Abstract
The ability to perceive environments supports an important foundation
for our self-developed robotic rat to improve kinematic performance and
application potential. However, the most existing visual perception of
quadruped robots suffers from poor perception accuracy in real-world
dynamic environments. To mitigate the problem of erroneous data
correlation, which is the main cause of low accuracy, the work presents
an approach that combines leg odometry (LO) and IMU measurements with
VSLAM to provide robust localization capabilities for small-scale
quadruped robots in challenging scenarios by estimating the depth map
and removing moving objects in dynamic environments. The method contains
a depth estimation network with higher accuracy by combining the
attention mechanism in the Transformer with the RAFT-Stereo depth
estimation algorithm. Besides, the method combines target identification
and segmentation with 3D projection of feature points to remove moving
objects in dynamic environments. In addition, LO and IMU data are fused
in the modified framework of ORB-SLAM3 to achieve highly accurate
localization. The proposed approach is robust against erroneous data
correlation due to moving objects and wobbles of quadruped robots.
Evaluation results on multiple stages demonstrate that the system
performs competitively in dynamic environments, outperforming existing
visual perception methods in both public benchmarks and our costumed
small-scale robotic rat.