<p>This study addresses the challenge of the limited availability of high-quality Turning Movement Count (TMC) data, which is essential for intersection analysis and traffic signal design. Due to the importance of TMC prediction, we utilize machine learning methods such as variational autoencoder (VAE), generative adversarial network (GAN), VAE-GAN hybrid, and long short-term memory (LSTM) to generate a synthetic dataset and address the issue. The synthetic dataset is enriched from the original dataset by introducing realistic and diverse samples to better generalize machine learning models in the TMC prediction task. Firstly, a vision-based tracking system is developed to collect trajectories of vehicles from six intersections in Las Vegas over one hour and estimate TMC data. Secondly, VAE, GAN, VAE-GAN hybrid, and LSTM are utilized to generate a larger synthetic dataset from the limited collected one to learn data distribution and produce new, realistic TMC sequences. The popular machine learning methods are trained on the synthetic data, and the obtained models are fine-tuned on the actual real-world TMC for the prediction comparison with the original models that were only trained on the actual TMC data. The experimental results show that the support vector regression (SVR) and the LSTM methods are superior in the TMC prediction task. Moreover, the synthetic dataset can improve the TMC prediction accuracy by 7% based on the minimum average error. This approach provides a cost-effective solution for transportation planners and engineers to enhance intersection-level traffic analysis and decision-making in data-scarce environments, supporting better signal control strategies and infrastructure planning.</p>