Fine-Tuning YOLOv8s for Unified Human and Face Detection in Crowded Environments

Accurate detection of human bodies and faces in densely populated scenes remains challenging due to occlusions and overlapping instances. This paper presents a lightweight object detection solution built upon the You Only Look Once version 8 small (YOLOv8s) architecture, fine-tuned for challenging urban scenes where occlusion, density, and limited computing resources are common. Leveraging an enhanced dataset with detailed person and face annotations, our model achieves a good mean Average Precision at IoU threshold 0.5 (mAP@0.5) of 57.61%, with particularly robust performance on full-body human detection (Average Precision (AP) = 73.5%). Despite moderate face detection accuracy (AP = 42.1%), qualitative results demonstrate solid performance under real-world constraints. The model's compact size and high inference speed make it ideally suited for deployment on edge devices, such as mobile cameras and embedded Artificial Intelligence (AI) systems. A compelling use case is explored through the lens of crowd monitoring in Jamaa El-Fna square in Marrakech, a bustling and high-density public space that demands real-time situational awareness. This work offers a practical tool for urban analytics and public safety, and it lays the foundation for future improvements in face detection, post-processing, and real-time system integration.