Algorithms for predicting pedestrians' trajectory in Automated Driving and Driving Assistance systems

Driving Assistance systems and Automated Driving systems are going to become crucial technologies in the next years to increase people’s safety. This thesis is part of a project born from the collaboration between Centro Ricerche Fiat (also known as CRF and part of FCA group) and Politecnico di Torino to find a reliable tool to analyze pedestrians’ movement and to predict their future positions. This is a very difficult task, because pedestrians’ movements are guided from a multitude of elements (the environment, nearby pedestrians, elements on the sidewalks and so on...) and are not easily predictable, even from another human. This project aims to detect pedestrians on a total field of view of 360° around the car using a multi-camera setup, to track them and, lastly, be able to predict their trajectories after some seconds of observation. This work started analyzing publicly available datasets, in order to understand which types of data are necessary to accomplish that task, and the amount of data that are required and how to correctly use them. Then, it was decided to use a Social-LSTM network to predict the trajectories of the pedestrians. With such an architecture it is possible to consider both the past positions of the pedestrians and pedestrians in their neighbourhood. This architecture has been studied in order to improve it and adapt it to our task. To do that, it was necessary to modify the network to work with pedestrians’ trajectories in a world coordinate reference system, to find the hyper-parameters of the network that give the best possible solution with the input data and, then, to add the position of the car as an input of the LSTM, in order to let it know where the car, and so the road, is. Moreover, to adapt this architecture on an FCA vehicle, all the possible modifications from the original architecture, such as the necessity to label the images (that in the publicly available datasets are already labelled), the camera calibration, the translation of detected pedestrians into world coordinates and the pre-process of data to use the LSTM network, have been considered.