Vision-based toddler tracking at home

Abstract

This paper presents a vision-based toddler tracking system for detecting risk factors of a toddler's fall within the home environment. The risk factors have environmental and behavioral aspects and the research in this paper focuses on the behavioral aspects. Apart from common image processing tasks such as background subtraction, the vision-based toddler tracking involves human classification, acquisition of motion and position information, and handling of regional merges and splits. The human classification is based on dynamic motion vectors of the human body. The center of mass of each contour is detected and connected with the closest center of mass in the next frame to obtain position, speed, and directional information. This tracking system is further enhanced by dealing with regional merges and splits due to multiple object occlusions. In order to identify the merges and splits, two directional detections of closest region centers are conducted between every two successive frames. Merges and splits of a single object due to errors in the background subtraction are also handled. The tracking algorithms have been developed, implemented and tested.