1、Introduction

DL解決VO問題：End-to-End VO with RCNN

2、Network structure

a.CNN based Feature Extraction

　　論文使用KITTI數據集。

　　CNN部分有9個卷積層，除了Conv6，其他的卷積層後都連接1層ReLU，則共有17層。

b、RNN based Sequential Modelling

　　RNN is different from CNN in that it maintains memory of its hidden states over time and has feedback loops among them, which enables its current hidden state to be a function of the previous ones.

　　Given a convolutional feature x_k at time k, a RNN updates at time step k by

　　hk and yk are the hidden state and output at time k respectively.

　　W terms denote corresponding weight matrices.

　　b terms denote bias vectors.

　　H is an element-wise nonlinear activation function.

　　LSTM

Folded and unfolded LSTMs and internal structure of its unit.

　　is element-wise product of two vectors.

　　σ is sigmoid non-linearity.

　　tanh is hyperbolic tangent non-linearity.

　　W terms denote corresponding weight matrices.

　　b terms denote bias vectors.

　　ik, f k, gk, ck and ok are input gate, forget gate, input modulation gate, memory cell and output gate.

　　Each of the LSTM layers has 1000 hidden states.

3、損失函數及優化

　　The conditional probability of the poses Y_t = (y₁, . . . , y_t) given a sequence of monocular RGB images X_t = (x₁, . . . , x_t) up to time t.

　　Optimal parameters :

　　The hyperparameters  of the DNNs:

　　(p_k, φ_k) is the ground truth pose.

　　(p?_k, φ?_k) is the estimated ground truth pose.

　　κ (100 in the experiments) is a scale factor to balance the weights of positions and orientations.

　　N is the number of samples.

　　The orientation φ is represented by Euler angles rather than quaternion since quaternion is subject to an extra unit constraint which hinders the optimisation problem of DL.

DeepVO: Towards End-to-End Visual Odometry with Deep Recurrent Convolutional Neural Networks