A comparison of machine learning methods for queue length detection

Abstract

Queues are formed by people waiting for a service in public institutions and they can be defined as orderly groups of people. Automatically counting the number of people waiting in a queue through video camera footage would provide these institutions with valuable information with regards to customer service quality. In this paper, our goal is to compare several machine learning methods for finding the total number of people waiting in a queue given video camera frames. We approached this problem as a regression task. We used a subset of the Collective Activity Dataset and compared three different methods. The first two methods used bounding box coordinates and orientations provided by the dataset, while the last method utilized the bounding box coordinates to extract feature maps from the frames using RoiAlign. The first method used XGBoost, while the latter methods used Convolutional Neural Networks (CNNs). Results show that the method using RoiAlign presents the best prediction performance in terms of mean squared error and mean absolute error, compared to other methods.