Fractional Data Distillation Model for Anomaly Detection in Traffic Videos

Timely automatic detection of anomalies like road accidents forms the key to any intelligent traffic monitoring system. In this paper, we propose a novel Fractional Data Distillation model for segregating traffic anomaly videos from a test dataset, with a precise estimation of the start time of the anomalous event. The model follows a similar approach to that of the typical fractional distillation procedure, where the compounds are separated by varying the temperature. Our model fractionally extracts the anomalous events depending on their nature as the detection process progresses. Here, we employ two anomaly extractors namely Normal and Zoom, of which former works on the normal scale of video and the latter works on the magnified scale on the videos missed by the former, to separate the anomalies. The backbone of this segregation is scanning the background frames using the YOLOv3 detector for spotting possible anomalies. These anomaly candidates are further filtered and compared with detection on the foreground for matching detections to estimate the start time of the anomalous event. Experimental validation on track 4 of 2020 AI City Challenge shows an s4 score of 0.5438, with an F1 score of 0.7018.