Motorcycle violations at zebra crossings remain common due to limitations in manual traffic enforcement, highlighting the need for automated detection systems based on computer vision. This study aims to analyze the effect of image resolution variations on YOLO11n performance in detecting motorcycles and zebra crossing violations by separately evaluating 3 outcomes, vehicle detection, violation detection, and computational time. The dataset comprised 1,069 images collected at the flyover Sekip traffic intersection in Palembang, Indonesia, with 70% allocated for training and 30% for testing. YOLO11n models were trained using 50 to 250 epochs and evaluated on test videos at 5 resolutions: 240p, 360p, 480p, 720p, and 1080p. At 150 epochs, the model achieved a precision of 0.914, recall of 0.971, and f1-score of 0.941 based on the confusion matrix, along with an mAP50 of 0.987 and mAP50-95 of 0.886. In test videos containing 128 motorcycles according to ground truth, the vehicle detection rate increased from 14,06% at 240p to 85,15% at 1080p, while computational time increased from 123,9 to 332,4 seconds. Based on predefined criteria requiring a detection rate of at least 80% with the lowest computational time, 720p was selected, detecting 105 of 128 motorcycles (82,03%) in 221,9 seconds. These findings are limited to one location, daytime conditions, and a single camera, restricting generalizability.
Copyrights © 2026