Real-time driving monitoring system on a single-board computer utilizing deep neural networks integrated with MiDAS

Abstract

In recent years, road safety has emerged as a critical concern due to the increasing number of accidents attributed to driver negligence and fatigue. This thesis addresses this pressing issue by proposing a Real-Time Driving Monitoring System designed for deployment on a single-board computer. The system employs a combination of cutting-edge technologies to comprehensively assess driver safety during operation. The system’s core objective is to discern whether the driver is operating the vehicle in a safe manner. To achieve this, three distinct input streams from specialized cameras are utilized. The first input stream leverages YOLOPv2, a state-of-the-art object detection model, to accurately detect road lanes and determine if the vehicle remains within the designated lane. This real-time feedback is crucial for preempting potential lane departure incidents. The second input stream employs Monocular Depth Estimation with MiDAS, a robust and efficient technique for gauging the distance of objects in close proximity to the vehicle. By aggregating depth measurements and calculating a mean depth value, the system establishes an empirical threshold. Instances where the mean depth falls below this threshold are indicative of potential collision risks, prompting the system to identify the driver as unsafe. Furthermore, the third input stream utilizes the front-facing camera to monitor driver behavior and detect signs of drowsiness. Through a combination of facial feature analysis and eye tracking, the system can accurately determine if the driver exhibits signs of fatigue or inattentiveness. Should the driver display drowsiness for a duration surpassing the specified threshold, an alert is triggered, thereby mitigating the risks associated with driver fatigue. In the event that any of the aforementioned conditions persist for a predetermined duration, the system activates an alert protocol. This protocol includes the illumination of LED indicators and the sounding of a buzzer, providing immediate feedback to the driver and drawing attention to the potential safety hazard. By combining these advanced technologies in a single-board computer-based system, this thesis presents a comprehensive approach to real-time driving monitoring. The integration of YOLOPv2 and MiDAS with deep neural networks ensures accurate and timely detection of potential safety risks, thereby contributing significantly to the enhancement of road safety standards.