Hog features-Understanding

I went to the Internet to find information about hog, and found that there was less understanding and it was longer. I wrote an article for your convenience, hope to help the comrades who strive to extract features:

Hog is histogram of Oriented Gradient, which is a feature description sub for target detection. This technique counts the number of partial direction gradients of an image, this method is similar to the edge direction histogram and scale-invariant feature transform. The difference is that hog's calculation improves accuracy based on consistent space density matrix. Navneet Dalal and Bill triggs first proposed hog in cvpr in 05 years for pedestrian detection of static images or videos.

Hog feature principle:

The core idea of hog is that the shape of the detected local object can be described by the distribution of light intensity gradients or edge directions. By cutting the entire image into a small connected area (called cells), each cell generates a gradient histogram or the pixel edge direction in the cell, the combination of these histograms can represent the descriptive sub (the target of the checked targets. To improve accuracy, the local Histogram can be standardized as a measure by calculating the light intensity of a large area (called block) in the image, and then use this value (measure) normalize all cells in this block. this normalization process has completed better illumination/shadow immutability.

Compared with other descriptive descriptors, the descriptive descriptors obtained by hog maintain geometric and optical conversion immutability (unless the object direction changes ). Therefore, hog descriptive descriptions are especially suitable for human detection.

In layman's terms:

The hog feature extraction method is to convert an image:

1. grayscale (the image is regarded as an X, Y, Z (grayscale) three-dimensional image)

2. Divide the data into small cells (2*2)

3. Calculate the gradient (orientation) of each pixel in each cell)

4. Calculate the gradient histogram of each cell (number of different gradients) to form the descriptor of each cell.

Let's talk about the application and difference of hog, sift and PCA-SIFT:

Hog has no rotation and scale immutability, so the calculation amount is small. In sift, each feature needs to be described using a 128-dimension vector, so the calculation amount is relatively large.

So how does one apply hog in the Pedestrian detection platform?

To solve the scale-invariant problem: Scaling images at different scales is equivalent to scaling templates at different scales.

To solve the problem of rotation-invariant: Create templates of different directions (generally 15*7) for matching.

In general, images of different scales are matched with Templates (15*7) in different directions, and each point forms an 8-direction gradient description.

Because of its large amount of computing, sift does not need to be compared with the pedestrian detection, and the PCA-SIFT method filters out the information in many dimensions, only 20 main components are retained, therefore, it is only applicable to physical examination buckets with little behavior changes.

Method	Time	Scale	Rotation	Blur	Illumination	Affine
Sift	Common	Best	Best	Common	Common	Good
PCA-Sift	Good	Good	Good	Best	Good	Best
Surf	Best	Common	Common	Good	Best	Good

Other explanations about sift:

Http://blog.csdn.net/abcjennifer/article/details/7639681

Http://blog.csdn.net/abcjennifer/article/details/7372880

Http://blog.csdn.net/abcjennifer/article/details/7365882

For many other discussions and exchanges on computer vision, please stay tuned to this blog and Sina Weibo sophia_qing.

Hog features-Understanding