MACHINE LEARNING SYSTEM FOR ANOMALY DETECTION OF PEDESTRIAN FLOW

Abstract

In this study, a practical system which consists of infrared sensors and computer programs for machine learning was proposed for anomaly detection of pedestrian flow. Utilizing time-series data collected by the sensors at an entrance of a building in a university as training data for machine learning, the proposed system successfully classified the data into two groups, based on characteristics of flow of visitors to the building. Subsequently, the system was applied to data set which is collected in real-time, and it was shown that the system can detect properly abnormal events that occur only with low probability.