BOsampler: a Plug-and-play Module for Stochastic Human Trajectory Sampling Promoting
The unpredictable nature of human motion necessitates that trajectory prediction systems employ a probabilistic model to capture the multi-modal aspects and deduce a limited set of potential future paths. However, most current techniques depend on Monte Carlo random sampling for inference, which is inadequate for representing realistic trajectories with a finite number of samples, due to the long-tail effect of the predicted distribution.
To enhance the sampling process for stochastic prediction, we introduce a plug-and-play module , which utilizes Bayesian optimization to adaptively discover plausible trajectories in an unsupervised manner. This approach functions as a sequential design strategy, with each new prediction relying on previously obtained samples. We represent the trajectory sampling as a Gaussian process and develop an acquisition function to assess the potential value of sampling. This function incorporates the original distribution as a prior and pro- motes the exploration of paths within the long-tail region. This sampling technique can be incorporated into existing stochastic prediction models without the need for retraining. Experiments conducted on a variety of baseline methods illustrate the efficacy of our proposed approach.
Z. Chen, "BOsampler: a Plug-and-play Module for Stochastic Human Trajectory Sampling Promoting", M.S. Thesis, Machine Learning, MBZUAI, Abu Dhabi, UAE, 2023.