Long-Term Trajectory Prediction of the Human Hand and Duration Estimation of the Human Action

In the frameworkof human-robot collaborative assembly, it is important to predict the long-term human hand trajectory for collision avoidance and to estimate the durations of the human actions for collaborative task planning. Many existing works predict long-term human trajectory by a preset time horizon, while in this letter, our prediction horizon depends on how far in the future we could predict the human's actions. To be more specific, we predict the human trajectory and estimate the durations for the human's current action and future actions. To address this problem, we propose a recognition-then-prediction framework. First, we present a hierarchical recognition algorithm to infer the human intentions for the current action and the future actions. Next, we propose to use the sigma-lognormal function to model and predict the human movement, and from this model we estimate the action durations. To accommodate different human behaviors, we also propose an online algorithm to adapt the movement model by using the observed trajectory, the human intentions and the scene layout. The effectiveness of the proposed framework is supported by experimental validations on the human trajectory data for conducting a computer assembly task.