In this paper, we evaluate the dataset created for the development of traffic volume measurement AI, focusing on whether the expected accuracy is exhibited by examining the dataset distribution. We train the representation of the dataset using a Variational Autoencoder (VAE) and are able to quantitatively demonstrate the distance between dataset distributions by using the Wasserstein distance as a metric. Furthermore, for time periods where the distance between dataset distributions is significant, we propose a self-learning method based on domain adaptation using composite images. As a result, we observe an increase of 2.3% in average precision (AP) and a time period with a 16.1% increase in the match rate with actual traffic volume. The process implemented in this paper, evaluating a custom dataset based on data distribution followed by self-learning of the model, can be considered a useful method when constructing custom datasets and employing AI in practical applications.