Research On Local Image Feature Extraction And Matching In Long-term Visual Localization

Long-term visual localization is a fundamental problem in computer vision and robotics,providing continuous and accurate position and orientation information for cameras in various intelligent systems.Existing long-term visual localization solutions predominantly focus on schemes based on the extraction and matching of local image features,with a core emphasis on the effective and efficient detection and description of these features.However,in complex long-term work environments,factors such as variations in imaging conditions and dynamic obstacles significantly impact the sustained effectiveness of local image feature extraction and matching.Additionally,mobile terminals equipped with long-term visual localization systems typically have limited computational and storage resources,necessitating the design of efficient algorithms,further intensifying the challenges associated with local image feature extraction and matching.Despite numerous studies proposing local image feature extraction and matching schemes from various perspectives,they struggle to simultaneously address the challenges of long-term effectiveness and computational efficiency.Typical local image features include feature points and line segments,both possessing strong complementarity.The extraction and matching of these features usually involve their detection,followed by the generation of corresponding descriptors,ultimately matched through the computation of descriptor similarities.In response to this,this dissertation conducts relevant technical research on the detection and description of image feature points and line segments,subsequently constructing a long-term visual localization system that integrates both point and line features.The specific innovations are outlined below:1.Addressing the long-term effectiveness challenge in image feature point extraction for visual localization,this dissertation introduces a corner and blob detection algorithm based on level line differences.The key insight is that corners,blobs,and non-feature points exhibit distinct patterns in the neighborhood support of level lines and their differences.The proposed algorithm enhances the long-term effectiveness of feature point detection.Furthermore,to address this challenge,a dedicated evaluation benchmark for contour-based corner detection algorithms is developed.The key lies in constructing two contour datasets,utilizing them as inputs to assess various contour-based corner detection algorithms,thereby avoiding the intertwining of contour extraction quality and corner detection performance presented in existing evaluation schemes.2.To overcome the challenges of long-term effectiveness and computational efficiency in image feature point matching for visual localization,this dissertation proposes an illumination-insensitive binary descriptor based on the invariance of local image patch relationships during drastic changes in illumination.The critical finding is that the relative relationships between statistical features of local image patches in the neighborhood of corresponding points often remain invariant under severe illumination changes.The dissertation constructs this descriptor through integral images in multiple feature data channels and spatial granularities,achieving long-term effective and computationally efficient illumination-insensitive feature point matching.3.Confronting the challenges of long-term effectiveness and computational efficiency in image line segment extraction for visual localization,this dissertation designs a line segment detection algorithm based on the dual consistency of edge point coordinates and level lines.The primary insight is that image line segments require dual consistency constraints of edge point coordinates and level lines.By decoupling the direction and position of line segments and employing the level lines and coordinates of edge points separately for fitting,the dissertation enhances the long-term effectiveness and computational efficiency of the line segment detection algorithm.4.To address the challenges of long-term effectiveness and computational efficiency in image line segment matching for visual localization,this dissertation proposes an illuminationinsensitive line binary descriptor based on band differences.The core insight is that adjacent/symmetric band differences within the same granularity and feature data channel are insensitive to illumination changes.The dissertation achieves long-term effective and computationally efficient illumination-insensitive line segment matching by using integral images for rapid feature statistics and comparison across multiple granularities and feature data channels.5.To validate the practicality and feasibility of the overall work,this dissertation designs a long-term visual localization system that integrates point and line features.To enhance the long-term effectiveness and computational efficiency of joint matching of image feature points and line segments in visual localization,the dissertation introduces a feature-matching filtering and multi-camera pose estimation algorithm based on weighted M-estimator sample consensus.The core lies in integrating various feature attributes in two-dimensional and three-dimensional spaces to customize weights reflecting the importance of feature matching.The proposed algorithm achieves long-term effective and computationally efficient feature-matching filtering and camera pose estimation.Experimental results based on real and virtual datasets demonstrate the excellent performance of the proposed image feature point and line segment detection and description algorithms in terms of long-term effectiveness and computational efficiency.Moreover,demonstration results on a real long-term driving dataset illustrate the long-term accurate localization capability of the designed long-term visual localization system across seasons.