NeuroLoc: Encoding Navigation Cells for 6-DOF Camera Localization

TL;DR

NeuroLoc tackles robust 6-DOF camera localization from a single image by emulating biological navigation systems. It integrates a Visual Encoder, Hebbian Storage Module for historical scene refinement, and a Pose Regression Module that employs Biologically Plausible Direction Attention and a 3D Grid Module to predict both camera pose and grid-center positions. The approach achieves state-of-the-art performance among single-image methods on indoor and outdoor benchmarks, with significant gains in textureless and dynamic environments. This brain-inspired framework enhances robustness and accuracy, offering practical benefits for autonomous navigation in challenging real-world settings.

Abstract

Recently, camera localization has been widely adopted in autonomous robotic navigation due to its efficiency and convenience. However, autonomous navigation in unknown environments often suffers from scene ambiguity, environmental disturbances, and dynamic object transformation in camera localization. To address this problem, inspired by the biological brain navigation mechanism (such as grid cells, place cells, and head direction cells), we propose a novel neurobiological camera location method, namely NeuroLoc. Firstly, we designed a Hebbian learning module driven by place cells to save and replay historical information, aiming to restore the details of historical representations and solve the issue of scene fuzziness. Secondly, we utilized the head direction cell-inspired internal direction learning as multi-head attention embedding to help restore the true orientation in similar scenes. Finally, we added a 3D grid center prediction in the pose regression module to reduce the final wrong prediction. We evaluate the proposed NeuroLoc on commonly used benchmark indoor and outdoor datasets. The experimental results show that our NeuroLoc can enhance the robustness in complex environments and improve the performance of pose regression by using only a single image.

Figures (5)

• Figure 1: An overview of the proposed NeuroLoc framework. It includes a visual encoder (extracting scene features from a single image), a Hebbian Storage Module (storing and reading scene information), a pose regression module (directional attention is used to map attention features to camera poses, and a 3D grid module is used to predict grid center positions).
• Figure 2: Overview of the Hebbian storage module. The input features will be expanded into index vectors, context vectors, and inactivate vectors, and then the storage matrix will be updated using Hebbian rules (persistent storage of scene features). The inactive vector is multiplied by the Hebbian matrix to obtain the activated vector.
• Figure 3: The left image shows the activation status inside the feature after embedding direction encoding. The image on the right shows the true direction in the real world.
• Figure 4: The left figure shows that we constructed $n$ 3D grids in 3D space, where $g_x$, $g_y$, and $g_h$ represent the grid boundaries (map boundaries) on the three coordinate axes, and the red dots represent the center positions of the 3D grids. The figure on the right shows that the center of our 3D mesh is directly calculated in the real world.
• Figure 5: Saliency maps of two scenes selected from Oxford RobotCar for straight driving (up) and turning (down).