Visual odometry and pedestrian trajectory forecasting: proof of concept for autonomous driving systems

Smart cities and smart mobility are part of our present and future. New technologies are being developed to hinder the increasing traffic, pollution and energy consumption; one of these are fully autonomous driving vehicles, that are expected to be available globally by 2040. Perception and decision making tasks are only some of the numerous challenge that this technology have to improve in order to be efficient and safe for everyone. This thesis aim is to develop a proof of concept able to detect, track and forecast pedestrian trajectories, to improve safety in assisted or automated driving scenario, and is part of a project being conducted by Centro Ricerche Fiat (CRF-FCA) and Politecnico di Torino. One of the required input to be able to forecast pedestrian trajectory is the ego-vehicle position, therefore a particular focus has been on the development of a localization system, using a multi-camera visual odometry algorithm (VO), in combination with classical inertial odometry. To deal with multi-camera VO, first a single-camera structure was implemented; the results have then been extended to correctly work with the full camera setup provided by the prototype. The proof of concept is tested and analyzed using a convenient automotive public dataset, nuScenes, that has a similar camera setup with respect to the CRF-FCA prototype. A final report have been then carried out in order to evaluate how the localization system affect the pedestrian trajectory forecasting error with respect to the ground truth.