Efficient vehicle detection and counting are crucial for effective highway management, yet the varying sizes of vehicles present significant challenges to accurate detection. This paper introduces a vision based vehicle detection and counting system, supported by a novel high-definition highway vehicle dataset comprising 57,290 annotated instances across 11,129 images. Unlike existing public datasets, our dataset includes annotations for smaller vehicle objects, providing a comprehensive data foundation for deep learning-based vehicle detection. We leverage the capabilities of deep convolutional networks (CNNs), which excel in learning image features and can simultaneously perform multiple tasks, such as classification and bounding box regression. Our detection approach is rooted in the You Only Look Once (YOLO-V4) model, combined with the Deep SORT algorithm for real-time vehicle tracking. YOLO-V4 generates vehicle detections from video frames, while Deep SORT enhances tracking accuracy by mitigating false predictions from YOLO-V4. The video input is processed frame-by-frame, allowing for real-time vehicle detection and tracking. The proposed model is rigorously trained using a blend of public and custom-collected datasets, ensuring robust performance in diverse traffic scenarios. This work represents a significant advancement in vehicle detection and tracking methodologies, contributing to improved accuracy in highway management systems.