基于图像感兴趣区域的机器人闭环检测算法

A Loop Closure Detection Algorithm for Robots Based on Region Proposals of Interest of Image

  • 摘要: 基于深度学习的机器人闭环检测算法在复杂光照条件下具有一定的鲁棒性,但在视角变化明显的场景下检测效果不佳,为此本文提出一种利用图像感兴趣区域的闭环检测新方法.首先,通过多尺度感兴趣区域网络(MSRPN)获得图像中的感兴趣区域,用改进的PlaceCNN(基于Place数据集的卷积神经网络)提取感兴趣区域的特征.然后,采用先粗匹配后细匹配原则,提出一种基于RPOI_PlaceCNN(基于图像感兴趣区域的PlaceCNN)的闭环检测算法,并利用双向匹配对之间的空间约束,去除不正确的匹配对,以提高闭环检测的整体准确性.在GardensPoint、Mapillary、Norland三种公开数据集上对方法的有效性进行了实验验证.实验结果表明,本文提出的闭环检测算法在光照、视角和不同变化组合引起的显著变化场景下依然能表现出较强的鲁棒性.

     

    Abstract: The loop closure detection algorithm of robots based on deep learning shows certain robustness under complex illumination, but it is prone to be affected by the scene with obvious changes of view angle. Therefore, a new loop closure detection method using RPOIs (region proposals of interest) of image is proposed. Firstly, the RPOIs of image are obtained by the improved MSRPN (multi-scale region proposal network), and the feature of interested region proposals is extracted by the improved PlaceCNN (Place dataset based convolutional neural network). Then, considering the shape similarity of region proposals, a loop closure detection algorithm based on RPOI_PlaceCNN (RPOI based PlaceCNN) is proposed by adopting the principle of coarse matching firstly and fine matching secondly. The space constraint between bidirectional matching pairs is used to remove incorrect matching pairs, which can improve the overall accuracy of loop closure detection. The effectiveness of the proposed method is experimentally verified on three public datasets, i.e. GardensPoint, Mapillary and Norland datasets. The experimental results show that the loop closure detection algorithm proposed can exhibit strong robustness in the situations of significant scene changes caused by illumination, view angle and different combinations of changes.

     

/

返回文章
返回