In this set of experiments, we investigate applications of our algorithm to summarization of surveillance videos. The algorithm is modified to build the feature matrix using only H and S elements in the HSV color space. The sample videos are taken from CCTV Camera Pros
Video 1 shows a day view of a car, someone approaches the car, steals something from inside and runs away. Video 2 is similar but shot at night. For all videos the first frames which are totally dark and the final advertisements are dropped and not processed. From each video a %1 summary is produced.
| Video | #Frames | Key Frames | Summary Image |
| #1 | 490 | 2, 4, 7, 146, 200 | summary1 |
| #2 | 540 | 2, 4, 205, 416, 522 | summary2 |
| #3 | 420 | 10, 132, 396, 416 | summary3 |
| #4 | 9000 | N/A | summary4 |
Summary of Video 1 did not include any frames from the second half of the video. After comparing the sequence with that of Video 2, it was noticed that unlike Video 2, Video 1 starts when the person is in the scene. Therefore, we guess that the frames from the second half of the video seem 'normal' and not detected as significant frames. To verify our guess, we tried Video 3 which similarly to Video 1 is shot in day light but starts with no person in the scene. Since the key frames from Video 3 does not include the person running away, our guess is incorrect. Longer videos are needed for further investigation.
Video 4 is a segment of a documentary from CBC. It shows that the summarization technique performs well on different types of video.
