Defesa de Tese de Doutorado – Elton Figueiredo de Souza Soares
Título do Trabalho: Privacy-Preserving Training of Monocular Depth Estimators for Autonomous Vehicles with Federated Self-Supervision, Bayesian Optimization, and Continual Learning
Resumo: Artificial intelligence has been used to enhance intelligent transportation systems, with a particular emphasis on autonomous vehicles. Computer vision is a crucial area of research for successful deployment of those vehicles. Image-based depth estimation task has gained significant attention in this area due to its cost-effectiveness and wide range of applications. Its monocular variant has been the main focus of research due to its versatility, since it requires only one camera. Although latest methods leveraged self-supervised learning successfully, they did not consider key requirements for use in autonomous vehicles, such as privacy preservation, reduced network consumption during training, and optimized computation cost distribution. Meanwhile, federated learning has been shown to help address these requirements in multiple domains. Thus, we propose three incremental methods that combine federated learning and self-supervision to enable the training of monocular depth estimators with equivalent efficacy and superior efficiency compared to current methods. The first method proposed uses standard federated averaging as aggregation function, while the second one applied Bayesian optimization to enhance the aggregation. On the third method, we applied standard federated averaging combined with regularization and experience replay to tackle the continual learning in autonomous vehicle scenarios. Our experiments, conducted using public benchmarks, demonstrate that our solutions achieve state-of-the-art performance, with greater efficiency than prior solutions.
Palavras – Chaves: Monocular Depth Estimation, Federated Learning, Intelligent Transportation Systems, Autonomous Vehicles, Computer Vision, Artificial Intelligence
Banca Examinadora:
– Carlos Alberto Vieira Campos, UNIRIO (Presidente)
– Sidney Cunha de Lucena, UNIRIO
– Carlos Eduardo Ribeiro de Mello, UNIRIO
– Miguel Elias Mitre Campista, UFRJ
– Joel José Puga Coelho Rodrigues, UFPI
Data e Hora da Defesa: 04/02/2025 às 16:00
Local da Defesa: Auditório PPGI (https://meet.google.com/jkd-ztey-vjn)
