Deteksi Objek Secara Real-Time Berbasis YOLOv8 dan Algoritma DeepSORT

Abstract

Manual monitoring of traffic density and pedestrian movement is often inefficient and prone to errors. This study aims to develop and evaluate an automated real-time object detection and tracking system as a solution to these issues, focusing on four main classes: cars, motorcycles, trucks, and people. The method employed involves the implementation of the YOLOv8l deep learning architecture for high-precision object detection, combined with the DeepSORT algorithm for tracking and re-identification to maintain unique IDs for each observed object. The training data were collected from various CCTV recordings and enriched through augmentation techniques, after which the model was trained using the Google Colab platform. For functional testing, the system was implemented in a local Flask-based web application capable of processing input from videos, webcams, and YouTube live streams. The evaluation results indicate that the model achieved very high and balanced performance, with overall accuracy, precision, recall, and F1-score values reaching 0.98. Although challenges related to Frame Per Second (FPS) stability were identified due to heavy computational loads, the system as a whole proved to be functional, reliable, and effective as an automated solution for object monitoring in public environments.