DCA-SSeqNet: Enhanced IMU Dead Reckoning via Dilated Channel-Attention

This study proposes DCA-SSeqNet (Dilated Channel-Attention Short Sequence Network), an improved lightweight deep learning-based IMU navigation method to address error accumulation and computational complexity in inertial navigation. Compared with Basic-CNN and traditional dead reckoning, DCA-SSeqNet combines CNN’s sequential feature extraction with dynamic channel attention weighting, enhancing IMU data modeling and optimizing the Kalman filter’s pseudo-measurement noise covariance. Its lightweight design reduces computational overhead, making it suitable for embedded systems. Experiments on the KITTI dataset and in self-built environments demonstrate that DCA-SSeqNet effectively suppresses IMU error accumulation, improves short-sequence prediction accuracy, and reduces Absolute Trajectory Error (ATE) by 30.4% compared to Basic-CNN, verifying its superior performance and generalization in complex motion scenarios.