Lightweight Neural Networks with Attention Mechanism for Loop Closure Detection in Visual SLAM

Abstract

Visual Simultaneous Localization and Mapping (SLAM) systems face significant challenges in loop closure detection, particularly in dynamic environments with varying illumination and viewpoint changes. Traditional methods relying on handcrafted features and bag-of-words models demonstrate limited robustness and computational efficiency. This research proposes a novel lightweight neural network architecture incorporating attention mechanisms to enhance loop closure detection performance while maintaining real-time computational requirements. The proposed method integrates an efficient channel attention module within a compressed MobileNetV2 backbone, enabling accurate feature extraction with reduced computational overhead. Experimental validation on standard datasets including TUM RGB-D, KITTI, and New College demonstrates superior performance compared to conventional approaches. The lightweight design achieves a 30.8% improvement in computational efficiency while maintaining comparable accuracy metrics. The attention mechanism effectively focuses on discriminative features, improving robustness to environmental variations. Results indicate that the proposed approach successfully addresses the trade-off between computational complexity and detection accuracy, making it suitable for resource-constrained robotic applications. The method demonstrates enhanced generalization capabilities across diverse indoor and outdoor scenarios, contributing to more reliable autonomous navigation systems.