VDI-Net: Viewpoint Changes and Dynamic Interference Immune Place Recognition Network for Intelligent Vehicles Based on Multiview Images

visual place recognition (VPR) improves the localization accuracy of agents in complex environments by extracting effective environmental representations for place matching, and it does not rely on additional high-precision, high-cost sensors or digital maps. However, current VPR research still faces limitations when simultaneously addressing the challenges of viewpoint changes and dynamic target interference, where it is not easy to obtain stable and reliable viewpoint-invariant environmental representations. To address these challenges, inspired by the human ability to recognize scenes, this article customizes a place recognition network called VDI-Net based on multiview images, immune to viewpoint changes and dynamic interference. Specifically, a dynamic target filtering (DTF) module is proposed to effectively filter out dynamic targets in complex environments. To tackle the challenge of viewpoint changes, a vision-surround Mamba module (VSM) is introduced to enhance the rotational invariance of features across different viewpoints. In the comparative validation of the nuScenes dataset and our real-vehicle collection dataset, the experimental results show that our proposed method achieves satisfactory performance in addressing the challenges of viewpoint changes and dynamic target interference. The proposed method outperforms current representative VPR baselines and surpasses some classical LiDAR-based and multimodal place recognition baselines.