In highly shared urban traffic environments, it is essential to protect Vulnerable Road Users (VRU) to avoid collisions with motorized transport. One approach is to predict the intention or the future trajectories of the VRU from their previous path in order to send warnings in case of danger, or even to brake the cars in case of using driver assistance systems. The main objective of this paper is to investigate the short-term and particularly long-term prediction abilities of the AI-based predictors assisted with environmental maps, if applicable. By comparing and evaluating the performance of Polynomial Regression (PR), Gaussian Process Regression (GPR), Convolutional Neural Network (CNN), and Sequence-tosequence neural networks (SeqToSeq) applied on an open access data set (i.e., Stanford Drone Dataset (SDD)) as well as some simulated data, we can conclude that the SeqToSeq generally performs better than other methods (Average Displacement Error is 25% lower and Final Displacement Error is 20% lower compared to a first order PR). By adding the environmental maps (navigation map and diffusion map), the pedestrian’s turnings are better predicted despite the fact that there is little improvement on other metrics. This can be explained by an insufficient amount of training data involving environmental maps in this research work. Thus it is still promising by adding more training data with environmental maps in the future.