The solution consists of three subcomponents.
The first one is the human action recognition tool, that is a deep learning-based model, which takes as input a sequence of images and outputs the classes (run, walk, fight, panic) of the human actions that are detected.
The second tool is the person re-identification tool that is able to identify the same person of interest in different locations and times from videos of multiple cameras. The deep
learning network detects the people that are most similar with the person of interest based on their matching distance.
The last module is the people counting tool, that count the number of people in a crowded scene using a deep learning model that receives an image and provides a number and a heat map that describes the density of the people presence in the scene.