An important application in surveillance is to apply computerized methods to automatically detect anomalous activities and then notify the security officers. Many methods have been proposed for anomaly detection with varying degree of accuracy. They can be characterized according to the approach adopted, which is supervised or unsupervised, and the features used. Unfortunately, existing literature has not elucidated the essential ingredients that make the methods work as they do, despite the fact that tests have been conducted to compare the performance of various methods. This paper attempts to fill this knowledge gap by studying the videos tested by existing methods and identifying key components required by an effective unsupervised anomaly detection algorithm. Our comprehensive test results show that an unsupervised algorithm that captures the key components can be relatively simple and yet perform equally well or better compared to existing methods.
For the sake of visualization, each anomalous feature point detected by OCCAM is visualized as a circle with radius equal to four pixels as shown in the following figures. OCCAM's localization results vs. ground truth. Red pixels are true positives (TP), blue pixels are false postives (FP), yellow pixels are false negatives (FN), and the rest gray pixels are true negatives (TN).
REMEMBER:
UCSDped1 and UCSDped2 are the only datasets that have the ground truth at pixel level.
UCSDped1
UCSDped2
UCSDped1_Test003 | UCSDped1_Test004 | UCSDped1_Test018 |
UCSDped1_Test019 | UCSDped1_Test024 | UCSDped1_Test032 |
UCSDped2_Test001 | UCSDped2_Test002 |
UCSDped2_Test003 | UCSDped2_Test006 |
We encourage you to refer to Abuolaim's thesis for broader review of related work, analysis, results, and discussions: