Detects the location and dimensions of a person from a video input. The algorithm receives an image and estimates the top-left and bottom-right corners of the human area through convolution operation.

Estimates the position of human joint points. In an image of a person, the position of the joints is calculated as a probability value in the NxM Matrix of the same ratio as the image. The algorithm then estimates a more accurate position through post-processing.

Tracks the same person in a series of images. For people detected in a continuous image frame, features are extracted using Appearance Feature and Motion Feature are used by the matching algorithm determine whether the two are of the same person.

Recognizes a person from video input by learning his/her distinguishing features. An N-dimensional vector is extracted by performing a convolution operation on the human region detected in the image. The vector that is extracted is then used to represent an individual feature of a person, meaning that a specific person can be recognized based on those features.

Recognizes abnormal behaviors related to safety issues. The algorithm receives an RGB image or a sequence of joint position coordinates and performs a convolution operation to extract spatial and temporal feature vectors for the action. The actions are categorized based on an N-dimensional feature vector that includes all the features of a person's posture at each time step (Spatial) and how the posture changes over time (Temporal).