Detecting and Tracking of Individuals in Crowded Environments

This research tackles the challenge of detecting and tracking individuals in complex environments using computer vision techniques. The analysis of individuals in populated environments is essential for managing spaces and ensuring safety in large institutions such as universities. In this work, we propose developing a pedestrian detection method based on the exploration of three pre-trained versions of YOLO models to maximize the detection of individuals in crowded scenarios. The proposed method was trained and validated on a private, purpose-built dataset that reflects real-world university surveillance scenarios with both low and high densities. The best-selected model was based on the YOLOv10 model with a confidence threshold of 0.5, yielding a mean mAP50 score of 0.9132, precision of 0.9131, and recall of 0.8644 using a ten-fold cross-validation strategy during the training stage. This result was statistically similar to other models regarding detection performance, but lower in algorithm complexity, highlighting that the simplicity and straightforward architecture of the YOLOv10 model did not detrimentally affect its localization capabilities. In the test stage, using a hold-out set, the proposed method achieved a higher mAP50 of 0.9267, a precision of 0.9218, and a recall of 0.8844. These results indicated its successful learning capability without incurring overfitting and a robust performance in detecting individuals under varied campus conditions, highlighting the effectiveness of the YOLOv10 architecture as a foundation for person detection in academic environments and supporting its practical suitability for accurate and real-time pedestrian monitoring.