People Counting in Sample Video Footage Using CNN Integrated with YOLOv5

Accurate people counting in dynamic environments remains challenging due to variations in lighting, complex backgrounds, and occlusion. This study proposes a video-based people counting system leveraging a Convolutional Neural Network (CNN) integrated with the YOLOv5 object detection model. The system applies a structured preprocessing pipeline, including frame extraction, normalization, and noise reduction, to enhance data consistency before detection. The model was evaluated using ten real-world campus video sequences to assess detection reliability and counting accuracy. Experimental results demonstrate that the proposed method achieves high precision and recall for real-time detection across diverse scenarios. Performance degradation was observed in frames containing dense crowds or low illumination, indicating limitations under extreme conditions. These findings validate the feasibility of lightweight CNN-based detectors for surveillance and monitoring applications, while highlighting the need for larger datasets and optimized training strategies to improve robustness in more complex environments.