2D Silhouette Properties
A method for activity recognition in a supportive home environment with multiple overlapping cameras is presented. The persons position and a coarse representation of it’s shape were tracked using the bounding box of the persons silhouette. Other authors have tracked people with ellipses and over head mounted cameras to achieve activity summarization. The position and view direction of the individual cameras were chosen in a way that image y-coordinates roughly relate to z-coordinates in real world. Thus this representation offers a reasonable accurate description.
Extracting the persons silhouettes is the first step in the activity recognition system. An enhanced color mean and variance motion detector with additional shadow and highlight detection provides a fast and reliable extraction of the silhouettes even with challenging lightning conditions.
Objects are tracked based on their bounding box in each camera independently. The bounding box centroid, width to height ratio and the current speed are tracked. Based on these features a rule based fall detection for each view was applied. Inactivity is detected if the centroid position does not change significant in the last second. Since the image y-coordinates roughly match real world z-coordinates, the changes of the bounding box in a fall are a decrease in y-direction and probably an increase in x-direction. This is similar to , however there ellipses were tracked and the image evidence was provided by a single over head mounted camera. Based on these assumptions, each camera recognizes fall individually. The fall detection module fuses the individual camera results with a simple majority vote. Figure 1 shows the graphical output of the prototype for the three object states normal, intermediate and falling.
