Multi-camera pedestrian detection and tracking for automated driving systems

Pedestrian detection and tracking is an essential task for automated driving systems. In recent years, several groups around the world have proposed a wide range of solutions with a variety of sensors and methods. The following thesis is developed in collaboration with the FCA, it describes the selection and adaption of the data set based on the FCA's and Politecnico di Torino research interests. Moreover, the thesis proposes an overall pedestrian detection and tracking system using only a Multi-camera set up with a total field of view that covers the 360° around the car. The thesis focuses on the development process for the detection and tracking system. It includes the evaluation performed to select a neural network that will be fed with the video sequences coming from the 6 cameras located in the car. The method used to project the 2D detections into points in the world coordinates according to the ego vehicle position and implementation of a tracking algorithm to get the notion of the path of each pedestrian. Finally, the overall system is evaluated in the nuscenes tracking challenge by comparing its tracking results with different system baselines.