Real-Time Human Group Detection and Clustering in Crowded Environments Using Enhanced Multi-Object Tracking

Group detection is a critical yet challenging task in video-based applications such as surveillance analysis, especially in crowded and dynamic environments where complex pedestrian interactions occur. Traditional trajectory-based methods often struggle with occlusions and overlapping behaviors, leading to inaccurate group identification. To address these limitations, we propose a novel algorithm that integrates an optimized YOLOv8 model with DeepSORT tracking, enhancing both detection accuracy and real time performance. Our approach uniquely combines high-precision object detection with stable multi-object tracking, ensuring consistent identification of individuals and groups over time, even in high-density scenarios. Additionally, we introduce an innovative method of constructing an adjacency matrix by integrating Euclidean distances and bounding box diagonal ratios, which is transformed into a graph to intricately analyze and predict complex group dynamics in real time. Experimental results on real-world airport CCTV footage demonstrate that our method significantly outperforms existing approaches, achieving higher precision and recall rates. Furthermore, the algorithm operates efficiently on standard hardware, indicating strong practical feasibility for real-time applications in public spaces. While challenges such as misclassification due to incomplete data annotations and occlusions remain, our study showcases the potential of integrating spatial and temporal data to advance real-time group detection and tracking, aiming to improve crowd management systems in public spaces.