基于自监督学习的温室移动机器人位姿跟踪

为实现温室环境下机器人行进过程中的位置及姿态跟踪,该研究提出一种基于时序一致性约束的自监督位姿变换估计模型。模型用软遮罩,处理视频帧间静止造成的位姿预测值收缩现象,进一步用归一化遮罩,解决非刚体场景和目标遮挡问题。设计了一种星型扩张卷积,并基于该卷积,为模型构建自编码器。在采集自种植作物为番茄的日光温室视频数据上开展训练和测试试验。结果表明,与不采用遮罩处理的模型相比,采用软遮罩的模型,位置和姿态估计相对误差分别减少5.06个百分点和11.05个百分点,采用归一化遮罩的模型,这2项误差则分别减少4.15个百分点和3.86个百分点,2种遮罩均可显著提高模型精度;星型扩张卷积对降低模型误差是有效的,在网络参数不变的前提下,该卷积使姿态估计相对误差减少7.54个百分点;时序一致性约束使姿态估计均方根误差下降36.48%,每百帧累积姿态角误差降低54.75%,该约束可用于提高模型精度及稳定性;该研究的位置及姿态估计相对误差分别为8.29%和5.71%,与Monodepth2相比,减少了8.61%和6.83%。该研究可为温室移动机器人导航系统设计提供参考。

Self-supervised pose estimation method for a mobile robot in greenhouse

Simultaneous localization and mapping (SLAM) play a vital role in implementing autonomous navigation of mobile robots in an unknown environment. Especially, visual odometry (VO) is a core component for a localization module in the SLAM system. The pose and velocity of a robot can, therefore, be estimated using computational geometry. Furthermore, the learning-based VO has gained great success in joint estimation camera ego-motion and depth from videos. In this study, a novel self-supervised VO model was proposed to realize the autonomous operation of a mobile robot in a greenhouse. The consistency constraint of temporal depth was also introduced for the learning framework using the binocular baseline supervision. Stereo video sequences were selected to train the model. The pose network after training was then used for pose estimation. A pre-test found that the stillness between video frames caused the prediction value of the model to shrink. Therefore, a soft mask was used in photometric re-projection error to remove the static region from the apparent difference measurement, and the non-rigidity scene and occlusion were further solved with normalized mask planes. Meanwhile, a new type of star dilated convolution (SDC) was also designed, where the filter was used to extract image features from the center 3×3 solid kernel and eight directions of 1-D kernel. The computational cost of SDC was thus less than that of the regular convolution of the same receptive field. Moreover, SDC was superimposed on spatial dimensions using depth-wise convolution with different dilation rates, particularly without the necessary to modify the existing deep learning framework. A convolutional auto-encoder (CAE) with residual network architecture was constructed using the SDC and inverse residual module (IRM), further serving as the backbone network for the VO model. With the aid of a binocular camera, the video sequences were collected in the solar greenhouses with tomato as the crop. The stereo video dataset was constructed to carry out the training and testing experiments. The static samples of the video sequence were removed from the image apparent difference measurement with a soft mask. The results showed that the mean relative errors (MREs) of translation and rotation estimation in the model were cut down by 5.06% and 11.05%, respectively, while the mean square root errors (RMSE) were reduced by 24.78% and 30.65%, respectively. Once a normalized mask plane was utilized in the model to deal with non-rigidity scenes and occlusion, the MREs of translation and rotation estimation were reduced by 4.15% and 3.86%, respectively. It inferred that both masks significantly improved the accuracy of the model. Meanwhile, the SDC-based IRM (SDC-IRM) reduced the MRE of rotation by 7.54% under the unchanged network parameters. Since the SDC-IRM structure presented significant effectiveness in reducing model error, the increase of perceptive field was an effective way to improve the accuracy of the model. The MRE and RMSE of rotation estimation were reduced by 2.74% and 36.48%, respectively, whereas, the mean cumulative rotation error per hundred frames (MCRE) decreased by 54.75%, when the consistency constraint of temporal depth was used in the model, indicating high accuracy and stability of pose estimation. The MRE of rotation estimation was reduced by 7.30% when extending the expansion factor of IRMs. The data demonstrated that the increase of receptive fields in the SDC kernel contributed to the higher accuracy of rotation estimation. Nevertheless, there was no longer obvious improvement of the model, when the maximum dilation rate was more than 6. More importantly, the calculation speed was up to 56.5 frames per second in the final pose estimation network. The MREs of translation and rotation estimation were 8.29% and 5.71%, respectively. The pose estimation performed better, compared with the previous VO model under similar input settings. This finding can provide sound support for the design of the navigation system for mobile robots in a greenhouse.