A Novel Method Based on Local Region Division Using Point Clouds for Place Recognition

In the field of robotics, place recognition is a critical part of global localization and loop closure detection. This technology will describe the scene information effectively, and use a succinct pattern to represent the places for retrieval and matching. In this paper, we proposed a new place recognition method with deep learning tools combining local region division and handcrafted local features. Firstly, we proposed two strategies for local region division. The first strategy adopted the uniform sampling method from the point clouds and grouped the neighbors for every region. The second strategy extracted key points from point clouds and grew regions from these points. The two strategies were combined effectively to adapt to various environments. Then, we integrated handcrafted local descriptors into the network pipeline for more local structure information, and finally obtained a global feature descriptor with strong descriptive power and distinctiveness. Through experiments on public datasets, it is showed that the global descriptor obtained by our proposed method can achieve higher recognition accuracy than the same type of methods.