Violence Detection and Localization in Video Through Subgroup Analysis

Emmeke A Veltmeijer, Morris E Franken, Charlotte Gerritsen

Keywords: violence detection, violence localization, subgroup analysis, subgroup tracking

TL;DR: Proposed is a novel approach to integrate social subgroups into violence detection, improving social awareness in surveillance videos and localizing groups involved in violent events.

Abstract:

In an era of rapid technological advancements, computer systems play a crucial role in early Violence Detection (VD) and localization, which is critical for timely human intervention. However, existing VD methods often fall short, lacking applicability to surveillance data, and failing to address the localization and social dimension of violent events.
To address these shortcomings, we propose a novel approach to integrate social subgroups into VD. Our method recognizes and tracks subgroups across frames, providing an additional layer of information in VD. This enables the system to not only detect violence at video-level, but also to identify the groups involved. This adaptable add-on module can enhance the applicability of existing models and algorithms.
Through extensive experiments on the SCFD and RWF-2000 surveillance datasets, we find that our approach improves social awareness in VD by localizing the people involved in an act of violence. The system offers a small performance boost on the SCFD dataset and maintains performance on RWF-2000, reaching 91.3% and 87.2% accuracy respectively, demonstrating its practical utility while performing close to state-of-the-art methods. Furthermore, our method generalizes well to unseen datasets, marking a promising advance in early VD.

Primary Area: representation learning for computer vision, audio, language, and other modalities

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 7409