基于深度神经网络的井下无人机视觉位姿估计

Visual post estimation of underground UAV based on deep neural network method

无人机将在未来少人或无人采矿中发挥重要作用,而位姿估计则是实现井下无人机自主巡检的关键.针对井下巷道照度分布不均匀和动态复杂环境的特点,提出采用慕尼黑工业大学深度图像(TUM RGB-D)数据集对深度神经网络模型进行预训练的方法,提取巷道特征路标点.为实现具有真实尺度信息的位姿估计,首先利用机载相机三维深度数据流恢复网络特征点深度,然后建立帧间匹配巷道路标点最小二乘模型,最后采用奇异值分解的方法获得无人机位姿.开发了手持移动传感器数据采集系统,完成传感器相对位置标定,采集真实巷道环境数据并进行实验.实验结果表明,相比ORB-SLAM2位姿估计结果,提出的无人机位姿估计方法针对巷道复杂环境数据其定位精度可提高71%以上,定位误差约为13cm.

UAV(Unmanned Aerial Vehicle)will play a significant role for coal mining with few people or unmanned coal mining in the future,and pose estimation is the key to the autonomous coalmine inspection by UAV.According to the characteristics of uneven illumination distribution in the underground tunnel scene and dynamic complex environment,a method of pretraining deep neural network using TUM RGB-D dataset was proposed to extract roadway features.Firstly,with the aim of realizing pose estimation with real-scale information in underground roadways,the depth of network model feature points was recovered from 3Ddepth data stream of airborne camera.After that,the least square model of inter-frame matched landmark points was constructed.Finally,the singular value decomposition method was adopted to estimate the UAV pose.A hand-held mobile sensor data collection device was developed and calibrated,which was used to collect the real tunnel environment data and to verify the proposed pose estimation method.The experimental results show that,compared with ORB-SLAM2 pose estimation results,the proposed UAV pose estimation method for underground complex environment can improve the localization accuracy by more than 71%and the localization error is about 13cm.