基于深度学习的视觉SLAM闭环检测方法

针对在摄像机视角、光照、气候、地貌等条件的大 幅度变化或者存在快速移动物体的 复杂场景下,视觉即时定位与地图构建(simultaneous localization and mapping,SLAM)的精确性和鲁棒性较低等问题,闭环检 测作为解决SLAM位姿漂移的重要环节,提出了一种基于神经网络的闭环检测方法。该方 法通过传感器获取视觉图像的数据,不同于传统方法的特征提取,采用改进三重约束损 失函数训练Darknet提取图像特征,构造对应特征向量矩阵。由于Darknet借鉴了残差网络(resnet)的思想,在具有较深网络层数的同时,仍保持较高的准确率,减少了特征提取 误差。经过自编码器方法对数据进行降维处理,通过余弦相似度计算,设定合理阈值,能够 更快的得到闭环检测结果。最后通过在两个公开视觉SLAM闭环检测数据集,New College数 据集和光照及角度变化更明显的City Centre数据集上进行实验,结果表明复杂环境下本文 提出的方法比现有闭环检测方法,能够得到更高准确率和速率,更好满足了视觉SLAM系统对 消除累计误差和实时性的要求。

The closed loop detection method of vision SLAM based on deep learning

For the camera perspective,lighting,climate,landform and other cond itions of large changes or the existence of fast moving objects in the complex s cene,the accuracy and robustness of simultaneous Locali-zation and mapping (SLAM) are low,a closed-loop detection as a solution to SLAM pose an import a nt link of the drift,combined with this paper proposes a closed-loop detection method based on neural network.In this method,visual image data is obtained th rough sensors.Different from the traditional method of feature extraction,the improved triple constraint loss function is adopted in this paper to train Darkn et to extract image features and construct corresponding feature vector matrix.B ecause Darknet borrowed from the idea of residuals network (RESNET),it has a de ep network layer while maintaining a high accuracy rate,which greatly reduces t he error of feature extraction.Through the self-encoder method to reduce the di m ension of the data,through the cosine similarity calculation,set a reasonable threshold value,can get the closed-loop detection results faster.Last through t he two open visual SLAM closed loop testing data set,the new college data colle ction and light and the angle change is more obvious city centre data set on the experiment,the results show that the proposed method under complicated environ ment than the existing closed loop detection method,can get higher accuracy and speed,and better meet the visual SLAM system to eliminate the accumulated erro r and real-time requirements.