Журнал дистанционного зондирования и ГИС

Абстрактный

Real time visual subject tracking and classification by combining motion signal analysis and tri-dimensional shape feature classifiers with group induction boosting algorithms

Lucas Agudiez Roitman

This paper provides a novel and unprecedented approach for integrating motion features in the detection and classification of moving subjects in a static environment. More specifically, author measure the impact of the use of trajectory history, rotation history, blob orientation, motion frequency in the three axes, motion acceleration, segmentation errors and flickering scores and how they can influence classification of moving people, pets and other objects. They apply our method to data captured by a combined color and depth camera sensor. They find that, while some motion descriptors slightly improve accuracy, the use of them in conjunction outperforms previous approaches in the classification and tracking of real world moving subjects in real-time.