Object Tracking in Distributed Video Networks: a New Approach Using Multi-dimensional Object Signatures

Computers have evolved a long way from huge vacuum tube-based machines to today?s small but immensely powerful devices. They have long been investigated as the foundation for an artificial vision system. The computer vision discipline has seen a rapid development over the years from basic motion detection systems to advanced model-based algorithms. Our work is one such improvement over past algorithms. It is based on the principle of multi-dimensional object signatures. Object signatures are constructed from individual attributes extracted via video processing. The lack of a comprehensive object model limits the application of current algorithms to controlled situations. In real-world conditions, such algorithms perform less efficiently due to inherent assumptions of constancy of attributes. Our approach assumes an environment where the object attributes may vary. The variations in accuracy is handled by using weights for each attribute based on local conditions at a sensor location. The results of our tests establish the validity of our approach as higher match accuracy was obtained with our multi- dimensional approach than with single-attribute based logic.