Comparison and evaluation of YOLO models for vehicle detection on bicycle paths

Abstract

Non-permitted vehicles have taken over bicycle lanes in various Latin American cities as an alternative escape from traffic. Still, they do not foresee the risk to which they expose users of smaller vehicles, such as cyclists. Technological advancement has made researchers use deep learning (DL) to solve various problems in a city's traffic. However, no research has been found focusing on any issue of vehicles allowed or prohibited to travel on a bicycle lane. Therefore, in this article, the you only look once (YOLO) algorithm was used, taking the lightest models from the YOLOv4 to the most recent version, YOLOv8, to detect 05 classes of vehicles that transit or interfere in a bicycle lane, such as bicycles (Bi), motorcycles (Mo), electric motorcycles (ME), electric scooters (SE) and motorcycle cabs (Mt). When testing with the test images, the YOLOv8m model in 50 epochs, using a batch size of 32 and SGD optimizer, was the most optimal, obtaining F1 results with 88.00%, mAP@0:50 of 94.80% and mAP@0.50:0.95 of 76.60%, also had a training time of 1:28h using a Nvidia T4 GPU from Google Colab.