Activity recognition based on multiple motion trajectories

Abstract

We propose a method for activity recognition based on multiple motion trajectories. Motion trajectories generated from body parts (hand, feet, and joints) are used as features. We not only recognize each activity but also temporally locate the start and end point of its duration. Input sequences are divided into separate temporal segments based on the number of detected trajectories. Segments with same number of trajectories are temporally segmented using the HMM model for each movement (activity). The experimental results show that our approach can successfully locate each activity in continuous video sequences.