Department of Computer Engineering
Anomaly Detection in Egocentric Traffic Videos
(Supervisor: Asst.Prof.Dr.Shervin Rahimzadeh Arashloo)
Computer Engineering Department
Abstract: Video anomaly detection (VAD) has been extensively studied. However, research on egocentric traffic videos with dynamic scenes is still open to dig in. For detecting anomalies, One-Class Classification (OCC) methods were used widely in the literature and regarding the provided information for OCC models, previous studies were categorized either as Frame-Level or Object-Centric. Since in Object-Centric models, only the bounding boxes of objects get predicted, they have less computational complexity compared to Frame-Level models and as a result, we prefer this approach in our problem. Based on the research we conducted up until now, Kernel methods were not studied for this application and we will encode information from past frames, e.g., ego and objects’ motion using kernel methods to predict the objects' future bounding boxes.
DATE: 25 April, Monday @ 15:50