Our AUV utilises Fast underwater image enhancement (FUnIE) for Improved Visual Perception, an image enhancing algorithm, to improve the quality of the image captured by our AUV’s monocular camera. An enhanced image allows for computer vision model to better identify and classify detected objects.

UNDERWATER IMAGE ENHANCEMENT FEATURES

• Designed specifically for underwater image enhancement
• Uses a generative adversarial network (GAN) architecture
• Capable of improving visibility and color correction in underwater imagery
• Performs real-time enhancement of underwater images and videos
• Reduces the haze, color cast, and noise common in underwater photography
• Trained on paired and unpaired underwater image datasets
• Designed to be lightweight compared to other enhancement methods
• Preserves structural information while enhancing visual quality
• Can be used as pre-processing for underwater computer vision tasks

Our chosen computer vision model for visual-based guidance is GroundingDINO. GroundingDINO uses zero-shot object detection, allowing our AUV to identify any objects of interest captured by the camera underwater, feeding the data back to the ROS mainframe for decision making.

COMPUTER VISION FEATURES

• Zero-shot object detection capabilities, allowing it to detect objects without specific training
• Text-to-box grounding, enabling users to specify objects to detect using natural language
• Strong generalisation abilities across different domains and datasets
• Combines visual and textual embeddings for multi-modal understanding
• Pre-trained on large-scale datasets including COCO and Objects365
• Capable of detecting both common and uncommon objects

The entire software stack is built using ROS Noetic. Additionally, a simulations platform is also created using Gazebo and ROS.

The ROS stack offers modularity and easy integration with hardware. It also provides access to a plethora of open-source libraries that can propel development.

SIMULATION FEATURES

• Support operation in ROV mode (remote controlled operations) and AUV mode (autonomous operations).
• Accurate physics (collisions, buoyancy, etc.).
• Custom worlds with tuning parameters for environment (eg. turbidity, lighting)
• Support for simulations of expensive sensors (eg. sonar sensors, dvl, etc.)
• Custom objects can be added.
• Supports CV models as well as other software algorithms.
• Support for SLAM algorithms (we tested mapping with ROVIO, DSO, and ORB-SLAM2)

This base platform allows spontaneous testing of software algorithms (navigation, perception, etc.) and reduces the deployment time.

Sensors used for AUVs can range from anywhere between S$3000-S$150,000. The platform helps with estimating the performance of different expensive sensors and compare them for custom missions.

The following software has also been tested on hardware –

1. Mapping Algorithms like DSO and ORB-SLAM2.
2. Bounding Box based autonomous navigation.
3. Remotely operated navigation where inputs were sent through an Xbox controller.
4. Data acquisition from sonar pinger, camera, barometer, IMU and processing of data for different algorithms.

FMonocular SLAM based map creation

Simulations environment