Real-time sonar object detection for an autonomous survey-robot

Customizeable, real-time Sonar Object Detection for an autonomous survey-robot

Surveying and search operations are carried out worldwide in lakes, canals or coastal waters to detect objects hidden underwater: scrap metal in shipping routes or old military loads, such as aerial bombs or underwater mines. To support these search operations, Evologics developed the Sonobot 5, one of the fastest floating sonar USVs (unmanned surface vessels) in the world. Among other things, it supports the police and coast guard in locating and recovering drowned people.

The Client

EvoLogics is a hightech company based in Berlin with strong maritime bionics and robotics R&D experience. They offer solutions for underwater communication and monitoring, including the Sonar USV with the highest speed on water worldwide. This USV should find and identify mission critical objects assisted by real-time object detection.

The Challenge

Autonomous robotic sonar surveys produce large amounts of complex, difficult to interpret sonar data. Points of interest (POI) are obscured by background scatter and noise, hidden in mountains of data.

Moreover, survey missions take place off-grid with limited hardware and limited connection to data processing infrastructure.

How do we assist the survey operators in finding POIs? Can the Bot scan interesting positions automatically from different angles? How can we train and deploy models to support the search for these limited edge devices, and with user specific object types?

The Solution

We developed a web portal that manages the clients' models, data, and updates for their embedded devices. In this platform the clients can configure and train customized models, choosing their ideal trade-off between speed and accuracy, sharing communal data or keeping their data confidential.

During the survey missions the POIs are detected and highlighted live and in real-time. The survey operators can also do the same anytime with replay data.
They can also set confidence filters for each detected class and optimize for clustered or solitary objects at detection time.

The updates for the models and detectors are modular, saving bandwidth and providing flexibility.

The Results

We finetuned a state-of-the art object detection algorithm for objects in sonar images.

• For this, we employed transfer learning to maximize data utilization and minimize training time.
• Our python-based deep learning framework enables both quick development and optimal model performance across platforms.
• We deployed the trainings pipeline and data storage for training data, models and artefacts in the cloud.
• This scalable and widely customizable cloud stack allows any number of clients to work in parallel. The resource deployment and management is fully automated.
• The deployment of the detection model occurs embedded in the survey Sonobots using edge-AI technology.
• Continuous delivery pipelines serve modular updates for both models and embedded devices.