Comparing Sparse and Dense Optical Flow Methods to Detect Traffic Anomalies, Based on Orientation

Abstract

In intelligent traffic systems, it is often important to detect anomalous events in order to facilitate the avoidance of accidents and the improvement of traffic safety. Automatic anomaly detection helps human operators in detecting anomalous traffic events. For this
study vehicle orientation is proposed as an approach to recognize anomalous events in traffic situations, by analyzing traffic surveillance video images. The study also compares the use of sparse optical flow and dense optical flow methods to obtain orientation features. A One-Class Support Vector Machine model is built to classify feature points into anomalous or “usual” events. Experiments conducted have demonstrated that the proposed method could reliably detect recorded anomalous events, such as vehicles driving against the traffic direction or committing illegal lane crossings.