DRIFT-Net: Drive-State-Aware Weighted Interpolation for Error Forecasting and Trajectory Correction Network

Abstract

Accurate vehicle localization is essential for autonomous driving. However, vehicle position estimation becomes challenging when localization information from sensors such as the Global Navigation Satellite System (GNSS) and Light Detection and Ranging (LiDAR) is unavailable, degraded, or unreliable. In such situations, dead reckoning can provide continuous localization, but accumulated yaw estimation errors often result in significant position error. To address this issue, this paper presents DRIFT-Net, a DRive-state-aware Weighted Interpolation for error Forecasting and Trajectory correction Network, to improve dead-reckoning-based vehicle localization. DRIFT-Net employs a Convolutional Neural Network–Long Short-Term Memory (CNN–LSTM) architecture to extract driving features and incorporates an expert layer that performs state-based weighted interpolation of expert predictions to estimate yaw error. A Kalman filter is subsequently applied to suppress noise and refine the final vehicle position estimate through smoothing. Experimental results in low-speed and high-speed driving scenarios demonstrate that the proposed approach improves localization accuracy. The proposed method achieved yaw Root Mean Square Error (RMSE) values of 0.339 deg and 0.148 deg in the respective scenarios, confirming improved performance compared with existing methods.