Summarization of surveillance videos by using periodic recurrence information

Video surveillance and tracking systems have become an integral part of our daily life. As a consequence an enormous amount of data is being produced and recorded on a daily basis, which makes storage, analysis and information extraction extremely difficult. Some of the recent studies have especially focused on cost-effective yet efficient video summarization methods. Video summarization methods for surveillance systems primarily aim at detecting and evaluating human actions within the videos. Most of the time human actions do consist of periodic actions. In this study we propose a novel approach to video summarization, which is based on the detection of the periodicity of different actions and summarizing the video using this information. Our proposed method is based on the popular string match algorithm, longest common subsequence, to determine the shortest period of recurring human actions. The concatenation of the shortest periods then produces the summary video. In the study, the differential features of the video frames were extracted by using human position data. This features were utilized to create vectors which represent the video. Periodical human actions were detected by using this vectors in Longest Common Subsequence algorithm for the process of searching. The summary video was created using one period of each different periodical motion found. By taking extra periods into account the information recurrence rate of the generated summary video is approximately 5%. The success of the system in finding the actual periodic motions was around 90%.