To promote the construction of freeway intelligence and improve the effectiveness and safety of human-machine co-driving, this study proposed an estimating method for the take-over reaction time of human-machine co-driving on freeways. A straight section of a closed freeway was used as the driving simulation scenario, and feature variables of the driving behavior were selected to estimate the take-over reaction time with the random forest method. The results show that the mean absolute percentage error of the proposed model is controlled within 13.56%, and its error fluctuation magnitude and dispersion degree are at better levels than those of other regression models. The importance of the feature variables was ranked, and it is found that the take-over reaction time budget has the greatest influence on the take-over reaction time prediction model, followed by driver proficiency and surrounding vehicle conditions. This research provides a reference for the development of smart freeways and the evaluation, design, and improvement of driving control take-over strategies.