Fine-Tuning YOLO for Pedestrian Detection in Webcam Images to Analyse Mobility Patterns

Kurzfassung

Object detection models based on convolutional neural networks have advanced considerably in recent years and are increasingly used in a wide range of applications. In the context of mobility research, they can be used to analyse footage of open webcam data and therefore serve as an additional data source. This thesis evaluates the performance of fine-tuned YOLO models on open webcam images from 41 locations across Germany for the detection of pedestrians. A manually annotated dataset of 820 images was used to train and compare several model configurations, varying in model size and image size, and detection performance was further assessed across a range of confidence thresholds. The best result was achieved by YOLO26s at an image size of 1024, reaching a mean Average Precision (mAP50) of 0.578 and a maximum F1 Score of 0.639. In a subsequent application, the best-performing model was used to analyse open webcam footage from selected urban locations to derive temporal and spatial patterns of pedestrian activity. The results show clear differences between locations and time periods, demonstrating the potential of YOLO-based pedestrian detection with open webcam data for mobility monitoring in public spaces.