<p>The rapid rise in surveillance technologies and the increasing number of cameras in open areas have heightened the need for an efficient way to identify individuals and recognise groups in congested areas. The present research proposes a new version of YOLOv5, called Crowd Scale YOLO, which is specifically designed to identify pedestrian classes in crowded scenes. We deal with the inherent problem of extremely large scale variation, extreme occlusion, and challenging boundary definitions by combining three essential architectural elements (1) CSPNet backbone to reduce computational cost and improve feature extraction, (2) Bidirectional Feature Pyramid Network (BiFPN) to comprehensively aggregates feature on a large scale, and (3) Adaptive Spatial Feature Fusion (ASFF) to reduce the issue of feature inconsistency when comparing the extremely small and large pedestrians. The suggested architecture provides a fine-grained classification of the pedestrians into the categories of individuals, groups, and crowds- this is essential in intelligent surveillance systems that require behavioral analysis and crowd control. Extensive CrowdHuman and Mall dataset testing shows mAP of 0.5 at 94.13% and 92.25%, respectively, with high recall (99.37%) on the crowd class, indicating better performance under heavy occlusion. Extensive ablation experiments validate the contribution of each element, with the full architecture producing 3.37%, 5.76%, and 2.39% improvements over baseline settings on crowd, group, and single classes, respectively. With about 55 FPS, Crowd Scale YOLO offers a state-of-the-art trade-off between accuracy and speed, able to operate in a real-time surveillance system and outperforming both base YOLOv5 (89.6% mAP) and recent improved versions (93.7% mAP) with a smaller computational footprint.</p>