Pedestrian trajectory forecasting in automated driving systems

In recent years machine learning has become even more diffused in very different applications; especially deep learning techniques, in combination with modern sensors such as lidar, GPS, radar and cameras, allow to develop on-vehicle systems able to help humans in preventing accidents while driving. The aim of this thesis is to investigate the use of the scene information in an Automated Driving scenario for trajectory forecasting. This work is part of a project carried out by CRF (Centro Ricerche Fiat) and Politecnico di Torino with the objective of developing a module for self driving cars, able to detect and track all pedestrians in the surrounding area of the vehicle, to perform trajectory forecasting, and to elaborate a future path plan based on the predictions, in order to safeguard all the road users. Working on one of the most famous dataset for AD (nuScenes), which provides also high definition semantic maps, this map information has been elaborated and encoded as input of an existing lstm network (developed by my colleagues working on this project) able to consider also the body and head pose and the social interactions. Some other modification has been introduced, with the final goal of improving the accuracy of trajectory forecasting.