Copyright © 2026 Authors retain the copyright of this article. This article is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

@article{194757,
        author = {Peddeni Mokshith and Pulla Chaitanya and Moriboyina Deekshith Kumar and Dara Manish Kumar},
        title = {AI-Based Abnormal Behavior Detection in Public Gatherings},
        journal = {International Journal of Innovative Research in Technology},
        year = {2026},
        volume = {12},
        number = {10},
        pages = {5416-5425},
        issn = {2349-6002},
        url = {https://ijirt.org/article?manuscript=194757},
        abstract = {Public environments such as railway stations, sports arenas, protest sites, and religious gatherings frequently attract large crowds where unusual behavior can rapidly develop into dangerous situations such as panic movement, physical conflicts, or stampedes. Monitoring these environments using traditional CCTV systems is challenging because security personnel must observe numerous video feeds simultaneously, which often leads to delayed reactions, operator fatigue, and reduced situational awareness. To address these limitations, this study proposes a real-time artificial intelligence framework for detecting anomalous activities in crowded public spaces through pose-based computer vision analysis. The system employs YOLOv8m-Pose to estimate human body keypoints and BoTSORT to maintain consistent multi-object tracking and identity association across consecutive frames. Each detected individual is modelled using the 17-keypoint representation defined in the COCO dataset, allowing the system to analyze body posture, movement patterns, and interactions between individuals in the scene. The proposed approach introduces a three-layer anomaly detection strategy. The first layer focuses on temporal behavioral patterns through tracked detectors, the second layer identifies immediate anomalies from individual frames using instant detectors, and the third layer analyzes collective crowd behavior through aggregate detectors. Using these mechanisms, the system can recognize ten types of abnormal events, including falling, running, loitering, stumbling, sudden stopping, crouching, physical fighting, stampede-like motion, ignored fallen persons, and abnormal crowd density. To enhance reliability, the framework incorporates a multi-signal quorum voting mechanism along with hysteresis-based state stabilization to reduce noisy predictions and maintain consistent anomaly labelling across frames. Experimental evaluation demonstrates that the proposed system can effectively perform real-time anomaly detection in crowded environments, indicating that pose-driven behavioral analysis is a promising approach for intelligent video surveillance applications.},
        keywords = {Computer vision, crowd anomaly detection, intelligent video surveillance, multi-object tracking, pose-based behavior analysis.},
        month = {March},
        }