Anomalous Motion Detection of Vehicles on Highway using Deep Learning

Abstract

Research in visual anomaly detection draws much interest due to applications in surveillance. Common data sets for evaluation are constructed using a stationary camera overlooking an area of interest. Despite the challenges of learning from a single class of data in an unsupervised learning paradigm, previous research shows promising results in detecting spatial as well as temporal anomalies in crowded environments. The advent of self-driving cars provides an opportunity to apply visual anomaly detection in a more dynamic application yet no data set exists to evaluate anomaly detection models in this setting. This thesis presents a novel anomaly detection data set for the problem of detecting anomalous traffic patterns from dash cam videos of vehicles on highways. I evaluate state-of-the-art unsupervised deep learning anomaly detection models as well as propose novel variations and discuss the exacerbated challenges of this new data set.