Vision-Based People Counter Using CNN-Based Event Classification

Abstract

This article proposes a convolutional neural network (CNN)-based people counter that classifies a given frame cube to a specific event that indicates people entering or exiting a target area to measure instantaneous people count. For the training of the proposed CNN, a training input frame cube and its corresponding class label that represents a specific event are generated using the proposed counting rules. For mitigating the overfitting problem that may occur in the training of the proposed CNN, data augmentation, and postclass correction using foreground distribution with event probabilities are applied. The experimental results indicate that the proposed method improved the F1 score and accuracy for the cumulative people counting results by up to 9.0% and 14.8%, respectively, compared with those of the benchmark methods, even though it calculated the cumulative count by summing instantaneous people counts, while the benchmark methods were optimized for the calculation of the cumulative count.