基于注意力卷积神经网络的视觉里程计

Visual Odometer Based on Attention-convolution Neural Network

传统的视觉里程计(visual odometry,VO)要求图像含有大量的纹理信息,且求解过程较为复杂。针对以上问题提出基于注意力卷积神经网络的视觉里程计,对相机进行端到端的位姿估计,利用注意力机制提高模型估计轨迹的精度。首先,使用注意力-卷积神经网络(convolutional neural networks,CNN)模块提取图像特征;然后,将特征输入到门控循环单元(gated recurrent unit,GRU)学习图像的时序连接性;最后,通过全连接层降维输出相机位姿。在KITTI数据集上完成实验,并与其他方法进行对比,结果表明卷积网络中加入注意力机制可以有效提高轨迹估计的精度,且误差低于其他视觉里程计算法。

The traditional visual odometry requires a lot of texture information in the picture,and the solution process is complex.To solve the above problems,a visual odometer based on attention-convolution neural network is proposed to estimate the pose of camera end-to-end.Firstly,the attention-convolutional neural networks are used to extract the features of the images,then,the features are input to the gated recurrent unit to learn the temporal connectivity.Finally,the camera pose is output through full connection layer dimensionality reduction.The experiment is completed on KITTI data set and compared with other methods.The results show that adding attention mechanism to convolution network can effectively improve the accuracy of trajectory estimation,and the error is lower than other visual mileage calculation methods.