LPMP: A Bio-Inspired Model for Visual Localization in Challenging Environments.

Area of Science:

• Computer Vision
• Robotics
• Neuroscience

Background:

• Autonomous vehicles require robust self-localization in dynamic environments.
• Visual Place Recognition (VPR) is crucial for place identification using visual data despite appearance changes.
• Existing VPR models like NetVLAD and CoHog serve as benchmarks.

Purpose of the Study:

• To introduce and evaluate the Log-Polar Max-Pi (LPMP) model, a novel neuro-cybernetic approach for VPR.
• To compare the LPMP model's performance against state-of-the-art VPR methods.
• To propose a new benchmark for VPR evaluation and assess the impact of environmental paving variations.

Main Methods:

• Developed the bio-inspired Log-Polar Max-Pi (LPMP) neural network model.
• Implemented unsupervised one-shot learning for environmental representation.
• Utilized distinct 'what' and 'where' pathways for processing visual information and generating visuospatial codes.
• Compared LPMP with NetVLAD and CoHog on the Oxford car dataset, considering different paving conditions.

Main Results:

• The LPMP model achieved localization performance comparable to or exceeding that of NetVLAD and CoHog.
• A new benchmark was established for evaluating VPR models across diverse environments.
• The study analyzed the impact of uneven paving on localization accuracy.

Conclusions:

• The neuro-cybernetic LPMP model offers a promising approach for visual place recognition in autonomous driving.
• The proposed benchmark facilitates more comprehensive VPR model evaluation.
• LPMP demonstrates robustness in challenging environmental conditions, including uneven paving.