Human activity recognition from movie sequences
Nhat-Tan Nguyen
Denis Laurendeau (Supervisor)
Problem: Human action recognition is an important topic in computer vision. The task of recognizing human actions poses several challenges. Human activity is extremely diverse. Many applications such as surveillance for security and for human computer interaction are in high demand and would benefit from a system that could achieve this task automatically. Automatic human activity recognition by computer involves, firstly, detecting and tracking moving objects from image sequences. The activities are then recognized from the characteristics of these tracked mobile objects. One of the major tasks in this process is concerned with how to link the gap between pixel level data and a high level abstract activity such as a verbal description.
Motivation: Our work, in this particular context, is to identify human action in a movie like: stand, walk, run, bend down, and jump. The challenges are that, in a movie, the camera is not always static and the scene is changing frequently. One advantage of working with a movie sequence is that we do not need to run our algorithm in real time.
Approach: Our method is composed of two steps. First, we perform background subtraction and camera motion estimation based on features of KLT 4 and Fuzzy C-mean clustering 4. Then combined with other information from skin detection, face detection and hierarchical clustering, we build the templates of activity descriptions. In out approach, an activity is represented by a set of pose and velocity vectors for main body parts. Then we use a voting approach to recognize the activity from the templates of activity.
Calendar: January 2006 -
Last modification: Sep 28 2007 2:44PM by ntnguyen


©2002-. Computer Vision and Systems Laboratory. All rights reserved