Priority sampling and visual attention for self-driving car

• Autor: Flores Benites, Victor
• Formato: tesis de maestría
• Fecha de Publicación: 2023
• Institución: Universidad Católica San Pablo
• Repositorio: UCSP-Institucional
• Lenguaje: inglés
• OAI Identifier: oai:repositorio.ucsp.edu.pe:20.500.12590/17744
• Enlace del recurso: https://hdl.handle.net/20.500.12590/17744
• Nivel de acceso: acceso abierto
• Materia: Visual attention; Self-driving; Non-identically distributed data distribution; End-to-end methods; https://purl.org/pe-repo/ocde/ford#1.02.01

End-to-end methods facilitate the development of self-driving models by employing a single network that learns the human driving style from examples. However, these models face problems such as distributional shift, causal confusion, and high variance. To address these problems we propose two techniques. First, we propose the priority sampling algorithm, which biases a training sampling towards unknown observations for the model. Priority sampling employs a trade-off strategy that incentivizes the training algorithm to explore the whole dataset. Our results show a reduction of the error in the control signals in all the models studied. Moreover, we show evidence that our algorithm limits overtraining on noisy training samples. As a second approach, we propose a model based on the theory of visual attention (Bundesen, 1990) by which selecting relevant visual information to build an optimal environment representation. Our model employs two visual information selection mechanisms: spatial and feature-based attention. Spatial attention selects regions with visual encoding similar to contextual encoding, while feature-based attention selects features disentangled with useful information for routine driving. Furthermore, we encourage the model to recognize new sources of visual information by adding a bottom-up input. Results in the CoRL-2017 dataset (Dosovitskiy et al., 2017) show that our spatial attention mechanism recognizes regions relevant to the driving task. Our model builds disentangled features with low cosine similarity, but with high representation similarity. Finally, we report performance improvements over traditional end-to-end models.