Ev-ReconNet: Visual Place Recognition using Event Camera with Spiking Neural Networks

In this paper, we utilize the advantages of an event camera to tackle visual place recognition (VPR) problem. The camera’s high measurement rate, low latency, and dynamic range make it well-suited overcome limitations conventional vision sensors. However, apply existing convolutional neural networks (CNNs) based algorithms such as NetVLAD, asynchronous stream should be converted a synchronous image frame, which causes loss in temporal information. To address problem, paper proposes method that employs characteristic spiking (SNNs) leverage nature streams. is images tensors our preprocessing module. SNNs-based reconstruction networks, are from CNNs, reconstruct edge regardless external environment changes. Visual conducted by matching features database those used feature extraction network study. evaluate performance VPR comparing previous methods for DDD17 Brisbane-Event-VPR dataset. Experimental results demonstrate accuracy proposed better than methods, especially datasets with adverse weather conditions. We also verify energy efficiency improved SNNs over CNNs. Our code available download on https://github.com/AIRLABkhu/EvReconNet.