Motion Event Recognition Using User Feedback

Abstract: With high associated costs, both in terms of time and economics, false alarms are a ubiquitous problem in camera surveillance. There exists many methods of lowering the rate of false alarms from motion detection, most of which are based around using static rules and filters such as area of interest, cross-line detection and filtering of swaying objects. This thesis explores the possibilities of using feedback from a user in the form of examples to filter new video sequences with detected motion, enabling the suppression of motion alarms that are of no interest to a user. Two approaches to achieve this are presented, a frame-by-frame method employing a one-class support vector machine and a sequence based recurrent relational network which aims to learn to recognize similar and dissimilar video sequences. Both approaches aim to fulfill the requirement of both being able to work with only positive data, and a limited amount of it at that. The report discusses the difficulties and challenges with working with data from only a single class, and the limitations that a limited amount of training data pus on such systems. Both approaches employ transfer learning in that a pre-trained network is used to extract high level abstract features from video frames. The results of experiments run show a surprising level of accuracy with even basic methods under the right circumstances, while also revealing problems with unexpected behavior when encountering some new, unseen data. The report further discusses the problems of establishing how predictable such a system can be, and how communication with a user can be done so as to properly convey how the system can be expected to act. The report makes the conclusions that technically there are indeed possibilities to recognize events in video sequences and enabling the filtering thereof upon example videos, but the problem of understanding what exactly it is in a sequence that a user may want to ignore and to communicate what exactly it is a system has learned to recognize is a very difficult one. The report concludes with suggesting a couple of avenues of future work, both when it comes to the general problem of using feedback from a user, and when it comes to the more technical aspects.