Real-Time Vehicle Detection and Counting Using YOLOv8 and ByteTrack Multi-Object Tracking on Surveillance Cameras

Abstract

Vehicle counting in tourist area parking facilities is typically performed manually, leading to counting errors and the inability to provide real-time capacity updates. This study proposes an automated real-time vehicle detection and counting system integrating YOLOv8s as the object detector and ByteTrack as the Multi-Object Tracking algorithm on surveillance camera footage to support parking capacity management. The dataset consists of 1,377 images across two vehicle classes, cars and motorcycles, prepared through a data-leakage-free pipeline with a 70/20/10 training-validation-test split followed by horizontal flip augmentation applied exclusively to the training subset. The model was trained for 50 epochs on an NVIDIA GeForce RTX 4050 GPU. A virtual counting line positioned at 70% of the frame height, combined with ByteTrack's persistent unique ID mechanism, enables precise vehicle entry and exit counting while preventing double counting. Evaluation results show that the YOLOv8s model achieved precision of 0.904, recall of 0.961, F1-score of 0.930, and mAP@0.5 of 0.969. Vehicle counting evaluation over a 30-minute test video yielded a counting accuracy of 96.875%, MAE of 2.25, and MAPE of 3.125%. The system operated at an average processing speed of 37.82 FPS, exceeding the real-time threshold of 25–30 FPS. These results indicate that the proposed system has the potential to serve as an alternative solution for automated parking capacity management at tourist area facilities.