Implementasi Model Yolov8 Dan SSD Untuk Sistem Penghitungan Kendaraan Pada Video Cctv Roro Bengkalis

Operational management at the RORO Bengkalis Crossing Port currently relies on manual monitoring, which is prone to human error and inefficient in providing real-time data. This research aims to design an automated vehicle counting system based on Computer Vision by integrating YOLOv8 and SSD (Single Shot Multibox Detector) Deep Learning models to address visual challenges such as occlusion and lighting variations. The system development follows the CRISP-DM methodology and implements a custom ensemble method named Confidence Aware Ensemble (CAE), which combines predictions from both models using coordinate weighting and cross-validation. The system is also equipped with the Custom IoU Tracker tracking algorithm and virtual line logic to accurately count Group II, IV, and V vehicles. Test results show that the ensemble method (CAE) successfully improved system sensitivity with an average Recall value of 88.7% during daylight operational hours, higher than the single YOLOv8 model (87.8%) and SSD (78.5%). While detection performance decreases under extreme visual conditions such as rain or night glare, this mechanism proved effective in recovering object detections missed by the primary model, making this system a reliable solution to support queue management efficiency at the port.