A Method for Long-Term Target Anti-Interference Tracking Combining Deep Learning and CKF for LARS Tracking and Capturing

Abstract

Autonomous underwater vehicles (AUV) recycling in an underwater environment is particularly challenging due to the continuous exploitation of marine resources. AUV recycling via visual technology is the primary method. However, the current visual technology is limited by harsh sea conditions and has problems, such as poor tracking and detection. To solve these problems, we propose a long-term target anti-interference tracking (LTAT) method, which integrates Siamese networks, You Only Look Once (YOLO) networks and online learning ideas. Meanwhile, we propose using the cubature Kalman filter (CKF) for optimization and prediction of the position. We constructed a launch and recovery system (LARS) tracking and capturing the AUV. The system consists of the following parts: First, images are acquired via binocular cameras. Next, the relative position between the AUV and the end of the LARS was estimated based on the pixel positions of the tracking AUV feature points and binocular camera data. Finally, using a discrete proportion integration differentiation (PID) method, the LARS is controlled to capture the moving AUV via a CKF-optimized position. To verify the feasibility of our proposed system, we used the robot operating system (ROS) platform and Gazebo software to simulate the system for experiments and visualization. The experiment demonstrates that in the tracking process when the AUV makes a sinusoidal motion with an amplitude of 0.2 m in the three-dimensional space and the relative distance between the AUV and LARS is no more than 1 m, the estimated position error of the AUV does not exceed 0.03 m. In the capturing process, the final capturing error is about 28 mm. Our results verify that our proposed system has high robustness and accuracy, providing the foundation for future AUV recycling research.