Guivant, J.; Narula, K.; Kim, J.; Li, X.; Khan, S. Compressed Gaussian Estimation under Low Precision Numerical Representation. Sensors2023, 23, 6406.
Guivant, J.; Narula, K.; Kim, J.; Li, X.; Khan, S. Compressed Gaussian Estimation under Low Precision Numerical Representation. Sensors 2023, 23, 6406.
Guivant, J.; Narula, K.; Kim, J.; Li, X.; Khan, S. Compressed Gaussian Estimation under Low Precision Numerical Representation. Sensors2023, 23, 6406.
Guivant, J.; Narula, K.; Kim, J.; Li, X.; Khan, S. Compressed Gaussian Estimation under Low Precision Numerical Representation. Sensors 2023, 23, 6406.
Abstract
This paper introduces a method to achieve optimal Gaussian estimation even when operating with low precision numerical representations of the full covariance matrix, such as 16-bit integer or 32-bit single precision formats, instead of more expensive double precision floating point representations (64 bits). The approximation's numerical error is accurately estimated in a conservative manner, resulting in minimal covariance inflation. When combined with a compressed estimator like the GCKF (Global Compressed Kalman Filter), which performs global updates at low frequency, this approach produces estimates that are nearly identical to the optimal ones. The deviation between the results obtained using this new approach and those obtained with the standard GCKF (which operates in double precision) is negligible. This new method is particularly relevant in high-dimensional estimation problems that require high-frequency operation. By leveraging the compressed operation of the GCKF, the estimation process applies the necessary approximation only during low-frequency global updates, using the single precision format instead of the more expensive double precision representations. The introduced numerical approximation is accurately estimated while maintaining a conservative approach, resulting in minimal covariance inflation. When combined with the GCKF, which performs global updates at low frequency, the estimates produced are almost identical to the optimal ones. This approach is particularly useful in high-dimensional estimation problems that require high-frequency operation. By representing the full covariance matrix in a low precision numerical format and applying appropriate scaling, memory usage is significantly reduced, offering substantial benefits for high-dimensional estimation problems and enabling real-time smoothing processes in such cases. By reducing the memory requirements for storing the full covariance matrix, the CPU avoids encountering a high number of page faults, resulting in improved processing performance. Lastly, the experimental section verifies the performance of the proposed approach in a single precision (32-bit floating point) GCKF SLAM process, assessing expected values and standard deviations of marginal PDFs.
Engineering, Electrical and Electronic Engineering
Copyright:
This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
