Multiple Regression Analysis of Explanatory Variables for Pedestrian Counting in Video Sequences

Abstract

In recent years, video processing-based methods for counting pedestrians in a crowded scene have been proposed. In these methods, the relationships between various features in video sequences and the number of pedestrians are obtained using training data. Then, the number of pedestrians is estimated from the video features and the pre-learning relationships. However, for each method, different video features are used and its performance is evaluated under different environments. Therefore, it is difficult to compare the effectiveness of video feature for estimating the number of pedestrians. In this report, we investigate the relationships between various video features and the number of pedestrians using multiple regression analysis under same evaluation environments. We use the size of moving objects, the form of moving objects, the number of optical flows, and the number of clusters of optical flows as video features. Through analysis using actual video sequences, we show that combining the number of optical flows and form of moving objects is effective for estimating the number of pedestrians under crowded situations.