Head and body pose estimation for pedestrian trajectory forecasting in automated driving systems

In the last years the interest on self-driving cars grew exponentially, in parallel with the great improvement of machine-learning and deep learning algorithms that, in combination with modern sensors such as GPS, radar and lidar, allow to design on-vehicle systems able to support or even to replace totally humans while driving. The objective of this thesis was to evaluate how the head and body pose of a person can be used in the context of Automated Driving to enhance the prediction of pedestrian paths in urban scenarios. This work is a part of a bigger project carried out by CRF (Centro Ricerche Fiat) and Politecnico di Torino, whose final aim is to realize a complete system for an automated vehicle that in the end will perform all what is necessary for a safe and effective driving: detection, tracking and path forecasting for all road users (pedestrians, cyclists, other cars), urban segmentation, signals recognition. After analysing one of the most famous dataset for AD, several algorithms were tested to obtain an estimation on pedestrians head and body poses as reliable as possible; then, starting from a LSTM network that was already able to predict pedestrians’ paths by using their past positions and the information about their neighbors, this model was modified in order to include the estimated poses, appropriately represented and encoded, in the forecasting module and to evaluate how they impact on the final prediction.