Visual SLAM system

VSlam Loop Closure Detection Based on Transformer (05/2021-04/2022)

• Designed and implemented transformer-based feature extraction with Pytorch and designed sequence matching at the back-end of LCD.
• Trained on Places365 with knowledge distillation, achieving top-1 of 53.28% and top-5 of 84.04%, comparable to CNN-based models.
• Improved the average precision by 3.18% compared to state-of-the-art CNN-based methods on NewCollege and CityCentre datasets.
• Authored and published a paper in 2022 IEEE International Conference on Advanced Robotics and Mechatronics.
  • 2022 ICARM Paper
  • 2022 ICARM Slide

Light Monocular Visual Odometry through Attentive Tensor-compressed LSTM Model for Robotic VSlam (09/2020-12/2020)

• Designed a CNN+T-LSTM model with attention mechanism to estimate the 6-DoF absolute-scale pose from the optimal flow feature for the monocular visual Odometry, making it friendly to be applied on edge devices.
• Achieved 1/7 the size of DeepVO and 23x faster than Flowdometry on KITTI dataset, while deployed it to a Raspberry Pi-based robot.
• Authored and published a paper in 2021 WRC Symposium on Advanced Robotics and Automation, awarded the Best Student Paper.
  • 2021 WRC Paper