Guglielmo Camporese

ritratto Guglielmo Camporese

Computer Science and Innovation for Societal Challenges, XXXV series
Grant sponsor

Dip. di Matematica, UNIPD

Lamberto Ballan

Gianluca Campana

Project description
Video and image understanding is a central problem in machine learning since it involves spatial, temporal and relational data. A combination of deep learning architectures such as a mixture of convolutional, recurrent and graph neural networks seems crucial on vision-related tasks. These techniques are directly applied to the “PRediction of activities and Events in an Urban Environment” (PREVUE) project in collaboration with the Comune of Modena in order to build a smart city. For instance, my project aims at building systems that learn the intra-interactions of the urban agents (cars, pedestrians, bicycles...) and the external interaction with the environment (the city). Challenging tasks of the project are real-time trajectory prediction from videos and general activity predictions such as collisions anticipation. Key topics of my project are the incremental learning, for extending the capacity of the models without forgetting the knowledge already learned, and the domain adaptation, for transferring specialized models’ ability on a task to non-specialized ones on a different task. The usage of the neural networks on real applications implies a series of constraints such as small inference time and small model size so one relevant aspect is to design light models and to implement pruning algorithms that can drastically reduce the dimensions without decreasing the performances. Another solution for real applications is to integrate into the intelligent systems an attention mechanism that can drive the focus of the model only on the relevant part of the data in order to solve the task, reducing the computational inference cost and improving the model accuracy.