Abandoned Object Detection in Video-Surveillance: Survey and Comparison

Area of Science:

• Computer Science
• Artificial Intelligence
• Computer Vision

Background:

• Abandoned object detection is crucial for video surveillance, with many systems developed to monitor public and private spaces.
• Current research often addresses individual components like segmentation and validation separately, limiting the understanding of their impact on the full pipeline.

Purpose of the Study:

• To formalize the abandoned object detection framework and review state-of-the-art approaches for each stage.
• To develop and evaluate a multi-configuration system to identify the best-performing combination of detection stages.
• To provide an online tool and a comprehensive dataset for the research community.

Main Methods:

• A systematic review of existing methods for foreground segmentation, stationary object detection, and abandonment validation.
• Development of a flexible, multi-configuration system enabling the selection and testing of various component alternatives.
• Creation of a heterogeneous dataset incorporating diverse challenges like illumination changes, shadows, and dense moving objects.

Main Results:

• Identification of the most effective configurations for abandoned object detection systems.
• Highlighting design choices that enhance robustness against errors in complex surveillance scenarios.
• Validation of the optimal configuration on previously unseen datasets, demonstrating generalizability.

Conclusions:

• The study provides a comprehensive analysis of abandoned object detection pipelines, offering insights into optimizing system performance.
• The developed multi-configuration system and dataset serve as valuable resources for advancing research in this field.
• Further research is needed to address remaining open challenges identified through experimental comparisons.