如何改进基于HOG特征的SVM行人检测分类器并解决假阳性问题

Let's walk through the key issues in your code and training setup that are likely causing the false positives, and how to fix them step by step:

• Critical Code Bug: Overwriting Video Frames
  Look at this line in your loop:

  img = imread("queue.jpg");
  

  This completely replaces the frame you grabbed from the video with a static image every iteration. No wonder you're getting consistent false positives—you're not actually testing your video! Delete this line immediately to use the real video frames.

• Use the Trained HOG Config File Instead of Manual Setup
  You commented out the hog.load() call and manually loaded the descriptor vector + set winSize. This is risky because the cvHOGClassifier.yaml file includes all the HOG parameters used during training (like blockSize, blockStride, cellSize, nbins) that you haven't set here. If these parameters don't match between training and testing, your feature extraction will be wrong, leading to garbage detections.
  Replace your manual HOG setup with:

  HOGDescriptor hog;
  if (!hog.load("/home/nvidia/Desktop/HOG/genfiles/cvHOGClassifier.yaml")) {
      cerr << "Failed to load HOG classifier!" << endl;
      return -1;
  }
  

  This ensures every HOG parameter matches exactly what was used during training.

• Fix detectMultiScale Parameters
  Your current call is missing critical parameters that control detection strictness:

  hog.detectMultiScale(img, detections, 0, winStride, padding);
  

  • The third argument hitThreshold is set to 0. This means any detection score above 0 is considered a hit, which is way too lenient. When you load the classifier via hog.load(), the correct threshold from training is already embedded—use hog.getDefaultPeopleDetectorThreshold() or just omit this parameter to use the default trained threshold.
  • Add a scaleFactor (e.g., 1.05) to control how much the detection window scales each step, and minNeighbors (e.g., 3) to filter overlapping weak detections.
    Updated call:

  hog.detectMultiScale(img, detections, foundWeights, 0, winStride, padding, 1.05, 3);
  

  Also, using foundWeights lets you filter detections by their confidence score later (e.g., only keep detections with weight > 0.5).

• Check Color Space Consistency
  You convert the frame to grayscale with cvtColor(img, img, CV_BGR2GRAY);, but did you train your classifier on grayscale images? If your training data was color, this mismatch will break feature extraction. Either:

  • Remove the grayscale conversion if you trained on color images, or
  • Ensure your training pipeline also converted images to grayscale.

• Improve Training Data & Pipeline
  Even with fixed code, your model might have false positives due to training issues:

  • Negative Sample Diversity: 3000 negative samples might not be enough, especially if they don't match the scene in your test video. Add more negative samples from the same environment as your test footage, and use hard negative mining (run the initial detector on negative images, collect false positives, and retrain the SVM with these) to drastically reduce false positives.
  • Sample Quality: Double-check your positive samples—make sure they're all correctly cropped to 64x128 and centered on the target object. Misaligned or incorrect positives will confuse the SVM.
  • Class Balance: 15000 positives to 3000 negatives is a 5:1 ratio. While not extreme, you might want to balance this more (e.g., 2:1) or use class weights during SVM training to prevent the model from biasing towards the majority class.

• Simplify Detection Filtering
  Your custom overlapping box filtering can be replaced with OpenCV's built-in groupRectangles function, which is more robust:

  vector<Rect> found_filtered;
  vector<double> weights_filtered;
  groupRectangles(detections, foundWeights, 2, 0.2); // Adjust parameters as needed
  

  This will merge overlapping detections and keep only the most confident ones.

内容的提问来源于stack exchange，提问作者user8219368