来自相机的 Qt 视频帧损坏

Question

EDIT: The first answer solved my problem. Apart from that I had to set the ASI_BANDWIDTH_OVERLOAD value to 0.

I am programming a Linux application in C++/Qt 5.7 to track stars in my telescope. I use a camera (ZWO ASI 120MM with according SDK v0.3) and grab its frames in a while loop in a separate thread. These are then emitted to a QOpenGlWidget to be displayed. I have following problem: When the mouse is inside the QOpenGlWidget area, the displayed frames get corrupted. Especially when the mouse is moved. The problem is worst when I use an exposure time of 50ms and disappears for lower exposure times. When I feed the pipeline with alternating images from disk, the problem disappears. I assume that this is some sort of thread-synchronization problem between the camera thread and the main thread, but I couldnt solve it. The same problem appears in the openastro software. Here are parts of the code:

MainWindow:

MainWindow::MainWindow(QWidget *parent) : QMainWindow(parent){

mutex = new QMutex;
camThread = new QThread(this);
camera = new Camera(nullptr, mutex);
display = new GLViewer(this, mutex);

setCentralWidget(display);

cameraHandle = camera->getHandle();

connect(camThread, SIGNAL(started()), camera, SLOT(connect()));
connect(camera, SIGNAL(exposureCompleted(const QImage)), display, SLOT(showImage(const QImage)), Qt::BlockingQueuedConnection );

camera->moveToThread(camThread);
camThread->start();
}

The routine that grabs the frames:

void Camera::captureFrame(){
    while( cameraIsReady && capturing ){
        mutex->lock();
        error = ASIGetVideoData(camID, buffer, bufferSize, int(exposure*2*1e-3)+500);
        if(error == ASI_SUCCESS){
            frame = QImage(buffer,width,height,QImage::Format_Indexed8).convertToFormat(QImage::Format_RGB32); //Indexed8 is for 8bit 
            mutex->unlock();
            emit exposureCompleted(frame);
        }
        else {
            cameraStream << "timeout" << endl;
            mutex->unlock();
        }
    }
}

The slot that receives the image:

bool GLViewer::showImage(const QImage image)
{
    mutex->lock();
    mOrigImage = image;
    mRenderQtImg = mOrigImage;

    recalculatePosition();

    updateScene();

    mutex->unlock();
    return true;
}

And the GL function that sets the image:

void GLViewer::renderImage()
{
    makeCurrent();
    glClear(GL_COLOR_BUFFER_BIT);

    if (!mRenderQtImg.isNull())
    {
        glLoadIdentity();
        glPushMatrix();
        {
            if (mResizedImg.width() <= 0)
            {
                if (mRenderWidth == mRenderQtImg.width() && mRenderHeight == mRenderQtImg.height())
                    mResizedImg = mRenderQtImg;
                else
                    mResizedImg = mRenderQtImg.scaled(QSize(mRenderWidth, mRenderHeight),
                                                      Qt::IgnoreAspectRatio,
                                                      Qt::SmoothTransformation);
            }
            glRasterPos2i(mRenderPosX, mRenderPosY);
            glPixelZoom(1, -1);
            glDrawPixels(mResizedImg.width(), mResizedImg.height(), GL_RGBA, GL_UNSIGNED_BYTE, mResizedImg.bits());
        }
        glPopMatrix();
        glFlush();
    }
}

I stole this code from here: https://github.com/Myzhar/QtOpenCVViewerGl

And lastly, here is how my problem looks:

This looks awful.

Answer 1

The image producer should produce new images and emit them through a signal. Since QImage is implicitly shared, it will automatically recycle frames to avoid new allocations. Only when the producer thread out-runs the display thread will image copies be made.

Instead of using an explicit loop in the Camera object, you can run the capture using a zero-duration timer, and having the event loop invoke it. That way the camera object can process events, eg timers, cross-thread slot invocations, etc.

There's no need for explicit mutexes, nor for a blocking connection. Qt's event loop provides cross-thread synchronization. Finally, the QtOpenCVViewerGl project performs image scaling on the CPU and is really an example of how not to do it. You can get image scaling for free by drawing the image on a quad, even though that's also an outdated technique from the fixed pipeline days - but it works just fine.

The ASICamera class would look roughly as follows:

// https://github.com/KubaO/stackoverflown/tree/master/questions/asi-astro-cam-39968889
#include <QtOpenGL>
#include <QOpenGLFunctions_2_0>
#include "ASICamera2.h"

