Deep learning based on LSTM model for enhanced visual odometry navigation system

Abstract

UAVs are employed for military, commercial, environmental, and other objectives. Flying in complex situations might strain the GPS (GNSS). Using INS alone increases positional inaccuracy. Despite cameras and sensors, drift persists. This work provides a GNSS-free UAV navigation system employing optical odometry, radar height estimates, and multi-sensory data fusion. Our monocular VO with optical flow leverages LSTM networks. We use optical flow to determine the vehicle's forward speed and LSTM to correct drift. A five-set LSTM model trained on GNSS data produces the velocity difference. The suggested technology was flight-tested. Experiments indicate the system can counteract lost GNSS signals' effects on forward and lateral speed. When GNSS signals are lost, the proposed strategy reduces average forward and lateral velocity errors to 63.01% in 30 s, 62.26% in 60 s, 58.76% in 90 s, and 54.33% in 113 sec.

Keywords

• Deep learning
• Long Short-Term Memory Networks
• Multi-sensor data fusion
• INS
• VO