Detection, classification, and trajectory prediction of vehicles and barriers using deep learning-based models and non-linear curve fitting via car-dash camera

Abstract

Road accidents remain a major concern in transportation systems, especially in areas where unsuitable road design, limited signage, and poorly positioned barriers increase collision risks. This study presents a framework for vehicle and barrier detection, classification, and trajectory prediction using deep learning-based models and a non-linear curve fitting model via car dash camera video streams. A customized dataset including the public COCO dataset and a privately collected barrier dataset was used to enhance detection diversity. This dataset was divided it into seven classes including person, bicycle, car, motorcycle, bus, truck, and barrier. To detect, classify, and tracking objects, the You Only Look Once version 8 (YOLOv8) model is applied. After that, the non-linear curve fitting model, implemented through Support Vector Regression (SVR) or Regressor Chain (RC), predicts future object trajectories. Model performance is evaluated using Mean Squared Error (MSE), Mean Absolute Error (MAE), and Root Mean Squared Error (RMSE), where lower values indicate higher prediction accuracy. The integration of YOLOv8 and non-linear curve fitting model enhances the ability to support collision risk assessment and promotes safer autonomous driving systems in complex road environments.