Computer Vision-Based Dangerous Scenes Detection System for Advanced Driving Assistance

This thesis presents a comprehensive study on the development of a computer visionbased system for Advanced Driver Assistance Systems (ADAS). The research initially explored a classical computer vision approach, which involved employing detection and tracking algorithms and monocular depth estimation to perceive the external environment. Moreover, the study focused on integrating the driver’s attention state through its gaze projected from a pair of eye-tracking glasses onto the external scene, through a roof-mounted camera. The primary objective was to analyze the driver’s behavior by comparing its gaze with the locations of vulnerable road users, thereby proposing an initial safety scheme. Despite the promising conceptual framework, it was observed that many of the conventional methods for extracting indirect features were not sufficiently robust in real-world driving scenarios. These methods struggled particularly under varying light conditions and during critical situations, highlighting significant limitations. In response to these challenges, the research transitioned to a deep learning-based approach. The core investigation centered on the capabilities of a Vision Transformer (ViT) in extracting human decision-making biases inherent in detecting dangerous driving situations. This approach was further improved by employing semi-supervised learning techniques, which leverage the vast amounts of easily accessible unlabeled data, thus addressing the challenges associated with the expensive and labor-intensive labeling process. The outcomes of this research show the substantial potential of the attention mechanism, on which vision transformer is based, in enhancing the robustness and reliability of ADAS. The study also opens up new themes for future research by identifying and addressing the challenges encountered during the development process. These findings underscore the critical role of computer vision in the advancement of ADAS, emphasizing its significance in improving driver assistance systems through more accurate and reliable perception mechanisms. In conclusion, this thesis contributes to the field of ADAS by demonstrating how modern computer vision techniques can be effectively integrated into driver assistance systems. It highlights the potential of deep learning, especially in overcoming the limitations of traditional methods, and points to future innovations for safer and more efficient driving experiences.