Some time ago began to understand hog and SVM pedestrian recognition, saw a lot including Dalal predecessors of the article and experience sharing, hog theory has some preliminary understanding.
The full name of HoG is histogram of oriented Gradient, which is the gradient direction histogram. is to calculate the gradient direction of each pixel, which is counted as a histogram to represent the target.
The following is a brief introduction to the use of Hog + SVM to achieve target detection of the simple steps
STEP1: Obtains the positive sample set and obtains the hog characteristic descriptor with the hog computation characteristic. For example, pedestrian detection can be done by Irina and other pedestrian sample sets to extract the traveler's descriptors.
STEP2: obtains the negative sample set and obtains the hog characteristic descriptor with the hog computation characteristic. Negative sample images can be randomly cropped with images without detection targets. Typically, negative samples are much larger than the number of positive samples.
STEP3: The model is obtained by using SVM to train positive and negative samples.
STEP4: The model is used to detect the negative sample difficult cases. The negative samples in the training set are measured in multi-scale, and if the classifier detects non-target by mistake, the image is added to the negative sample. (hard-negative mining)
STEP5: re-training model with difficult examples.
STEP6: The final classifier model test set is applied, the different scale of each image is scanned by sliding, and the descriptor is extracted and classified by classifier. If the target is detected, use the bounding box. After the image scan is complete, the non-maximum suppression is applied to eliminate overlapping unwanted targets.
Detection and analysis of HoG SVM targets