K-Nearest Neighbor-Based Video Anomaly Detection Approaches

In the previous chapter, we described a probability-based video anomaly detection approach to generate a set of potential frame-level video anomalous images or image subsequences. In this chapter, we look at another popular method of generating more complex frame-level (although not necessarily so) sequences of anomalous videos, named k-nearest neighbor-based anomaly detection method, which relies on the distances to one’s k-nearest neighbors as a measure of deviation from being normal and can further eliminate the false positives from the results detected by the probability-based methods if such false positives do exist. It is also a very useful approach for classification when existing images or image subsequences are already labeled and used to predict a label for the new-coming video subsequences. We will also look at Python anomaly detection package named PyOD.