Traffic Analysis Framework: Vehicle Counting and Traffic Volume Forecasting

This project examines the usage of recent machine learning techniques to forecast traffic flow using vehicle counts produced by an object detection system. By combining a recently developed object detection model, YOLOv8 with BYTETrack were used for vehicle counting and tracking. The forecasting phase utilised various models trained on processed data, including XGBoost, LSTM, GRU, and ConvLSTM. The XGBoost model with PCA dimensionality reduction performed better across training and validation sets. The project's main objective is to use the count produced from the YOLOv8 as an input to the forecasting model to provide the vehicle count for the next hour. Using data analysis, the project addressed research questions, which covered insights about the impact of weather, time and the number of lanes on traffic flow. The findings confirm that time components, like hour and day of week, affect traffic, while weather surprisingly has minimal impact. Moreover, roads with more lanes were found to handle more vehicle counts within the hour, which shows the importance of infrastructure planning. The project also faced multiple challenges, such as data processing limitations and overfitting, which confirms further improvements can be applied, such as using cloud platforms to bypass the hardware limitations so that processing the whole dataset into the deep learning models can be possible. On the other hand, even with these challenges, the project’s results are satisfactory for traffic forecasting, which provides numerous recommendations to reduce congestion. The project also implemented a web interface to interact with the entire system and present the findings from the dataset.