YOLO-based Human Detection and Tracking Application for Contact Tracing using Cameras in the Philippine Fisheries Development Authority Lucena Fish Port Complex

Computer vision had a tremendous growth in recent years and is currently applied in different devices for specific purposes such as autonomous vehicles [1] and video surveillance [2] which both incorporate object detection. One of the object detection techniques is human detection, which has been the subject of various research in deep learning. This study aimed to utilize a human detection application capable of detecting and tracking people who violate social distancing protocol for more than fifteen (15) minutes in the Philippine Fisheries Development Authority - Lucena Fish Port Complex. The study consisted of the following processes: data gathering, building the dataset, configuring the Darknet-53 framework, training of YOLO V3 model, building the GUI, enabling GPU support, integration of the model with the GUI, formulation of distance and tracking algorithms, camera calibration and utilization of both CPU and GPU. The study showed that the model which was trained for 6,000 iterations, using a dataset of 2,038 images and 7,975 annotations, has achieved the highest accuracy results with 95% Precision, 97% Recall, 96% F1-Score, and 98.15% Average Precision. These results indicate that a larger dataset and longer training time are factors in increasing the model’s accuracy. Furthermore, the application attained 12.5 fps in performance when utilizing both CPU and GPU of i7 6th gen. and NVIDIA GTX 1070, respectively. As a result, higher hardware resources are recommended to achieve real-time human detection and tracking using a YOLO V3 model.