Optimasi Deteksi Objek pada Video dengan Kompresi Region of Interest menggunakan Model YOLOv8

The demand for real-time object detection systems, such as those used in video surveillance and autonomous vehicles, drives the need for efficient data storage and transmission without compromising accuracy. One promising approach is Region of Interest (ROI)-based video compression, which preserves visual quality in important areas. This study aims to evaluate the impact of video compression on object detection accuracy using the YOLOv8 model through statistical analysis using Analysis of Variance (ANOVA), and to compare the effectiveness of uniform and ROI-based compression methods. Videos from the VIRAT Video Dataset were compressed using the Constant Rate Factor (CRF) parameter and evaluated based on mAP_50, mAP_50_95, and file size. ANOVA results indicate no statistically significant differences between the two methods. At CRF 50, file size can be reduced by over 60%, but mAP_50 accuracy drops below 50% due to quality degradation in non-ROI areas, which disrupts the spatial context required by the model. This study contributes by examining the compression tolerance limits of YOLOv8 and reveals that overall visual quality, rather than just object-focused quality, plays a crucial role in model performance. These findings have important implications for real-time applications such as CCTV and autonomous vehicles, where a balance between compression efficiency and detection accuracy is critical. Future studies may explore adaptive ROI approaches that consider dynamic object movement.