Muis Muhtadi, Achmad Hamdan, Hakkun Elmunsyah, Azhryl Akbar Assagaf
Real-time object detection in applications such as video surveillance and autonomous systems requires efficient video compression techniques that minimize storage requirements while preserving detection accuracy. This study evaluates the impact of video compression on object detection performance using the YOLOv8 model, with a comparative analysis of uniform compression and Region of Interest (ROI)-based compression. Video samples from the VIRAT dataset were encoded using various Constant Rate Factor (CRF) settings and assessed in terms of detection accuracy, measured by mean Average Precision over multiple Intersection over Union (IoU) thresholds (mAP_50_95), and compression efficiency, measured by the Compression Ratio (CR). Experimental results demonstrate that uniform compression consistently achieves higher detection accuracy and greater compression efficiency compared to ROI-based compression across all CRF values. ANOVA statistical analysis yielded a p-value =1. 8 7 4 × 1 0-8, confirming that these differences are significant. These findings indicate that, contrary to initial expectations, uniform compression provides superior performance in both accuracy and compression. Future work will explore adaptive ROI strategies that account for dynamic object behavior to improve the balance between compression efficiency and detection fidelity. © 2025 IEEE.
State University of Malang, Electrical Engineering and Informatics, Malang, Indonesia; State University of Malang, Electronic Systems Engineering Technology, Malang, Indonesia
