LoCS-Net: Localizing convolutional spiking neural network for fast visual place recognition

Area of Science:

• Robotics and Artificial Intelligence
• Neuromorphic Engineering
• Computer Vision

Background:

• Visual place recognition (VPR) is crucial for robot navigation but faces challenges like perceptual aliasing and dynamic scenes.
• Current artificial neural network (ANN) based VPR methods are computationally inefficient.
• Spiking neural networks (SNNs) offer potential computational efficiency but face training and real-time performance issues.

Purpose of the Study:

• To develop an efficient and tractable end-to-end convolutional SNN model for VPR.
• To improve the training and inference performance of SNNs for VPR tasks.
• To demonstrate the real-world deployment capabilities of SNNs for VPR on neuromorphic hardware.

Main Methods:

• Developed an end-to-end convolutional SNN model for VPR using backpropagation for training.
• Employed rate-based approximations of leaky integrate-and-fire (LIF) neurons during training, switching to spiking LIF neurons for inference.
• Utilized an ANN-to-SNN conversion strategy for on-chip deployment on neuromorphic hardware (Intel Kapoho Bay).

Main Results:

• Achieved superior performance on Nordland (78.6% precision at 100% recall) and Oxford RobotCar (45.7%) datasets compared to SOTA SNNs.
• Demonstrated significant improvements in training and inference times.
• On-chip performance on neuromorphic hardware showed continued outperformance over SNN counterparts, with notable energy efficiency.

Conclusions:

• The proposed SNN model offers a simpler training pipeline and enhanced performance for VPR.
• The approach facilitates rapid prototyping and real-world deployment of SNN-based VPR solutions.
• This work represents a significant advancement towards prevalent SNN-based robotics solutions.