Abstract—During the transition from manual driving to (partially) automated driving, conflicts between human and automation should be minimized. An important factor to reduce the human-machine interaction risk is by increasing the trust drivers have in the system. A system including human-like behaviour tends to be trusted more. Accurate models of human driving behaviour are important in safety studies and the development of human-compatible automated driving. While global models exist showing the relationship between road geometry and speed control, research on models that use observable visual inputs for drivers is scarce. Previous research used a heuristic approach to develop such a model. This work extends on that basis, using machine learning to develop a model for speed control based on angles and time margins that can be extracted from the driver’s visual field. The model is developed using two optimization methods with a single hidden layer feed-forward neural network structure and training data obtained from 14 participants in a fixed-base driving simulator. Validation runs have been done by the same participants in order to validate the individual and general (one-size-fits-all) longitudinal driving models. The validation shows that a general model is able to accurately capture human longitudinal acceleration behaviour on single-lane curved high-speed roads (100km/h).