FFSCore-LSTM: An enhanced LSTM-based camera relocalization networks via front feature smoothing core. (31st March 2023)

• Record Type: Journal Article
• Authors: Wang, Dawei; Liu, Gang; Durga Prasad, Bavirisetti; Xiao, Tiantian; Yang, Yejun
• Is Part Of: Measurement. Volume 210(2023)
• Publication Date: 2023-03-31
• Subjects: Localization -- Camera pose estimation -- Deep learning -- SLAM; Weights and measures -- Periodicals; Measurement -- Periodicals; Measurement; Weights and measures; Periodicals; 530.8
• DOI: 10.1016/j.measurement.2023.112542
• Languages: English
• ISSNs: 0263-2241

Abstract

Highlights: A reversible network to denoise pre-processing image sequences before they are used for training. A modified LSTM for dimensionality deduction to improve the recognition accuracy and reduce training time. Optimization objective function and constraint equation to improve its ability to encode environmental features. Abstract: At present, many traditional camera relocalization methods are being replaced by CNN + LSTM architectures, but CNN + LSTM based camera relocalization methods always contain feature noise when encoding environmental features, which affects the localization performance after structured dimensionality reduction. To solve this problem, we propose a CNN + LSTM architecture called FFSCore-LSTM which contains a front feature smoothing core and a structured dimensionality reduction module. The front feature smoothing core performs feature denoising before structured dimensionality reduction. The risk of overfitting problem of traditional LSTM can be effectively prevented by modified LSTM via reducing the training parameters, and a loss function called PSREJ-Constraint Loss is designed to improve the accuracy and real-time performance of the modified LSTM network. Moreover, a real industrial environment dataset containing indoor and outdoor scenes called CVInd is provided, and the proposed FFSCore-LSTM is validated on this dataset. In addition, the comparison experiments on Cambridge and 7-scenes datasets with the state-of-the-art methods are performed. The experimental results show that FFSCore-LSTM has a better comprehensive performance than other methods in scenes of different scales, including prediction accuracy and stability.