There have been many implementations in recent years for identifying the human form within a given photo frame, however one particular method stands out called Histograms of Oriented Gradient (HOG) descriptors. Given a training set, the HOG algorithm is capable of eliminating information irrelevant to human detection. The human form can be shown in many different poses, perspective, ambient lighting and backgrounds, however one of the most important characteristics that is common to all are edges and corners. The edge and gradient structure information is defined locally in small regions. The HOG technique consists of counting occurrences of gradient directions in localized cells (or pixel matrices). We then normalize these local histograms. It is suggested that the human form can be represented by using the distribution of local intensity gradients.
For my MATLAB implementation, check out my Github.
The person identification system can be divided into two main functions which are:
- TrainHOG – will compute the necessary weights from training data.
- PredictHOG – will use the previously mentioned weights for predicting whether input images are human or not.