Pedestrians trajectory forecasting for Automated Driving and Driving Assistance systems

A key functionality of Automated driving and Driver assistance systems is the ability to detect obstacles in order to constantly adapt the vehicle trajectory to a changing environment. To guarantee a safe path planning has always been the main challenge of all the self-driving car projects, in which artificial intelligence software represent the core part. The most important type of road obstacle are of course pedestrians, which present a particular walking behavior that take into consideration many factors from neighbors' path to environmental influences and road characteristics, but sometimes car detectors fails to interpret their information causing risky situations. The problem of recent automated car systems is that they base their decision just on the present moment failing to consider the next steps of a human trajectory, which is something important that has to be taken into account in order to better decide whether to stop the vehicle or not in front of a potential dangerous scenario. For all these reasons the overall objective of the project is to develop an Artificial Intelligence module able to predict possible human trajectories, along with their level of confidence, in order to represent the last module of a detection, tracking and prediction architecture that can lead to a robust and complete obstacle detection. In this thesis a preliminary study on publicly available dataset has been done to analyze pedestrians' trajectory behavior, how to deal with them and the amount of data necessary to train a good prediction model. Whole data have been pre-processed to extract pedestrians' positions and normalize them with respect to the specific dataset characteristics, as they presented a monocular or multi-camera setting mounted on a moving vehicle. Guidelines have been provided to future works in which the acquisition of a custom dataset will be needed. As first step, a multi-camera calibration software has been developed to fulfill the need of a preliminary calibration process related to the automotive domain and adapted to the specific vehicle sensor setting. This is done with the aim of projecting all the detected pedestrians' position from pixel coordinates to world ones, so that trajectories can be considered with respect to a fixed reference system not relative to the ego-vehicle motion. After data pre-processing a prediction model based on the architecture of a Social LSTM Recurrent Neural Network has been developed and tested both on monocular and multi-camera dataset. The model, which achieves the state of the art for trajectory forecasting field, has been adapted to the automotive context representing an evolution compared to previous works in which pedestrians' paths were only considered in a fixed camera setting for surveillance goal.