MDPR-Net: Dynamic Target Interference Removal and Autonomous Vehicle Place Recognition Network for Multi-View Images

Accurate navigation and localization are essential for autonomous vehicles in complex environments. Visual place recognition (VPR) provides an efficient and cost-effective method for environmental representation. Our study introduces MDPR-Net, an autonomous vehicle positioning network utilizing 360-degree images. The dynamic interference removal module (DIR) eliminates dynamic targets filtering, ensuring precise environmental perception. Following DIR, a multi-view image encoder module (MIE) encodes the filtered panoramic images with shared weights, capturing comprehensive features. The image-relation attention module (IRA) then associates these features across multi-view images, enhancing the model's ability to understand the scene contextually. This approach is demonstrated on the nuScenes dataset, yielding promising results.