Dynamic Feature Elimination-Based Visual-Inertial Navigation Algorithm.

Area of Science:

• Robotics
• Computer Vision
• Sensor Fusion

Background:

• Traditional visual-inertial navigation systems (VINS) suffer from reduced accuracy in dynamic environments due to moving objects.
• Dynamic objects create interference, leading to significant pose estimation errors.

Purpose of the Study:

• To propose an improved VINS algorithm that enhances positioning accuracy in dynamic scenarios.
• To address the limitations of existing VINS by integrating multi-scale feature optimization and dynamic feature elimination.

Main Methods:

• Reconstructed the SuperPoint encoder with dual-branch multi-scale feature fusion and channel compression for robust feature extraction.
• Developed the Adaptive Simple Online and Realtime Tracking (ASORT) algorithm for real-time moving object detection and dynamic feature point filtering.
• Utilized an object detection network, Kalman filter, and Hungarian algorithm within ASORT to ensure only static features are used for backend optimization.

Main Results:

• The proposed method achieved an average improvement of 14.8% in absolute trajectory accuracy on the KITTI dataset compared to the original VINS-Fusion.
• The algorithm processes single frames in an average of 23.9 milliseconds, demonstrating efficiency.
• Successfully minimized pose estimation errors caused by dynamic interference.

Conclusions:

• The enhanced VINS algorithm provides an efficient and robust solution for navigation in highly dynamic environments.
• The synergistic combination of feature optimization and dynamic feature elimination effectively overcomes the challenges posed by moving objects.