An Illumination-Adaptive Visual SLAM Algorithm With Lightweight Object Detection for Complex Dynamic Environments
- 7 July 2025
- journal article
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 74 (00189456) , 1-22
Abstract
Existing visual simultaneous localization and mapping (SLAM) algorithms excel in stable illumination variation and static environments, but struggle in dynamic environments with varying illumination. Object detection, an essential component for dynamic feature removal, imposes substantial computational burdens due to model architectural complexity. This study proposed a robust visual SLAM framework that fuses an illumination-adaptive thresholding feature point extraction module with a lightweight object detection network (YOLO+ShuffleNetV2+ EMA) to effectively separate dynamic and static features by improving feature point extraction and matching accuracy. The fusion module is deployed at the SLAM front-end and utilizes an independent threading mechanism to decouple the dynamic feature point removal process from the main pipeline. Furthermore, a back-end point cloud mapping module is introduced into the SLAM framework, which collaborates with an RGB-D camera and the YOLO+ShuffleNetV2+EMA network to eliminate dynamic object point clouds. Experimental results on indoor and outdoor datasets demonstrate that the proposed SLAM framework significantly enhances feature point processing capability and localization accuracy under varying illumination and dynamic environments. Under illumination conditions varying from 30% to 130% within the same scene, the adaptive thresholding method consistently maintains an average feature point count exceeding 1700, with an average variation of 784 feature points across different environments. The lightweight object detection network achieves a 70.72% reduction in parameter count and a 71.33% decrease in model size. In terms of localization accuracy, the framework achieves a maximum absolute trajectory error reduction of over 95% and a maximum sum of squared errors reduction of over 94%. The trajectories generated by the proposed SLAM algorithm closely align with the ground truth and enable real-time computation using CPU-only processing.Keywords
Funding Information
- National Key Research and Development Program of China (2024YFB2505705)
- Science and Technology Plan Program of Zhejiang Communications Investment Group Company Ltd. (202303)
- Science and Technology Research Project of Yunnan Communications Investment & Construction Group Co., Ltd. (YCIC), China (YCIC-YF-2022-06)
This publication has 51 references indexed in Scilit:
- A comprehensive overview of core modules in visual SLAM frameworkNeurocomputing, 2024
- Large Separable Kernel Attention: Rethinking the Large Kernel Attention design in CNNExpert Systems with Applications, 2023
- PPLC-Net:Neural network-based plant disease identification model supported by weather data augmentation and multi-level attention mechanismJournal of King Saud University Computer and Information Sciences, 2023
- Dynamic visual simultaneous localization and mapping based on semantic segmentation moduleApplied Intelligence, 2023
- Fast visual inertial odometry with point–line features using adaptive EDLines algorithmMeasurement Science and Technology, 2022
- A review of visual SLAM methods for autonomous driving vehiclesEngineering Applications of Artificial Intelligence, 2022
- Improved FAST corner‐detection methodThe Journal of Engineering, 2019
- Microsoft COCO: Common Objects in Context2014
- Adaptive document image binarizationPattern Recognition, 2000
- Adaptive histogram equalization and its variationsComputer Vision, Graphics, and Image Processing, 1987