Visual Odometry Algorithm Based on Deep Learning

Fig. 5. Mean translation errors and mean rotation errors under different conditions. (a) Mean translation errors under different path lengths; (b) mean rotation errors under different path lengths; (c) mean translation errors under different speeds; (d) mean rotation errors under different speeds

Download full size

Fig. 6. VO trajectories in different scenarios. (a) Scenario 03; (b) scenario 04; (c) scenario 09

Download full size

Fig. 7. Practical experiment. (a) VO trajectory in actual scene; (b) practical map

Download full size

Table 1. Parameters of CNN
View table
Table 1. Parameters of CNN
Layer Receptive field size Padding Stride Number of channels
Conv1 7×7 3 2 64
Conv2 5×5 2 2 128
Conv3 5×5 2 2 256
Conv3_1 3×3 1 1 256
Conv4 3×3 1 2 512
Conv4_1 3×3 1 1 512
Conv5 3×3 1 2 512
Conv5_1 3×3 1 1 512
Conv6 3×3 1 2 1024

Table 2. Parameter calculation of different algorithms

View table

Table 2. Parameter calculation of different algorithms

Algorithm	Constitute	FLOPs /10¹⁰	Parameter /10⁸	t_rel/%	r_rel /(°)
VISO2_M	Conventional (monocular)	\	\	17.48	16.52
VISO2_S	Conventional (stereo)	\	\	1.89	1.96
DeepVO	CNN+LSTM	5.143(not include LSTM)	5.1016	5.96	6.12
Proposed algorithm	CNN+attention	5.186	2.6031	8.85	15.68