Vision-based vehicle trajectory data provides useful information for evaluating roadway safety. However, the prediction accuracies of using this approach are often not evaluated. This study investigated the accuracy of a video-based tool that predicted near-miss crashes at signalized intersections using two signalized intersections in Overland Park, Kansas. Video data from both intersections was collected on ten weekdays between February and March in 2021. About ten percent of the data were sampled for manual validation, including drawing vehicle trajectories and conflict spots, and measuring time in milliseconds. Two surrogate safety measures, namely Post-Encroachment Time (PET) and Time-to-Collision (TTC), were applied based on three conflict categories: critical conflicts, minor conflicts, and potential conflicts. In addition, three separate conditions were considered related to weather and traffic: rainy peak, sunny peak, and sunny off-peak conditions. Four performance measures were used to compare the manually and automatically extracted surrogate safety measures, including Mean Absolute Deviation, Root Mean Squared Error, Mean Absolute Percentage Error, and Root Mean Squared Log Error. An Analysis of Variation test was also conducted for each analysis.