Overtaking of heavy vehicles, like trucks, has a significant risk due to limited visibility and potential collisions. This research work addresses the challenge by accurately detecting and tracking vehicles in real time, predicting overtaking decisions and providing information to the drivers for safer driving. In recent years, deep learning methods have shown a robust performance compared to traditional techniques, and transformed how vehicles are detected and counted in various scenarios. This research focuses on enhancing safety during overtaking on single-lane roads by developing a vehicle detection and counting system using the MobileNet-SSD deep learning model. Through meticulous testing on Common Objects in Context (COCO) dataset, this system achieves an average accuracy of 98.7% in detecting and counting vehicles, demonstrating its efficacy and reliability. These results demonstrate the potential of intelligent vision-based systems to significantly improve safety and traffic .

Vehicle counting plays an important role in traffic management and surveillance systems. Counting the vehicles is challenging due to various factors, such as lighting variations, occlusions, and diverse vehicle types. Existing methods for vehicle counting often relied on manual or semi-automated methods, which are both labor intensive and are also bound to have some human error. Automating the counting process using deep learning algorithms can provide a more efficient and accurate solution. This work aims to detect and count vehicles in real-time video streams for which real-time recorded video datasets are used. A centroid-based tracking algorithm has been implemented to track the vehicles across consecutive frames, the algorithm associates the center points of vehicles over time, thereby enabling the tracking of individual vehicles as they move throughout the video. The evaluation demonstrates YOLOv8's high accuracy in detecting and counting vehicles in comparison with YOLOv5, which makes it suitable for diverse real-world applications such as traffic flow analysis, congestion management, and the system's real-time processing capabilities which further enhance its practical utility in surveillance systems.

Driver drowsiness is a critical factor contributing to road accidents worldwide. The accidents can be prevented if warning is provided on time to the drowsy driver. This research proposes a deep learning based approach for detection of drowsiness in drivers. The system leverages computer vision techniques, specifically the calculation of Eye Aspect Ratio (EAR), to monitor and detect signs of drowsiness by analysing facial landmarks. EAR is computed to determine eye closure patterns indicative of drowsiness. To enhance accuracy, a method for detecting sunglasses is integrated in the proposed system by utilizing Hue, Saturation and Value (HSV) color space and color masks to identify regions obstructing the eyes. Upon detecting drowsiness, the system triggers visual and auditory alerts to notify the driver, promoting timely intervention for safer driving. Experimental results demonstrate the system's effectiveness in various lighting conditions and validate its ability to differentiate between drowsy and alert states with high accuracy. The proposed approach not only addresses the challenge of detecting drowsiness in diverse conditions but also contributes to mitigating road accidents caused by driver fatigue, thereby enhancing overall road safety.