Robotic fish tracking - integration of unmanned surface vehicle and acoustic fish telemetry

Autonomous vehicle systems and acoustic fish telemetry are research areas of strong tradition and merits at the Department of Engineering Cybernetics. The project aims to enable close integration of these areas to create novel platforms for robotic search, localisation and tracking of marine life such as migrating fish and other small and evasive underwater objects.

Our FishOtter-system is specifically designed for robotic fish tracking applications (Lauvås et al. 2022, Lauvås and Alfredsen 2023) and consists of three cooperating autonomous surface vehicles (ASVs) of the type shown in Fig. 1. The purpose of the system is to enable autonomous search, localisation and persistent tracking of acoustically tagged fish in unstructured inshore coastal environments such as river estuaries, fjords, and archipelagos. The ASVs are small, highly controllable differential thrust catamarans, each equipped with a time-synchronised omnidirectional acoustic telemetry receiver. The receiver enables reception, decoding and accurate time-stamping of signals emitted from miniature underwater acoustic transmitters of the type commonly used in fish tagging studies. By forming a dynamic and spatially distributed receiver array, the vehicle system can estimate and track the position of a tag in real-time using a localisation filter with measurements of the signal arrival times and the vehicles’ own position as input. The vehicles continuously attempt to maintain an optimal receiver geometry while avoiding obstacles and grounding by employing a collaborative multi-agent formation control algorithm.

Figure 1: ASV FishOtter.

Two projects are currently available related to this system:

1) Target movement prediction model

To facilitate miniaturisation and preserve battery-life, fish tags are usually implemented as intermittent transmitters, with transmission intervals ranging from a few 10s of seconds to several minutes. Once a reliable target position is established, tracking performance (localisation accuracy and prevention of target-loss) is highly reliant on the vehicle system’s ability to predict target movement between consecutive tag transmissions and relocate to an optimal receiver geometry. The tag localisation mechanism is based on an exogenous Kalman filter (XKF) with inherent state-space representation and estimation of target kinematics. However, it is not obvious what model structure gives the best representation and predictions of a moving fish target. The project will explore different model representations and evaluate their performance in computer simulations and in real-data tracking scenarios.

2) Robust fish localisation filter

Tag position is estimated using an exogenous Kalman filter (XKF) with updates based on signal time-of-arrival measurements. The filter relies on simultaneous input from a minimum of three vehicles/receivers to produce a measurement update of the target’s horizontal position. However, practical experiments have shown that triple receptions may be compromised by random noise events that cause double or even single detection scenarios, which are not exploited by the current measurement model. However, with a properly modelled and initialised localisation filter, also partial detections represent valuable information about the target position that may be utilized. This project addresses the task of developing an extended and more robust measurement model for the XKF that incorporates all detection scenarios, potentially also including signal SNR measurements. Filter performance will be investigated in both computer simulations and with field data.

Ultra-low power embedded design projects

Some projects can be specified for students particularly interested in embedded hardware design (circuit design, layout and production of dedicated hardware), notably power-constrained/battery-operated solutions such as wireless sensor nodes/IoT. Several alternatives exist and can be discussed (contact supervisor for more information).

Fish farm behavioural data analytics

The goal of this project is to explore how unsupervised machine learning techniques can be applied to make inference from fish telemetry data from large-scale fish farms.

Background: Despite today's highly efficient salmon production, the potential for further optimisation is significant if more profound knowledge on the underlying behavioural processes that take place in the sea-cages could be obtained. Growth and development of a salmon population in a fish farm are ultimately a product of processes occurring at the individual level. However, fish behaviour in large-scale salmon farms depends on complex and partly unknown relationships involving environmental conditions, farm operations, the fish’ condition and physiological state, as well as social predispositions and interactions.

Figure 2: Ocean Farm 1.

Detailed insights into individual responses to different management regimes, operations, and environmental conditions, and not least the variations in such responses, would make it possible to increase the precision in farm management practices for better health, growth and welfare. Due to the dynamic nature of the aquaculture process, it would be particularly useful if such information could be captured in real-time, systematised, analysed and operationalised through advanced decision-support systems. This form of precision animal husbandry has been under development for many years under the name Precision Livestock Farming in land-based animal husbandry, and generally in agriculture as Precision Agriculture. A similar development is foreseen within aquaculture through the Precision Fish Farming paradigm (Føre et al. 2018).

Acoustic fish telemetry represents a unique method for observing and acquiring individual behavioural data histories from free-ranging fish. The method relies on equipping fish with miniature acoustic sensor transmitter tags and a collecting tag data over time using a dedicated receiver system. Fish telemetry has previously been applied to farmed salmon in commercial-scale production facilities such as Ocean Farm 1 (Fig.2) creating large datasets on individual behavioural histories along with environmental and production-related data. The goal of this project is to explore how unsupervised machine learning techniques can be applied to make inference from these datasets.