This study encompasses the recognition of urban traffic lights by conducting a comparative performance analysis of YOLOv5, YOLOv7, and YOLOv8 R-CNN models. Two datasets were used for this purpose: the Bosch Small Traffic Lights Dataset and BOSCHv2. The former consists of 13,425 images and 24,000 annotations related to traffic lights captured in the areas of Francisco Bay Area and Palo Alto, California, while the latter is a derivative containing annotations exclusively for medium and large-sized traffic lights. The experiments conducted with holdout cross-validation revealed that the YOLOv8 Nano and YOLOv8 Small models achieved the best performances on the BOSCHv2 dataset, with mAP@0.5 metrics of 0.826 and 0.851, respectively. These findings highlight the models' strong performance when applied to the proposed data preparation and contribute to the development of solutions for traffic light detection in Smart Cities.