Comparison of CNN models in vehicle counting for smart traffic lights

Continuous urban growth increases road congestion, which has a negative effect on the population. One way to reduce congestion is an effective control traffic light timing. Congestion is sometimes increased by the fixed timing of traditional traffic lights. Smart traffic lights adapt their timing based on the number of vehicles waiting at the intersection. Smart traffic lights with artificial vision don’t require complex technological infrastructure. Machine vision systems use convolutional neural networks (CNN) like YOLO or Faster RCNN for object detection in vehicle counting because of their adaptability to different scenarios. The aim of the present research is to determine which CNN is the most suitable for frame-based vehicle counting at traffic signalized intersections. This was achieved by evaluating the performance of YOLOv8M and Faster RCNN Resnet50 in 147 test scenarios. The performance metrics are the weighted average count error and the CNNs’ inference time. An ANOVA test with an alpha level of 0.05 was used to obtain the results, which indicate that the best CNN model in daytime conditions is YOLOv8M and that in nighttime conditions it is Faster RCNN Resnet50.