Crowd Density Estimation Method Using Reduced Features Set

摘要

A new crowd density estimation method with reduced feature set is proposed that enhances both prediction errors and computational time in terms of feature extraction, training, and prediction times. Based on the filter, wrapper, and embedded approaches to feature selection, 9 different sets of selected features were generating. Features categories were ranked based on the number of features selected across all selection methods. Using each set of features, 7 different regression models got trained. Each model performance was assessed based on MSE, MAE, and MRE. Feature selection methods were ranked based on its robustness in reducing the prediction errors. Based on the experimental results, the proposed method outperforms the previous works.