In the context of intelligent transportation systems and applications such as
autonomous driving, it is essential to predict a vehicle’s immediate future
states to enable precise and timely prediction of vehicles’ movements. This
article proposes a hybrid short-term kinematic vehicle prediction framework that
integrates a novel object detection model, You Only Look Once version 11
(YOLOv11), with an unscented Kalman filter (UKF), a reliable state estimation
technique. This study provides a unique method for real-time detection of
vehicles in traffic scenes, tracking and predicting their short-term kinematics.
Locating the vehicle accurately and classifying it in a range of dynamic
scenarios is achievable by the enhanced detection capabilities of YOLOv11. These
detections are used as inputs by the UKF to estimate and predict the future
positions of the vehicles while considering measurement noise and dynamic model
errors. The focus of this work is on individual vehicle motion prediction using
short-horizon kinematic cues. The publicly employable Lyft Level 5 dataset has
been used to validate the proposed method, indicating its efficacy in attaining
high prediction accuracy with low latency. The experimental results illustrate
that the accuracy, precision, root mean square error (RMSE), and mean absolute
error (MAE) are improved by 4.1%, 2.66%, 11.9%, and 13.3%, respectively, when
the performance of the enhanced algorithm is compared to that of the YOLOv11
combined with extended Kalman filter (EKF) algorithm. Integrating YOLOv11 with
the UKF leads to enhanced responsiveness and reliability of vehicle trajectory
predictions, which is profitable for autonomous vehicles and advanced
driver-assistance systems.
