A Robust Visual Place Recognition Model Based on Cuckoo Search Algorithm and Deep Learning

Abstract

Visual place recognition (VPR) utilizing deep learning (DL)is developed in the domain of computer vision (CV) by connecting the power of neural networks (NNs) for automatically recognizing and finding particular land marks or positions within visual scenes. Leveraging DL model, namely convolutional neural networks (CNNs) and recurrent neural networks (RNNs), this technology has achieved important developments in tasks like object localization autonomous navigation, and scene understanding. Consequently, this study presents a Cuckoo Search Algorithm with Deep Learning Driven Robust Visual Place Recognition (CSADL- VPR) method. This introduces CSADL-VPR approach for VPR that leverages Gaussian filtering (GF) based pre-processing, Cuckoo Search optimization algorithm (CSA), MobileNetV2 feature extractor, and Manhattan distance-based similarity measurement. The MobileNetV2 feature extractor could be exploited for extracting compact and distinctive features from input images. For enhancing the performance of the place recognition model, the CSA has been implemented for parameter tuning and feature selection (FS). The Manhattan distance metric is employed for evaluating the similarity among feature vectors extracted from query images and those in the reference datasets. A series of simulated analyses can be executed to ensure the enhanced performance of the CSADL-VPR approach. The experimental outcomes represented the improvements of the CSADL-VPR method on the place recognition task.