Probabilistic estimation of vehicle speed for autonomous vehicles using deep kernel learning

Perception systems of autonomous vehicles (AVs) play a crucial role in achieving different levels of autonomy and interpreting information from complex traffic environments. The inherent uncertainties are a persistent factor. While some environmental features can be directly measured by certain sensors, measuring accurately the velocity of moving vehicles presents a substantial challenge. To this end, this paper demonstrates the utilisation of a powerful non-parametric method, Gaussian process regression, in combination with a method that leverages deep neural networks, known as Deep Kernel Learning (DKL), to estimate the vehicle speed using other existing data in the simulation that are considered feasible in real-world scenarios. The methodology is experimentally evaluated in a single ring-shaped traffic simulation where an autonomous vehicle (AV) drives together with human-driven vehicles (HDVs). The study reveals that the approach significantly enhances the accuracy and confidence of speed estimation with 64.58% and 50% improvements in the root mean square error (RMSE) for 525 and 3000 training data, respectively. It outperforms the conventional Gaussian processes, which suffer from a large dataset.