Human behavior recognition using regression models

This paper proposes a method for human behavior recognition by estimating the human state, i.e., position and orientation, using regression models. In the method, human silhouette in video images is detected by background subtraction technique, and the upper part of human silhouette is used for extracting the image feature. Linear regression technique is introduced to create a model that associates the image feature with human state. Human state estimation from the currently observed image is being performed through this model. Experiments are conducted on indoor space where an Omni Directional Vision (ODV) sensor is installed to the ceiling of crossing hallway. The feasibility and accuracy of our method is discussed through the experimental results.