Aggregating Multiple Bio-Inspired Image Region Classifiers For Effective And Lightweight Visual Place Recognition (2312.12995v1)

Abstract: Visual place recognition (VPR) enables autonomous systems to localize themselves within an environment using image information. While VPR techniques built upon a Convolutional Neural Network (CNN) backbone dominate state-of-the-art VPR performance, their high computational requirements make them unsuitable for platforms equipped with low-end hardware. Recently, a lightweight VPR system based on multiple bio-inspired classifiers, dubbed DrosoNets, has been proposed, achieving great computational efficiency at the cost of reduced absolute place retrieval performance. In this work, we propose a novel multi-DrosoNet localization system, dubbed RegionDrosoNet, with significantly improved VPR performance, while preserving a low-computational profile. Our approach relies on specializing distinct groups of DrosoNets on differently sliced partitions of the original image, increasing extrinsic model differentiation. Furthermore, we introduce a novel voting module to combine the outputs of all DrosoNets into the final place prediction which considers multiple top refence candidates from each DrosoNet. RegionDrosoNet outperforms other lightweight VPR techniques when dealing with both appearance changes and viewpoint variations. Moreover, it competes with computationally expensive methods on some benchmark datasets at a small fraction of their online inference time.

• The paper introduces RegionDrosoNet, a bio-inspired VPR system that aggregates specialized DrosoNets to enhance place recognition on resource-constrained robots.
• The method divides images into non-overlapping regions and applies grid-specific DrosoNets, improving recognition across seasonal and viewpoint variations.
• Benchmark tests show that RegionDrosoNet achieves real-time performance comparable to heavier CNN-based systems while maintaining high efficiency.

Introduction

Visual place recognition (VPR) has become an essential technology for mobile robotics, allowing robots to determine their location in an environment based solely on visual inputs, typically camera images. Traditional VPR systems often rely on Convolutional Neural Networks (CNNs), but such models have high computational demands, which can be impractical for robots equipped with limited hardware capabilities.

Lightweight Approach to VPR

A novel VPR system known as RegionDrosoNet offers a balance between computational efficiency and VPR performance. Inspired by biological systems, specifically the fruit fly's olfactory circuitry, the system introduces a group of compact neural network models called DrosoNets. These models are trained to specialize in recognizing different subdivisions of an image, thereby diversifying their ability to match locations amid varying appearances and perspectives.

System Design and Functionality

RegionDrosoNet divides input images into multiple non-overlapping regions using different grid configurations, ensuring heterogeneous coverage. Each region feeds into a specialized group of DrosoNets, trained exclusively on that section of the image during the training phase. During operation, these DrosoNets generate multiple place confidence scores, which are then synthesized using a novel voting module that considers the top reference candidates from each. This approach allows RegionDrosoNet to reach consensus on the most accurate location match, increasing the reliability of VPR even when individual DrosoNets fail to recognize the correct place.

Performance and Evaluation

RegionDrosoNet's performance was evaluated using several benchmark datasets that test against seasonal variation, viewpoint changes, and different weather conditions. The system demonstrates a remarkable ability to outperform other lightweight VPR techniques, as well as bear comparison with more computationally demanding methods, all while maintaining a rapid inference time suitable for real-time applications.

Implications and Future Work

The implications of RegionDrosoNet's design are profound for the development of autonomous systems. By offering a VPR system that doesn't compromise on speed or accuracy, robots can potentially operate in various environments for extended periods without the need for high-end computational hardware. Future research directions could explore eliminating the need for environment-specific training, broadening the system's applicability across diverse deployment scenarios.