Learning objectives
The course aims to provide knowledge and skills on the development of Artificial Intelligence (AI) systems for the Automotive. It provides an initial overview, also of an economic impact type, on AI systems that are used in automotive, both for ADAS, for autonomous driving, for human-vehicle interaction and for industrial production. The course offers a detailed overview of the state of the art of AI technologies and above all of deep learning, oriented to visual understanding, action and interaction in the Automotive field, and provides advanced skills on the design of Deep Learning algorithms, also in reference to their optimization and deployment on architectures with reduced computational capacity. Computer vision techniques for the geometric understanding of spaces, machine learning techniques and practices for the design and implementation of completely connected, convolutive neural architectures, and for the understanding and generation of time sequences are presented. An important part of the course is dedicated to the study of architectures and solutions for the visual understanding of data from cameras positioned inside and outside the passenger compartment - with particular reference to motion estimation, object identification, segmentation semantics the prediction of trajectories, the understanding of the surrounding environment and the monitoring of the driver and his attention state.
Prerequisites
Basic knowledge of statistics, linear algebra, geometry introduced in the three-year and master's courses of the Engineering degrees. Knowledge of computer architecture and basic knowledge of the Python language (which, however, is taken up again in the course).
Course unit content
# History and impact of AI in Automotive
Introduction to the history of AI and to its economical impact in the Automotive domain. AI and Computer Vision for Automotive.
# Neural Architectures fundamentals
Gradient-based optimization, fully connected and convolutional architectures. Analysis of computational costs and computational graph management. Design practices, network surgery.
# On-board sensors and cameras, depth sensors
Analysis of commercial on-board sensors, depth, thermal and RGB cameras.
# Sequence modelling and prediction
Recurrent, Convolutional and Fully-Attentive architectures for sequence understanding and generation. Computational graph analysis. Self-attention and cross-attention. Applications to trajectory prediction and language-based interfaces.
Full programme
- - -
Bibliography
- Slides and scientific papers from international conferences and journals (CVPR; ECCV; ICCV; T-PAMi. IVPR, IEEE ITS; IEEE IV)
Teaching methods
Most of the lessons are frontal and use scientific slides and papers as didactic support; about 30% of the lessons are laboratory, with hands-on experiences on tensor computation libraries and the design and training of neural networks and AI algorithms for the Automotive. Upon completion, meetings and discussions with companies are organized.
Assessment methods and criteria
The exam takes place at the end of the course according to the official exam session calendar. The test consists of an oral test according to the modalities of an interview between teachers and student during which three questions will tend to be proposed, with increasing degree of depth (and difficulty) and on different areas of the program. During the interview it is not possible to consult the didactic material and / or other materials.
Oral exams take place over the course of the educational calendar of the training offer and for each exam the student who intends to take it must register using the ESSE3 Platform. The results of the oral exam will be published on ESSE3 and the student will have the right to accept or refuse the grade.
Other information
- - -
