Anomaly activity classification in the grocery stores

Abstract

Nowadays, because of the growing number of robberies in shopping malls and grocery stores, automatic camera’s applications are vital necessities to detect anomalous actions. These events usually happen quickly and unexpectedly. Therefore, having a robust system which can classify anomalies in a real-time with minimum false alarms is required. Due to this needs, the main objective of this project is to classify anomalies which may happen in grocery stores. This objective is acquired by considering properties, such as; using one fixed camera in the store and the presence of at least one person in the camera view. The actions of human upper body are used to determine the anomalies. Articulated motion model is used as the basis of the anomalies classification design. In the design, the process starts with feature extraction and followed by target model establishment, tracking and action classification. The features such as color and image gradient built the template as the target model. Then, the models of different upper body parts are tracked during consecutive frames by the tracking method which is sum of square differences (SSD) combined with the Kalman filter as the predictor. The spatio-temporal information as the trajectory of limbs gained by tracking part is sent to proposed classification part. For classification, three different scenarios are studied: attacking cash machine, cashier’s attacking and making the store messy. In implementing these scenarios, some events were introduced. These events are; basic (static) events which are the static objects in the scene, spatial events which are those actions depend on coordinates of body parts and spatio-temporal events in which these actions are tracked in consecutive frames. At last, if one of the scenarios happens, an anomalous action will be detected. The results show the robustness of the proposed methods which have the minimum false positive error of 7% for the cash machine attack and minimum false negative error of 19% for the cashier’s attacking scenario.