Risks of applying deep learning in autonomous vehicle systems: a literature review (#648)

Authors

Zegarra Ramos, Matthew Stephano

Haro Garcia, Luis Angel

Ayala Ñiquen, Evelyn Elizabeth

Roque Pisconte, Vanessa Del Carmen

Abstract

Although the application of deep learning in autonomous vehicle systems consolidated over time, its use could entail certain risks. Therefore, this study aimed to identify these risks in autonomous vehicle systems through a literature review. To achieve this, the PRISMA methodology was used for the collection and selection of studies, as well as the PIOC strategy for formulating research questions, in this study that did not use meta-analysis. Based on inclusion and exclusion criteria, 27 open-access articles from the Scopus database were selected. The results showed that the application of deep learning in autonomous vehicle systems encompassed key aspects such as environmental perception, object detection, and route planning. However, significant risks were also identified, such as inaccuracies in perception, vulnerability to attacks, detection errors, and lack of interpretability of the models. To mitigate these risks, detection and evaluation techniques such as cross-validation, sensitivity analysis, and testing in simulated environments were proposed. Additionally, tests were conducted in various scenarios and conditions, such as urban, suburban, rural environments, highways, and adverse weather conditions. The research concluded that although deep learning had the potential to improve vehicle autonomy and safety, it could also present significant risks. It was recommended that future work focus on developing and validating new techniques to address these risks, as well as establishing regulatory frameworks and standards to ensure the safety and reliability of autonomous vehicles powered by deep learning.