Trainable Hypervectors Encoding for Efficient 3D Loop-Closure Detection on Edge Devices

Abstract

Loop-closure detection plays a critical role in simultaneous localization and mapping (SLAM) systems. The primary task of loop-closure detection involves analyzing previously visited locations and correcting mapping errors, which typically stem from intrinsic noise in sensor data and accumulate over time. However, the burden of storing and querying/searching for previously visited information continues to increase with time. Consequently, reducing the amount of data stored becomes increasingly important. In this study, we propose a trainable hypervectors (THV) encoder, integrating quantization and a lookup table (LUT) to significantly enhance execution speed. Additionally, we employ a triangular mask in second-order pooling (SOP) for filtering extraneous features in the encoder and introduce binary quadruplet loss to efficiently train binary feature representations. We evaluate our method extensively on the KITTI, MulRan and Wild-Places datasets. The experiments demonstrate that our method substantially improves efficiency while maintaining accuracy. Moreover, our method effectively utilizes the 3D-NAND flash in-memory computing technique to improve execution performance.