Improvements to the operation of autonomous vehicles (AVs) in Smart Cities

Abstract

Autonomous vehicles (AVs) are vehicles that either partially or fully carry out the task of driving themselves. They use systems and technologies such as Advanced Driver Assistance Systems (ADAS), Adaptive Cruise Control (ACC), Lane-keeping systems, self-parking technology, and crash warning systems. AVs are an application of intelligent transportation systems (ITS). AVs are increasingly becoming commercial, and stakeholders in the transport sector forecast their commercial use by 2025 (Criddle, 2021). In this thesis, we proffered some design solutions to improve the safety of autonomous vehicles (AVs) operation in smart cities. We identified three key areas: security of stored and extracted data, deep learning (DL) model training, and safeguarding the inherent weakness with the human element. We considered the existing/related systems, the process, challenges, and improvement of each application area's current or related system.Firstly, we designed and implemented a forensic process for digital data extraction and storage from AVs, which protects against manipulation of the data by archiving, hashing, and encrypting the data. Secondly, we found critical errors and omissions in the Udacity selfdriving car dataset – a widely used dataset for open-source AV research. There were many illusive annotations, duplicated bounding boxes, and oversized bounding boxes. We improved the YOLOv4 network model's detection accuracy on common stationary and moving objects in a driving scene by relabeling missing objects in the Udacity self-driving car dataset and performing a series of preprocessing and augmentation steps. Our results showed an improved speed and accuracy in detection. Thirdly, we built a small-scale system that solves the challenges faced by standard cameras and computer vision algorithms on selfdriving cars by detecting and recognising active traffic lights under different environmental conditions. Our solution can accurately identify a red-yellow-green traffic light under harsh environmental conditions and act accordingly. Finally, to safeguard the inherent weakness with the human element in AVs, we developed a system that detects a drowsy driver by counting not only eye blinks but also counts yawns which is another indication of a drowsy driver. Our system uses the HAAR feature-based cascade classifier because though less accurate, the classifier is significantly faster than commonly used detectors like the HOG and Linear SVM detectors (Li, J. et al., 2011; Rahmad, C. et al., 2020; Zhu, Q. et al., 2006), and therefore ideal for AVs. All the proposed designs are implemented and analysed to demonstrate feasibility and measure their performance.

Date of Award	Jul 2021
Original language	English
Awarding Institution	University of Bedfordshire
Supervisor	Xiaohua Feng (Supervisor) & Dayou Li (Second supervisor)