University of Central Florida at TRECVID 2008 Content Based Copy Detection and Surveillance Event Detection

Omer Bilal Orhan, Jason Hochreiter, Jonathan Poock, Qinfeng Chen, Ajay Chabra, Mubarak ShahTRECVID 2008 · OpenReview

Abstract

This paper presents the methodologies and results from the University of Central Florida (UCF) team's participation in the TRECVID 2008 evaluation. The research focused on two distinct domains: Content-Based Copy Detection (CBCD) and Surveillance Event Detection (SED). For CBCD, UCF introduced a robust sequence matching framework utilizing Normalized Hu-Moment Invariants (NHMI) to overcome the limitations of traditional frame-by-frame matching. For SED, the team developed specialized detectors based on low-level vision heuristics. The results demonstrate that global image signatures and temporal sequence analysis provide superior robustness for copy detection, while surveillance tasks remain challenging due to the gap between motion detection and behavioral understanding.

Introduction

As digital video content proliferated in the late 2000s, the need for automated systems to manage and monitor this data became critical. The TRECVID 2008 evaluation provided a benchmark for two major challenges: identifying transformed video copies (CBCD) and detecting specific human actions in complex surveillance environments (SED). UCF's approach prioritized robustness and efficiency, moving away from computationally expensive local features toward global signatures and rule-based heuristics.

Content-Based Copy Detection (CBCD)

Normalized Hu-Moment Invariants (NHMI)

The core innovation in UCF’s CBCD system was the use of global image signatures based on Hu-moments. These moments are mathematically invariant to translation, scale, and rotation. To address sensitivity to lighting and gamma changes, the team normalized these moments against the first moment, creating a signature robust to significant quality degradation.

Sequence Matching and Alignment

Rather than matching individual frames, the system employed a sliding window to match sequences of NHMI features. By analyzing the temporal "rhythm" of a video, the system could identify matches even when individual frames were heavily distorted. A key verification step involved looking for a "V-shaped" error curve; as the search window aligns with the true match, the error reaches a distinct minimum before rising again.

Figure 3: Alignment of Normalized Hu-Moment Invariants between a query and target video sequence.

Handling Complex Transformations

The CBCD task included challenging transformations such as picture-in-picture (PiP), where a video is resized and embedded within another. UCF utilized Hough Lines to detect persistent horizontal and vertical boundaries, allowing the system to localize and extract the sub-video before feature extraction.

Figure 6: Examples of various video transformations in the TRECVID 2008 CBCD task, including picture-in-picture and text overlays.

Surveillance Event Detection (SED)

For the SED task, the team developed four specialized detectors based on physical assumptions and low-level motion analysis:

  • PersonRuns: Utilized optical flow to cluster feature points, tracking the movement of these clusters rather than individual skeletons.
  • ObjectPut: Analyzed a gridded scene for consistent downward optical flow over a temporal window.
  • TakePicture: Employed a heuristic to search for sudden, localized intensity spikes characteristic of camera flashes.

Experimental Results

CBCD Performance

The UCF system demonstrated significant dominance in the CBCD task, achieving the highest F1 scores for nine out of ten transformation types. The system proved highly efficient; despite indexing a corpus of approximately 20 million frames (generating 120 million features), the matching algorithm accounted for less than 20% of total processing time.

SED Challenges

While the CBCD results were stellar, the SED task highlighted the limitations of motion-based heuristics. The OpposingFlow detector achieved 75% recall but suffered from precision below 1%. Similarly, the PersonRuns detector caught 45% of events but struggled with false alarms (4% precision), as the system often confused fast walking with running.

Conclusion

UCF’s participation in TRECVID 2008 underscored the power of robust, global signatures for video copy detection. By focusing on sequence matching rather than frame-level detail, the system achieved state-of-the-art results across diverse transformations. However, the surveillance results indicate that motion alone is insufficient for high-precision event detection. Future work must integrate more sophisticated behavioral models and camera-specific tuning to bridge the gap between low-level vision and high-level activity understanding.

Summary updated

BibTeX
@inproceedings{orhan2008university,
  author = {Omer Bilal Orhan and Jason Hochreiter and Jonathan Poock and Qinfeng Chen and Ajay Chabra and Mubarak Shah},
  title = {{University of Central Florida at TRECVID 2008 Content Based Copy Detection and Surveillance Event Detection}},
  booktitle = {TRECVID},
  year = {2008},
  url = {https://openreview.net/forum?id=hZNfyRhJ0s},
}
Figure