Better-than-the-best fusion algorithm with application in human activity recognition
Nayeff Najjar, Shalabh Gupta · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2015
This paper introduces the Better-than-the-Best Fusion (BB-Fus) algorithm. The BB-Fus algorithm is a simple and effective information fusion algorithm that combines the information from different sources (be it sensors, features or classifiers) to improve the Correct Classification Rate (CCR). It can be observed that in most classification problems, different sensors or features might have different classification accuracies in separating different classes. Therefore, this paper constructs an optimal decision tree that isolates one class at a time with the best sensor to separate that particular class. The paper shows that the decision tree improves the overall CCR as compared to the use of any single sensor or feature for any 3-class classification problem. The efficiency of the BB-Fus algorithm is validated on the Opportunity data set to solve the human activity recognition problem where a set of 56 sensors (including a localization system, accelerometers, inertial measurement units and magnetic sensors mounted on various body parts; besides, accelerometers and gyroscopes mounted on different objects) are used. The CCR resulting from the BB-Fus algorithm is 96% while the best sensor achieved 94% CCR.