Research on Visual SLAM Algorithm Based on Improved LSD Line Feature Extraction Algorithm

Yuang Guan, Li Qian, Jinyang DuElectronics 2026 · DOI

TL;DR

This paper tackles the problem of over-segmentation and fragmentation in line feature extraction for Visual SLAM, a common issue with the standard LSD algorithm. They introduce an improved LSD algorithm (LSD_Advance) that enhances line feature quality and robustness through adaptive length filtering, intelligent grouping, and a novel merging strategy, leading to better SLAM performance without sacrificing real-time capabilities.

What Problem Does This Paper Address?

Simultaneous Localization and Mapping (SLAM) is crucial for robots to navigate and build maps of their surroundings. While visual SLAM (using cameras) is cost-effective and flexible, it struggles in environments with sparse features (low texture). Line features, being more robust and structurally informative, offer a solution. However, existing line feature extraction methods, particularly the widely used Line Segment Detector (LSD), suffer from over-segmentation, producing many short, fragmented line segments. This leads to increased false matches, degraded pose estimation accuracy, and overall system instability. This paper aims to improve the LSD algorithm to overcome these limitations, enhancing line feature quality and consistency for more robust Visual SLAM.

Methods

The authors propose three key improvements to the LSD algorithm, collectively termed LSD_Advance:

  • Adaptive Length Suppression: This strategy filters out short, potentially unreliable line segments. It dynamically calculates a minimum length threshold based on the current frame's line segment lengths and inter-frame changes, ensuring that only structurally relevant long lines are retained.
  • Adaptive Segment Grouping: To address the issue of fragmented segments, they introduce a two-stage grouping strategy:
  • Angle Grouping: Line segments are initially grouped into 18 clusters based on their orientation (10-degree intervals). This reduces the search space for subsequent fusion.
  • Endpoint Distance Grouping: Within each angle group, segments are further grouped based on the proximity of their endpoints. This helps identify segments that likely belong to the same underlying structure.
  • Intelligent Segment Merging: A novel similarity-based merging strategy is employed. Instead of a simple baseline selection, they define "baseline groups" and then fuse segments from non-baseline groups based on a similarity score that considers length, angle, and positional relationship. This aims to reconstruct more complete and accurate line segments.

Datasets

The paper evaluates its proposed method on the following datasets:

  • Kitti Dataset: Used for initial experiments comparing line feature extraction metrics and overall processing time.
  • YorkUrban Dataset: A dataset with indoor and outdoor images, used for evaluating line segment extraction performance using Average Precision (AP), Average Recall (AR), and F-score.
  • EuRoC Dataset: A benchmark dataset used for evaluating the performance of the improved algorithm in complex structured environments, focusing on average time consumption per frame, number of extracted line features, and matching rate with ground truth.

Key Insights and Ideas

  • The "fragmentation problem" of LSD is a significant bottleneck: The paper clearly articulates how over-segmentation directly impacts SLAM accuracy and robustness.
  • Adaptive thresholds are key for robust feature extraction: Dynamically adjusting parameters based on image content and inter-frame changes is more effective than fixed thresholds.
  • Hierarchical grouping is crucial for managing complexity: Breaking down the problem into angle-based and then endpoint-based grouping significantly reduces computational load for merging.
  • Similarity-based merging is superior to simple fusion: Considering multiple geometric attributes (length, angle, position) leads to more accurate reconstruction of line segments.
  • Improved line quality directly benefits SLAM: The paper demonstrates that better line features translate to fewer mismatches and a more stable SLAM system.

Key Results

  • Improved Line Quality: Achieves a 14.3% increase in the proportion of long line segments.
  • Enhanced Accuracy: An 8% improvement in Average Precision (AP) and a 15% rise in Average Recall (AR) on the YorkUrban dataset.
  • Competitive Performance: The LSD_Advance algorithm achieves the best overall accuracy metrics (AP, AR, F-score) compared to LSD, FLD, and EDLines.
  • Real-time Capability: While there's a moderate increase in per-frame processing time (from 19.5 ms to 25.5 ms), it remains within the acceptable range for most real-time Visual SLAM systems.
  • Reduced Feature Mismatches: Contributes to an overall 12.39% reduction in the feature mismatch rate in subsequent SLAM stages.

Limitations and Future Work

  • Increased Processing Time: The proposed LSD_Advance algorithm has a slightly higher per-frame processing time compared to the original LSD due to the added grouping and merging steps. However, the authors argue this is a favorable trade-off for the significant improvement in feature quality.
  • Future Work: The paper implicitly suggests that further optimization of the grouping and merging steps could potentially reduce the processing time further. Additionally, exploring the integration of these improved line features into more advanced SLAM frameworks (e.g., those incorporating deep learning for other aspects) would be a natural next step. The authors also highlight the potential for this method in complex, low-texture environments, suggesting further validation in such scenarios.

Summary updated

Abstract

Visual SLAM algorithms rely on image sequences to achieve autonomous localization and mapping, where line features act as crucial structural information to enhance system robustness in weakly textured or structured environments. However, conventional line feature-based methods, such as the Line Segment Detector (LSD) algorithm, are prone to over-segmentation during line segment extraction, resulting in a large number of redundant short segments and fragmented line pieces. This phenomenon increases the false matching rate, which in turn degrades the accuracy of pose estimation and the overall stability of the Visual SLAM system. To address the above issues, we perform comparative experiments on multiple public datasets between the proposed improved line feature algorithm and classical counterparts from dimensions of time overhead, line feature number and detection accuracy. The results show that the proposed algorithm incurs a 20% increase in overall time for line feature extraction and matching, yet achieves a 14% higher proportion of long line segments, an 8% improvement in Average Precision (AP) and a 15% rise in Average Recall (AR). It is thus verified that the proposed method retains real-time performance while remarkably improving its line segment matching success rate, with its localization accuracy and system robustness maintained or even enhanced.
BibTeX
@inproceedings{guan2026research,
  author = {Yuang Guan and Li Qian and Jinyang Du},
  title = {{Research on Visual SLAM Algorithm Based on Improved LSD Line Feature Extraction Algorithm}},
  booktitle = {Electronics},
  year = {2026},
  doi = {10.3390/electronics15051006},
}
Figure