基于CNN特征提取和增量式字典的VSLAM回环检测

VSLAM LOOP CLOSURE DETECTION BASED ON CNN FEATURE EXTRACTION AND INCREMENTAL DICTIONARY

在视觉SLAM系统中,传统的回环检测方法难以同时满足通用性和实时性。通过标志区域提取和CNN特征提取,提出在线构建增量式字典的回环检测方法。通过对图像进行随机扭曲来模拟运动产生的视角变化,结合GIST特征实现无监督的模型快速训练。通过局部标志区域的二进制特征实现快速检索,全局浮点特征实现选择最优匹配。实验表明,与传统方法相比,在100%准确率前提下,召回率提升约30%,整体查询时间约200 ms,内存占用约30 MB。在不同场景下检测更稳定,能够实现快速鲁棒的回环检测。

In the visual SLAM system,the traditional loop closure detection method is difficult to satisfy both versatility and real-time.Through marker region extraction and CNN feature extraction,we propose a loop closure detection method for constructing incremental dictionary online.The image was randomly distorted to simulate the change of the angle of view generated by the motion.This method combined the GIST feature to realize the unsupervised rapid training of the model.The fast retrieval was realized by the binary features of the local marker area,and the selection of the optimal matching was completed by the global floating point feature.Experiments show that compared with the traditional method,the recall rate is improved by about 30%under the premise of 100%accuracy,the overall query is about 200 ms,and the memory usage is about 30 MB.The detection is more stable in different scenarios,and is fast and robust.