Limitations with measuring performance of techniques for abnormality localization in surveillance video and how to overcome them?

Abstract

Now a days video surveillance is becoming more popular due to global security concerns and with the increasing need for effective monitoring of public places. The key goal of video surveillance is to detect suspicious or abnormal behavior. Various efforts have been made to detect an abnormality in the video. Further to these advancements, there is a need for better techniques for evaluation of abnormality localization in video surveillance. Existing technique mainly uses forty percent overlap rule with ground-truth data, and does not considers the extra predicted region into the computation. Existing metrics have been found to be inaccurate when more than one region is present within the frame which may or may not be correctly localized or marked as abnormal. This work attempts to bridge these limitations in existing metrics. In this paper, we investigate three existing metrics and discuss their benefits and limitations for evaluating localization of abnormality in video. We further extend the existing work by introducing penalty functions and substantiate the validity of proposed metrics with a sufficient number of instances. The presented metric are validated on data (35 different situations) for which the overlap has been computed analytically.