Enhancing Visual Place Recognition with Semantic Segmentation Filters

Visual Place Recognition (VPR) in the urban environment presents a significant challenge due to the existence of both static objects for example roads, terrains, and buildings, and dynamic objects such as pedestrians, vehicles, and sky. Dynamic objects, which appearance can change significantly with time, could negatively impact the VPR model. In order to address this, attentional mechanisms are often employed to assign relevance weights to features, thereby focusing the model on static objects such as buildings and roads. However, these approaches may not always accurately identify each feature. In this paper, we propose a semantic segmentation based approach that incorporates a semantic filter into the VPR model. This integration aims to exclude dynamic objects and accurately label every feature. This study demonstrates that this combined approach, leveraging the strengths of semantic segmentation, significantly improves the VPR model performance.