Applications of Computer Vision for the Improvement of Autonomous Vehicle Design

Abstract

This dissertation approaches methods of improving autonomous vehicle design through two separate lenses. In the first study, we investigate how technological transparency can improve driver trust in artificial intelligence and ultimately encourage the adoption of automated driving systems. Automated driving systems provide a means of reducing the inherent danger of operating a personal motor vehicle. However, barriers to adoption exist due to low trust in the artificial intelligence that powers the systems. To fill this deficit of trust, Chapter 3 proposes a deep learning-based visual alert system that allows passengers to monitor the artificial intelligence performance in real-time. Using a trained object detection model, we design a novel perception augmentation system for conveying information about the driving scene to the passenger through the lens of artificial intelligence. We conduct an empirical study that confirms that the proposed system improves the trust in the underlying artificial intelligence technology. Trust in artificial intelligence is also found to not only positively affect the perceived benefit from– and intention to use an automated driving system, but also negatively influence the perceived risk associated with using the technology. Perceived enjoyment from the autonomous vehicle is also found to have a strong effect on the perceived benefit from– and intention to use the system.

In the second study in this dissertation, we take a close look at methods to improve the quality of sensor data in automated driving systems. Although deep learning methods continue to set new state-of-the-art metrics on deblurring benchmarks, a comprehensive understanding of what losses are effective for the deblurring task is missing from the literature. The study in Chapter 4 provides an empirical foundation for the selection of a loss function when developing image deblurring models. Despite the popularity of mean squared error as a content loss function for image restoration tasks, we demonstrate that mean absolute error produces higher quality results to the human visual system. Furthermore, we show that deblurring models trained solely using a perceptual content loss produce outputs that are perceptibly better than the same model trained using a plain mean absolute error or mean squared error loss despite validation metrics that would indicate otherwise. Finally, we demonstrate that adversarial losses do not produce generators capable of confidently deblurring images in the absence of auxiliary loss functions; however, the combination of adversarial and content losses in some cases produces higher quality results than either constituent loss when trained in isolation. Compared to state-of-the-art methods, the best model developed in this work produces worse quantitative validation metrics, but visibly better results on real-world blurs in natural images.