针对存在明显光照变化或遮挡物等室外复杂场景下,现有基于深度学习的视觉即时定位与地图构建(visual simultaneous localization and mapping,视觉SLAM)回环检测方法没有很好地利用图像的语义信息、场景细节且实时性差等问题,本文提出了一种YOLO-NKLT视觉SLAM回环检测方法。采用改进损失函数的YOLOv5网络模型获取具有语义信息的图像特征,构建训练集,对网络重训练,使提取的特征更加适用于复杂场景下的回环检测。为了进一步提高闭环检测的实时性,提出了一种基于非支配排序的KLT降维方法。通过在New College数据集和光照等变化更复杂的Nordland数据集上进行实验,结果表明:室外复杂场景下,相较于其他传统和基于深度学习的方法,所提方法具有更高的鲁棒性,可以取得更佳的准确率和实时性表现。

Aiming at the problems that the existing deep learning based visual simultaneous localization and mapping (visual SLAM) loopback detection methods do not make good use of the semantic information of images,scene details and poor real-time performance in complex outdoor scenes with obvious illumination changes or occlusion objects,this paper proposes an YOLO-NKLT visual SLAM loopback detection method.The YOLOv5 network model with improved loss function is used to obtain image features with semantic information,construct the training set,and retrain the network to make the extracted features more suitable for loopback detection in complex scenes.In order to further improve the real-time performance of closed-loop detection,a KLT dimensionality reduction method based on non-dominated sorting is proposed.Through experiments on the New College dataset and Nordland dataset with more complex changes such as illumination,the results show that compared with other traditional and deep learn-based methods in outdoor complex scenes,it has higher robustness and can achieve better accuracy and real-time rate performance.

天津市高等学校自然科学研究项目(2017ZD13)资助项目