Análisis de rendimiento de you only look once, retinanet y single shot detector aplicado a la detección y conteo vehicular

Abstract

Artificial intelligence has experienced significant growth in recent decades. As a result, several architectures have been developed for object detection, classification, and recognition. Currently, there are several alternatives that fulfill these purposes; however, there is no rigid framework defining how these architectures are formed. This work presents an updated performance analysis of real-time vehicle detection and counting using You Only Look Once (YOLOv8) version 8, RetinaNet (RN), and Single Shot Detector (SSD). For such analysis, the Google Colaboratory was used as the main re-training environment. The Research-Action methodology was employed to develop the practical case, and a systematic literature review was also con-ducted to determine the state of the art in this problem domain. For feature extraction, RESNET-50 and MobileNet were used in RN and SSD, respec-tively. The results indicated that YOLOv8 (which has undergone the most adjustments since its inception) exhibits the best performance in terms of detection time, precision considering the frames that need to be analyzed to enable real-time usage, and ease of implementation.