MSGC-NetVLAD: a lightweight visual place recognition neural network for autonomous robots

Long-term localization in large-scale urban environments remains a core challenge for autonomous robots, especially when global navigation satellite systems (GNSS) become unreliable due to multipath effects and signal occlusion. Visual place recognition (VPR) offers a complementary solution by leveraging environmental visual cues to estimate position; however, existing deep VPR frameworks often rely on large backbone networks such as VGG16, resulting in excessive computation and memory demands. To address this limitation, we propose MSGC-NetVLAD, a lightweight VPR network designed for mobile robotic platforms. The proposed model integrates multi-scale group convolution (MSGC) modules and efficient channel at tention (E CA) mechanisms to capture hierarchical spatial features while minimizing redundancy. By combining multi-scale dilated convolutions and grouped operations, MSGC-NetVLAD achieves efficient feature extraction and compact descriptor aggregation through a NetVLAD pooling layer. Experimental results on benchmark datasets—Pitts30k and Tokyo24/7—demonstrate that our approach achieves up to 3% higher Recall@1 accuracy compared with lightweight baselines (e.g., MobileNetV3-NetVLAD), while reducing trainable parameters by over 60%. These results confirm that MSGC-NetVLAD provides a strong balance between recognition accuracy and computational efficiency, enabling practical deployment in mobile VPR systems.