PT-HMC: Optimization-based Pre-Training with Hamiltonian Monte-Carlo Sampling for Driver Intention Recognition

Abstract

Driver intention recognition (DIR) methods mostly rely on deep neural networks (DNNs). To use DNNs in a safety-critical real-world environment it is essential to quantify how confident the model is about the produced predictions. Therefore, this study evaluates the performance and calibration of a temporal convolutional network (TCN) for multiple probabilistic deep learning (PDL) methods (Bayes-by-Backprop, Monte-Carlo dropout, Deep ensembles, Stochastic Weight averaging - Gaussian, Multi SWA-G, cyclic Stochastic Gradient Hamiltonian Monte Carlo). Notably, we formalize an approach that combines optimization-based pre-training with Hamiltonian Monte-Carlo (PT-HMC) sampling, aiming to leverage the strengths of both techniques. Our analysis, conducted on two pre-processed open-source DIR datasets, reveals that PT-HMC not only matches but occasionally surpasses the performance of existing PDL methods. One of the remaining challenges that prohibits the integration of a PDL-based DIR system into an actual car is the computational requirements to perform inference. Therefore, future work could focus on optimizing PDL methods to be more computationally efficient without sacrificing performance or the ability to estimate uncertainties.