class ASICamera : public QObject {
   Q_OBJECT
   ASI_ERROR_CODE m_error;
   ASI_CAMERA_INFO m_info;
   QImage m_frame{640, 480, QImage::Format_RGB888};
   QTimer m_timer{this};
   int m_exposure_ms = 0;
   inline int id() const { return m_info.CameraID; }
   void capture() {
      m_error = ASIGetVideoData(id(), m_frame.bits(), m_frame.byteCount(),
                                 m_exposure_ms*2 + 500);
      if (m_error == ASI_SUCCESS)
         emit newFrame(m_frame);
      else
         qDebug() << "capture error" << m_error;
   }
public:
   explicit ASICamera(QObject * parent = nullptr) : QObject{parent} {
      connect(&m_timer, &QTimer::timeout, this, &ASICamera::capture);
   }
   ASI_ERROR_CODE error() const { return m_error; }
   bool open(int index) {
      m_error = ASIGetCameraProperty(&m_info, index);
      if (m_error != ASI_SUCCESS)
         return false;
      m_error = ASIOpenCamera(id());
      if (m_error != ASI_SUCCESS)
         return false;
      m_error = ASIInitCamera(id());
      if (m_error != ASI_SUCCESS)
         return false;
      m_error = ASISetROIFormat(id(), m_frame.width(), m_frame.height(), 1, ASI_IMG_RGB24);
      if (m_error != ASI_SUCCESS)
         return false;
      return true;
   }
   bool close() {
      m_error = ASICloseCamera(id());
      return m_error == ASI_SUCCESS;
   }
   Q_SIGNAL void newFrame(const QImage &);
   QImage frame() const { return m_frame; }
   Q_SLOT bool start() {
      m_error = ASIStartVideoCapture(id());
      if (m_error == ASI_SUCCESS)
         m_timer.start(0);
      return m_error == ASI_SUCCESS;
   }
   Q_SLOT bool stop() {
      m_error = ASIStopVideoCapture(id());
      return m_error == ASI_SUCCESS;
      m_timer.stop();
   }
   ~ASICamera() {
      stop();
      close();
   }
};

Since I'm using a dummy ASI API implementation, the above is sufficient. Code for a real ASI camera would need to set appropriate controls, such as exposure.

The OpenGL viewer is also fairly simple:

class GLViewer : public QOpenGLWidget, protected QOpenGLFunctions_2_0 {
   Q_OBJECT
   QImage m_image;
   void ck() {
      for(GLenum err; (err = glGetError()) != GL_NO_ERROR;) qDebug() << "gl error" << err;
   }
   void initializeGL() override {
      initializeOpenGLFunctions();
      glClearColor(0.2f, 0.2f, 0.25f, 1.f);
   }
   void resizeGL(int width, int height) override {
      glViewport(0, 0, width, height);
      glMatrixMode(GL_PROJECTION);
      glLoadIdentity();
      glOrtho(0, width, height, 0, 0, 1);
      glMatrixMode(GL_MODELVIEW);
      update();
   }
   // From http://stackoverflow.com/a/8774580/1329652
   void paintGL() override {
      auto scaled = m_image.size().scaled(this->size(), Qt::KeepAspectRatio);
      GLuint texID;
      glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT);
      glGenTextures(1, &texID);
      glEnable(GL_TEXTURE_RECTANGLE);
      glBindTexture(GL_TEXTURE_RECTANGLE, texID);
      glTexImage2D(GL_TEXTURE_RECTANGLE, 0, GL_RGB, m_image.width(), m_image.height(), 0,
                   GL_RGB, GL_UNSIGNED_BYTE, m_image.constBits());

      glBegin(GL_QUADS);
      glTexCoord2f(0, 0);
      glVertex2f(0, 0);
      glTexCoord2f(m_image.width(), 0);
      glVertex2f(scaled.width(), 0);
      glTexCoord2f(m_image.width(), m_image.height());
      glVertex2f(scaled.width(), scaled.height());
      glTexCoord2f(0, m_image.height());
      glVertex2f(0, scaled.height());
      glEnd();
      glDisable(GL_TEXTURE_RECTANGLE);
      glDeleteTextures(1, &texID);
      ck();
   }
public:
   GLViewer(QWidget * parent = nullptr) : QOpenGLWidget{parent} {}
   void setImage(const QImage & image) {
      Q_ASSERT(image.format() == QImage::Format_RGB888);
      m_image = image;
      update();
   }
};

Finally, we hook the camera and the viewer together. Since the camera initialization may take some time, we perform it in the camera's thread.

The UI should emit signals that control the camera, eg to open it, start/stop acquisition, etc., and have slots that provide feedback from the camera (eg state changes). A free-standing function would take the two objects and hook them together, using functors as appropriate to adapt the UI to a particular camera. If adapter code would be extensive, you'd use a helper QObject for that, but usually a function should suffice (as it does below).

class Thread : public QThread { public: ~Thread() { quit(); wait(); } };

// See http://stackoverflow.com/q/21646467/1329652
template <typename F>
static void postToThread(F && fun, QObject * obj = qApp) {
   QObject src;
   QObject::connect(&src, &QObject::destroyed, obj, std::forward<F>(fun), 
                    Qt::QueuedConnection);
}

int main(int argc, char ** argv) {
   QApplication app{argc, argv};
   GLViewer viewer;
   viewer.setMinimumSize(200, 200);
   ASICamera camera;
   Thread thread;
   QObject::connect(&camera, &ASICamera::newFrame, &viewer, &GLViewer::setImage);
   QObject::connect(&thread, &QThread::destroyed, [&]{ camera.moveToThread(app.thread()); });
   camera.moveToThread(&thread);
   thread.start();
   postToThread([&]{
      camera.open(0);
      camera.start();
   }, &camera);
   viewer.show();
   return app.exec();
}
#include "main.moc"

The GitHub project includes a very basic ASI camera API test harness and is complete: you can run it and see the test video rendered in real time.