Developing vehicles capable of operating without human drivers has long been a dream. Autonomous driving systems are the culmination of this dream, potentially becoming a reality if we can overcome challenges related to accuracy, response time, safety, and hardware efficiency.
Deep Neural Networks (DNNs) represent a significant advancement in data-driven perception models, revolutionizing computer vision technologies. If developed and trained properly, these models can provide accurate results within an acceptable inference time and become reliable components of autonomous driving systems.
This thesis presents our contributions to the development of accurate yet fast and hardware-efficient DNN models for safety-critical autonomous driving applications such as lane detection, stereo vision, and trajectory prediction. It particularly highlights our neural architecture search (NAS) methods to address the aforementioned challenges. Additionally, we introduce our distinct self-supervised learning method for lane detection, which is critical for training accurate models without relying on human-labeled data. This method significantly enhances the accuracy and generalization of lane detection methods across a wide range of images featuring different weather conditions, textures, and geographies.
The research presents several key contributions. Firstly, the thesis introduces a novel NAS method for 3D lane detection, named 3DLaneNAS, which provides accurate lane capturing on mountain roads and outperforms existing models in both accuracy and latency. Secondly, we propose a fast and accurate NAS method for stereo vision, named FastStereoNet, which is crucial for computing the distance of surrounding obstacles using cameras, and demonstrates superior accuracy and latency across GPUs, CPUs, and FPGAs. Thirdly, we develop a self-supervised lane detection method named CLLD that leverages strong lane priors and a cross-similarity operation to detect lanes robustly under diverse conditions. Fourthly, we introduce a NAS method for trajectory prediction using Lidar data, named TrajectoryNAS, which achieves state-of-the-art results in trajectory prediction with the lowest inference time. Finally, the thesis contributes to model compression by introducing advanced post-training quantization (PTQ)techniques, including cluster-affine transformations and output reconstruction, which significantly improve the accuracy of ultra-low-bit models while ensuring efficiency on resource-constrained hardware.
In conclusion, this thesis has successfully introduced several innovative methods tailored for autonomous driving applications, making substantial advancements in 3D lane detection, stereo vision, lane detection, trajectory prediction, and quantization. Each proposed approach demonstrates significant improvements over existing techniques in terms of accuracy, latency, generalization, and hardware efficiency across different driving conditions and platforms.
For future work, we aim to further enhance robustness and generalization by extending our methods with vision transformers and exploring hybrid NAS–quantization pipelines. Such approaches are expected to leverage deeper contextual information from visual data while further reducing computational demands, thereby improving the reliability of autonomous driving technologies in diverse real-world environments.
Västerås: Mälardalens universitet, 2026.
Autonomous vehicles; Computer Vision; Efficient Models; Self-supervised Learning; Quantization; Neural Architecture Search
2026-11-06, Kappa, Mälardalens universitet och digitalt, Västerås, 13:30 (English)