Adaptive human motion analysis and prediction

Abstract

Human motion analysis and prediction is an active research area where predicting human motion is often performed for a single time step based on historical motion. In recent years, longer term human motion prediction has been attempted over a number of future time steps. Most current methods learn motion patterns (MPs) from observed trajectories and then use them for prediction. However, these learned MPs may not be indicative due to inadequate observation, which naturally affects the reliability of motion prediction. In this paper, we present an adaptive human motion analysis and prediction method. It adaptively predicts motion based on the classified MPs in terms of their credibility, which refers to how indicative the learned MPs are for the specific environment. The main contributions of the proposed method are as follows: First, it provides a comprehensive description of MPs including not only the learned MPs but also their evaluated credibility. Second, it predicts long-term future motion with reasonable accuracy. A number of experiments have been conducted in simulated scenes and real-world scenes and the prediction results have been quantitatively evaluated. The results show that the proposed method is effective and superior in its performance when compared with a recursively applied Auto-Regressive (AR) model, which is called the Recursive Short-term Predictor (RSP) for long-term prediction. The proposed method has 17.73% of improvement over the RSP in prediction accuracy in the experiment with the best performance. On average, the proposed method has 5% improvement over the RSP in prediction accuracy over 10 experiments. © 2011 Elsevier Ltd. All rights reserved.