Implementation of The YOLOv5 Model For Crowd Detection : Model Analysis Model Performance Analysis

Abstract

One of several deep learning object detection model ideas is YOLO. High-level characteristics obtained by Convolutional Neural Networks (CNN) can be used to overcome the challenges of prior assumptions. Because of the inherent complexity of the human crowd, prominent places in a high-density crowd may exhibit characteristics that differ from those in a normal density crowd. The most efficient deep learning model architecture for use in real-time video applications is provided by YOLO. The CrowdHuman dataset is annotated by researchers using the Roboflow platform. The outcomes are annotated files in the YOLOv5 format including around 4,500 pictures. Photographs of crowds at various elevations and distances are included in this image dataset. The Roboflow platform is used by researchers to convert this annotation and get the required results. This research is divided into four stages: the evolving process, the training process, the testing process, and the evaluation process. The training results are quite positive because they have a mAP value larger than 0.4. However, during the evaluation phase, the model continues to have issues detecting small and moving objects. The device used is a YOLOv5s, a smaller version of the YOLOv5.