Fast human activity recognition based on structure and motion

We present a method for the recognition of human activities. The proposed approach is based on the construction of a set of templates for each activity as well as on the measurement of the motion in each activity. Templates are designed so that they capture the structural and motion information that is most discriminative among activities. The direct motion measurements capture the amount of translational motion in each activity. The two features are fused at the recognition stage. Recognition is achieved in two steps by calculating the similarity between the templates and motion features of the test and reference activities. The proposed methodology is experimentally assessed and is shown to yield excellent performance.

► A method is presented for the recognition of human activities using structure and motion.
► Templates are designed to capture the discriminative structural and motion information.
► Direct motion measurements capture the amount of translational motion.
► The extracted features are fused and recognition is achieved in two stages.
► The proposed method yields excellent results when applied on the INRIA database.