单级特征图融合坐标注意力的视觉位置识别方法

Visual Place Recognition with One Level Feature Fusing Coordinate Attention

针对视角和环境变化的场景中现有视觉位置识别方法存在的匹配遗漏和实时性差的问题,提出基于单级特征图融合坐标注意力的视觉位置识别方法。首先通过坐标注意力捕获特征的相对位置信息,然后利用扩张卷积和局部聚合向量网络(NetVLAD)构造多尺度特征融合的编码器,最后基于三元组损失训练网络。经Pitts30k和Nordland数据集验证,在位置识别试验中,与同基线的先进方法Patch-NetVLAD相比,所提出的方法能够获得同等的召回精度且检索速度提高19%。在回环检测试验中,所提出的方法达到了合理平衡鲁棒性和检索速度的目标。

To solve the problems of matching omission and poor real-time performance of existing visual place recognition methods in scenes with changing viewpoints and environments,this paper proposed visual place recognition method based on one level feature fused with coordinate attention.Firstly,the relative place information of features was captured by coordinate attention.Secondly,an encoder for multi-scale feature fusion was constructed using dilated convolution and NetVLAD.Finally,the network was trained based on triplet loss.Validated by Pitts30k and Nordland datasets,the proposed method achieves the same recall accuracy and 19%faster retrieval speed compared with the state-ofthe-art method Patch-NetVLAD of the same baseline in the test of position recognition.In the test of loop detection,the proposed method achieves a reasonable balance between robustness and retrieval speed.