This paper presents a case study of an integrated practice-oriented methodology to continuously monitor intersections' capacity from a data stream of vehicle trajectories after a one-time calibration with information automatically collected from video images through object recognition and tracking. Such an approach would allow a quick adaptation to upcoming changes in traffic flow physics with connected and automated vehicles, as well as the spread in time and space of the knowledge of effective capacities, filling the current gap in the practice that relies on highway capacity manuals or punctual measurements. Historical data from probe trajectories are used to extract common features of the signalized approach such as cycle time, green start and stop line position. Applying computer vision machine learning on short video recordings of the signalized approach we track vehicles and extract specific features such as queue density (equivalently spatial headway) and the relationship between vehicles speed profiles and vehicle sizes. These parameters help extract saturation flow rate values from newly observed probe trajectories. The same information extracted from videos serve as a comparison basis in spite of different time frame.