Qt debug output)."
                color: "#bbb"
                font.pixelSize: 12
                Layout.alignment: Qt.AlignHCenter
            }
        }
    }
}

Expected output (Android)

Visual

Camera feed with green bounding boxes
Label + confidence above each detection

Console (logcat / Qt Creator Application Output)

Example lines:

person score=0.82 box=[210,124,180,360]
car score=0.74 box=[520,260,280,170]

Practical Android tips (important)

Performance knobs (the first things to try)

Frame skipping: get the latest frame to process object detection when previous frame has been finished
Use a smaller model: yolov8n > yolov8s for mobile
Lower input size if your model supports it (e.g., 320×320)

Accuracy gotcha: letterbox vs plain resize

Many YOLO pipelines use letterbox resizing (padding). This tutorial uses plain resize in blobFromImage. If your boxes look shifted or scaled wrong, implement letterboxing and reverse mapping.

Debugging on Android

Watch logcat (Qt Creator shows it)
Print model load path and m_ready status
Temporarily lower m_confThres to 0.20

Conclusion:

In this guide, we built a practical, end-to-end real-time object detection demo on Android using Qt—from camera capture to ONNX inference and rendering the final annotated frame. By combining Qt Multimedia (camera + frame delivery), OpenCV DNN (ONNX inference), and a lightweight YOLO model at 320×320, we achieved a setup that’s both portable and fast enough for mobile devices.

A few details made the difference between a “demo that runs” and a pipeline that feels smooth:

Static-shape ONNX models are essential for OpenCV DNN on Android (dynamic/zero shapes often fail).
Letterbox preprocessing + correct reverse mapping keeps bounding boxes aligned with the original camera image.
Running inference in a dedicated QThread and applying backpressure (latest-frame-wins) prevents frame backlog, reduces stutter, and avoids crashes from excessive allocations.
Loading labels.txt and emitting detection strings gives you clean, readable output for logs or UI overlays.

Qt is also building API to easily inference AI Model (Qt AI Inference API) in your QML app

References:

OpenCV for Android: https://github.com/opencv/opencv/releases/download/4.12.0/opencv-4.12.0-android-sdk.zip
ONNX Model for Object detection: https://huggingface.co/giangndm/yolo11-onnx/resolve/main/yolo11n_320.onnx?download=true

Ready to get started?

Contact IVC for a free consultation and discover how we can help your business grow online.

Contact IVC for a Free Consultation

TECH

June 25, 2026

PostgreSQL Indexes: Performance Tuning

PostgreSQL is one of the most popular relational database management systems (RDBMS), known for its reliability, extensibility, and powerful feature set. While query performance is rarely a concern during the early stages of development, it often becomes a challenge as datasets grow from thousands to millions of rows.

As data volume increases, database queries can become a major bottleneck, leading to slower API response times and degraded application performance. One of the most effective ways to address this issue is through indexing.

Indexes are specialized data structures that allow PostgreSQL to locate data more efficiently, significantly reducing query execution time and minimizing expensive full table scans. When used correctly, indexes can dramatically improve the performance of large-scale applications.

In this article, we'll explore what PostgreSQL indexes are, how they work, and how different index types can be used to optimize database performance.

TECH

June 24, 2026

The IT Professional's Guide to Understanding AI Terminology (Part 2)

In Part 1, we explored the "Big Picture"—the Russian Doll hierarchy of AI, how tokens work, and why AI sometimes tells "confident lies." We saw the trap of Overfitting, where an AI memorizes a "red bird" instead of learning what a bird actually is.

But to be a true IT pro in 2026, you need to know more than just the structure. You need to understand the specialized "jobs" AI can do and the different ways we "teach" it. Let’s dive into Part 2 with the essential terms from our list, explained simply but with the detail you need for a technical meeting.

What does Critical Thinking mean to Developers?

Popular Articles

Fintech Development Outsourcing: The Ultimate Guide for 2026

Healthcare IT Outsourcing: A Complete Guide for Medical Software Development

DevOps Best Practices: What to Build and What to Outsource

What Is Adaptive Software Development (ASD)? A Practical Guide for Modern Engineering Teams

White Paper

COMPANY PROFILE

COMPANY PORTFOLIO

ASK ISB Vietnam ABOUT DEVELOPMENT

Step 4 — Object detection worker (OpenCV DNN + ONNX)

CONTACT US