Automated Detection of Multiple Actions Based on Hog Features
The main objective of our work is to present a method to remark multiple human actions (predefined set) in an
input video automatically. Having considered the recent experiences with the field we use Histogram of Gradient HoG
algorithm for feature extraction and a general classifier for classification and regression. Finally, we present a fully automatic
set up with action learning and classification obtained for a training data set. We have also implemented it for single action.
As a part of preprocessing, first the input video is divided into a number of frames and each frame is stored in a file and
displayed on the screen during the whole run. Next, the above frames are segmented into two parts from the middle and are
stored in different files. This process makes identification of features much easier. As a part of feature extraction each frame
is first converted into grace scale image and a fixed size, then HoG feature extraction algorithm is applied and each processed
frame is compared with the set of relevant images to complete the process. The classifier is used for classification which uses
feature vector obtained for each frame as an input. It should also be noted that the complete process of classification takes
place twice for a single video due to segmentation of each frame.
Index Terms— Histogram of Gradient HoG; preprocessing; feature extraction, classifier.