Summary |
Unmanned autonomous underwater vehicles (AUV) is developed to carry out underwater investigation, environmental exploration, biological recording, etc. AI technologies are integrated to achieve accurate object detection and positioning. Modified Inception module is used in object detection network enhanced by parallelized multi-dimensional feature extraction algorithm to optimally reduce the network size without sacrificing too much accuracy. Underwater object image augmentation method by 3-D modeling is developed to enlarge underwater object databank. Our methods can reach 10fps and 83.645% accuracy on 24GFLOPS RPi3 for deep-sea unmanned AUVs.
The developed AUV recognizes many underwater objects, including fish, turtle, fishnet, etc., with over 90% average accuracy. The error of auto depth≤0.2 m, way point navigation error≤1 m, static object distance estimation error≤5 m (speed≤0.2 m/sec), tracing of special object (e.g., diver), for underwater positioning are also proved. |
Scientific Breakthrough |
The major breakthrough of this project is aimed at embedding AI technologies in unmanned AUVs for deep sea operations. All of major KPI comparisons with international benchmark (IBK) are listed as follows.
1. The developed AI-based underwater object recognition can identify 5 different objects at the same frame/time with 82.05% accuracy (mAP). 2. Underwater image enhancement is elevated to 7.406 over IBK= 7.199. PCQI is increased to 0.9859 higher than IBK = 0.9076. 3. AUV navigation control system performs auto-heading +/-5 degrees (IBK=10 degree), auto-depth +/-0.1 m (IBK=0.2m), and auto-altitude +/-0.1 m (IBK=0.2m). 4. Databank augmentation based on 3D modeling generates images with 10 viewpoints (IBK=4) and increases by 1000 times (IBK=60). |