BYTECounter: Improving Vehicle Turning-Movement Counting

This paper proposes a methodology to enhance the accuracy of vehicle turning-movement counting in video footage. The methodology uses a three-component system comprising object detection, object tracking, and object turning-movement counting. To improve vehicle detection, a custom vehicle neural network called YOLOv7-vehicles-0.1 was trained using a unique vehicle dataset comprising COCO, KITTI, and OpenImages datasets. Additionally, a novel tracker named BYTECounter was developed, using two virtual drawing objects, Regions and Lines, to count vehicles' turning movements based on trajectory information. The performance of our system was evaluated using the UA-DETRAC dataset, achieving an overall Movement Counting Accuracy (MCAcc) of 93.5% with a processing rate of 82.86 frames per second. This paper demonstrates that BYTECounter outperforms the state-of-the-art tracker, BYTE, and highlights the superiority, in terms of counting metrics, of YOLOv7-vehicles-0.1 over YOLOv7 models trained on the COCO dataset.