Histogram of Oriented Gaussian (HOG) descriptor and detector

The class implements Histogram of Oriented Gradients object detector [Dalal2005].

Example

The basic usage is the following for computing HOG descriptors:

hog = cv.HOGDescriptor();
descriptors = hog.compute(im);

This computes descriptors in a "dense" setting; Each row is a feature vector computed from a window of size WinSize slided across the input image gradient. Each vector element is a histogram of gradient orientations (quantized into NBins directions). The histogram is collected within a "cell" of pixels of size CellSize, and locally normalized (see HistogramNormType and L2HysThreshold) over larger spatial "block" regions of size BlockSize overlapped according to BlockStride. See cv.HOGDescriptor.getDescriptorSize.

If you need to compute descriptors for a set of certain "sparse" keypoints (for example SIFT or SURF keypoints), use the Locations option of the compute method:

keypoints = cv.FAST(im);
descriptors = hog.compute(im, 'Locations', {keypoints.pt});

The next step in object recognition using HOG descriptors is to feed the descriptors computed from positive and negative images into a linear SVM classifier trained to classify whether a window is an object or not. This trained SVM model is represented by a set of coefficients for each element in the feature vector. This vector is assigned to the cv.HOGDescriptor.SvmDetector property.

Alternatively, you can use the default built-in people detector which is accessible by name as:

% detect and localize upright people in images
hog.SVMDetector = 'DefaultPeopleDetector';
boxes = hog.detectMultiScale(im);

In this case, there is no need to train an SVM model.

Either way, you simply call the detect methods which use the saved coefficients on the input feature-vector to compute a weighted sum. If the sum is greater than a user-specified threshold HitThreshold, then it is classified an object (a pedestrian in the case of the builtin detectors).

The cv.HOGDescriptor.detectMultiScale method detects at multiple scales (see NLevels and Scale) and directly returns the bounding boxes. The cv.HOGDescriptor.detect method returns a list of top-left corner points, where the detected object size is the same as the detector's window size.

References

[Dalal2005]:

Navneet Dalal and Bill Triggs. "Histogram of oriented gradients for human detection". CVPR 2005. http://lear.inrialpes.fr/pubs/2005/DT05/