Local Tetra Pattern-based Bilateral LPP in Robot Navigation Guidance Systems: A Robust Approach against Vision Sensor Noises

Abstract

Locality preserving projections (LPP), originally being a graph Laplacian-based dimensionality reduction technique, searches for a space to embed the higher-dimensional data, keeping the local neighborhood information of the data preserved. However, with conventional LPP, a major challenge lies in the fact that the projection matrix of LPP solely depends on the spatial distribution of the data. This distribution in the Euclidean space is susceptible to spatial factors like scaling, rotation, translation, etc. Apart from that, similar to most graph-based methods, the graph formulation in LPP is influenced by noise and outliers. In this paper, the concept of bilateral filtering has been utilized in the case of LPP, which incorporates feature weights obtained from the higher-dimensional data along with the Euclidean spatial kernel. It enables LPP to make use of the important intrinsic information from the feature space as well, which ameliorates its vulnerability to spatial distortions reasonably. Moreover, a local tetra pattern (LTrP)-based feature descriptor has been utilized here, which helps in extracting robust features from the vision sensor data corrupted with sensor noises. With extensive experimental studies, the proposed algorithm of LTrP-BLPP has been comprehended to be a robust variant of LPP against vision sensor noises.