The prediction of subway passengers' walking time during exit is an important basis for urban traffic operation and management. Accurate prediction can help alleviate subway congestion, optimize subway services, and improve passenger satisfaction. Firstly, video analysis software was used to extract the walking time and several characteristic variables of passengers during exit from surveillance videos. Secondly, in order to screen out the factors that have a significant impact on walking time, correlation analysis and optimal scaling regression model were used for factor analysis, and genetic algorithm was used to extract the optimal feature combination. Finally, the extracted features were used as input vectors and the Extreme Gradient Boosting model was used to predict walking time, with mean absolute error as the evaluation index. The experimental results show that the method proposed in this paper has good effect in predicting the behavior of subway passengers during exit, with a mean absolute error of 1.55 s, lower than the unoptimized Extreme Gradient Boosting model (1.87 s), support vector machine (2.03 s) and random forest (1.96 s) models.