Autonomous off-road vehicles: enhancing perception, remote control, and environmental impact through advanced AI techniques

This thesis addresses the critical challenges of deploying Autonomous Vehicles (AVs) in off-road environments, particularly for Search and Rescue (SAR) operations. It proposes a multi-faceted approach that encompasses enhanced perception, improved point cloud resolution, human-in-the-loop autonomy, and environmental sustainability.A novel dataset, the UWS Off-Road Dataset, was introduced, providing annotated Light Detection and Ranging (LiDAR) point cloud and image data specifically tailored for off-road scenarios. This dataset fills a critical gap in existing resources, enabling the development and evaluation of robust object detection algorithms for challenging terrains.Extensive experiments were conducted to evaluate state-of-the-art deep learning models for human detection, including both single-modal (LiDAR-based) and multi-modal (LiDAR and camera fusion) approaches. The results demonstrate the effectiveness of Point Pillars, a point-cloud-based model, in achieving high accuracy in off-road human detection.To enhance perception capabilities further, novel point cloud upsampling techniques using generative Artificial Intelligence (AI) were investigated. This approach improves the resolution and quality of LiDAR data, leading to more accurate and detailed environmental representations. The results demonstrate the effectiveness of the proposed Point Cloud Upsampling using Graph Convolutional Networks (PU-GCN) model in achieving superior performance compared to other upsampling methods.The thesis also explores human-in-the-loop autonomy by designing and implementing a remote driving framework on a commercial vehicle using 4G connectivity. This real-world validation demonstrates the feasibility of teleoperation as a complementary technology to fully autonomous driving, enabling human intervention in challenging or uncertain situations. Finally, the thesis investigates the environmental impact of AVs by developing a novel transformer-based model, CO2ViT, for accurate CO2 emission prediction. The model is evaluated in both standalone and Mobile Edge Computing (MEC) environments, highlighting its potential for real-time emission monitoring and optimization.Overall, this thesis contributes to the advancement of autonomous vehicle technology for off-road applications, with a particular focus on enhancing perception, safety, and sustainability. The developed datasets, algorithms, and frameworks offer valuable tools for researchers and practitioners working towards the deployment of robust and environmentally responsible autonomous systems in challenging off-road environments.