The development of efficient deep learning (DL) methods tailored for embedded hardware in intelligent transportation systems (ITS) is the primary focus of this research. Scalable architectures of composite DL networks are designed to address diverse tasks such as estimating recommended velocities for light electric vehicles (LEVs). Training and fine-tuning will be based on a custom dataset derived from monocular recordings combined with LEV sensor data. From a mathematical perspective, the work further investigates uncertainty quantification using Markov Chain Monte Carlo (MCMC) methods, particularly the Metropolis-adjusted Langevin algorithm, as a principled alternative to variational Bayesian approaches.