What question did this study set out to answer?

To enhance visual odometry performance in autonomous driving by addressing challenges in particle filters.

March 14, 2026

Attention-based Particle Resampling in Sequential Importance Resampling Filters for Enhanced Visual Odometry Performance

Puntos clave

To enhance visual odometry performance in autonomous driving by addressing challenges in particle filters.
Proposed a Particle Transformer framework combining weighted multi-head attention and a variational autoencoder for resampling.
Implemented an adaptive strategy based on effective sample size for balancing accuracy and efficiency.
Conducted experiments on the KITTI odometry dataset.
Achieved 3.0% lower absolute trajectory error compared to systematic resampling.
Obtained 21.2% lower relative pose error on the KITTI dataset.
Maintained high diversity with quick processing time of 6.4 ms per frame.

Resumen

Accurate state estimation is essential for autonomous driving in global navigation satellite system (GNSS)-denied environments. Particle filters, particularly sequential importance resampling (SIR), are widely used for sensor fusion but suffer from weight degeneracy and particle impoverishment during resampling. To address this, we propose Particle Transformer, a deep learning-based resampling framework combining weighted multi-head attention for particle importance modeling and a variational autoencoder (VAE) for diversity preservation. An adaptive strategy based on effective sample size (ESS) further balances accuracy and efficiency. Experiments on the KITTI odometry dataset show 3.0 % lower absolute trajectory error (ATE) and 21.2 % lower relative pose error (RPE) compared to systematic resampling, while maintaining maximum diversity with only 6.4 ms per frame.

Preguntar a la IA

Me gusta

Guardar