The multi-scale place recognition based on various coverage residual fingerprint for indoor scenarios

The place recognition is a crucial localization step for indoor robot. It is challenging to satisfy the real-time requirements of place recognition especially when dealing with large-scale maps. This paper proposes a multi-scale place recognition based on VCRF (Various Coverage Residual Fingerprint). Firstly, we construct VCRF solely from front-view images and comprise two components: FRF (Fusion Residual Feature) for current-location representation and DAF (District Aggregate Feature) for district area representation. Specifically, the FRF is obtained by integrating the top three residual features extracted from a ResNet using logistic regression, thereby enhancing feature discriminability. Then the DAF is constructed by averaging FRFs within each district to represent district area. Secondly, we propose a multi-scale place recognition utilizing VCRF to implement place recognition. It employs coarse-to-fine strategy involving DAF based coarse and Bayesian based fine localization. Thirdly, in Bayesian based fine localization, robot velocity is fused with FRF within a Bayesian model to obtain accurate localization. Finally, the proposed method was evaluated in indoor scenarios under various conditions. The F1-measure of proposed method achieved 0.9947 and 0.9983 for illumination invariant and illumination changing conditions, respectively. We compared the proposed method with several the-stated-of-art methods. The experimental results demonstrated that the proposed method achieves promising comprehensive performance.