6.4 Discussion
7.3.2 Sequence 2
The second sequence demonstrates manoeuvrability of fish and the system’s abil- ity to track these manoeuvres. The sequence is shown in Figure 7.8 and points lighter in colour are at the beginning of the sequence while darker points are to- wards the end of the sequence. There are two main interactions in the sequence: the first being a manoeuvre and change of speed by the Green fish, the second one being a circular movement by the Red fish along the sides of the tank while the Green fish is stationary. Only MDA data association has been considered in this sequence as differences between GNN and MDA have been addressed in Sequence 1.
(i) Ground truth
(ii) MDA data association
Figure 7.8: Tracking results for Sequence 2. Strength of the colour determines the temporal order. Lighter points are at the beginning of the sequence, darker points are at the end of the sequence.
Complex Manoeuvres
This manoeuvre starts with the Green fish swimming from the bottom of the image to the top of the image, towards the Red fish (Fig. 7.9). When the Green fish reaches the top of the image, there is an interaction with the Red fish (the Green fish touches the tail of the Red fish). The Green fish turns around and both fish begin to travel to the bottom of the image. As Red fish reaches the bottom and starts to move towards the left side of the image along the wall of the tank, the Green fish suddenly accelerates, performs a turn and starts following the Red fish.
The tracking system copes well with the sudden change of direction by the Green fish, as the Particle filter can deal with non-linear motion. The filter also deals with acceleration of the Green fish but activation of the Coupled Partition component generates inaccuracy in tracking. The system recovers and correctly associates both targets at the end of the sequence.
Passing a Stationary Fish
This case looks at the Red fish swimming around the tank while the Green fish is stationary. In Figure 7.10, the Red fish begins its circling motion from the top right of the image, approaches the stationary Green fish and then continues along the wall of the tank downwards and to the right hand side of the image. The sequence continues in Figure 7.11 where the Red fish goes past the Green fish for the second time. The Green fish remains essentially still; its slight motion is mainly to maintain its station in relation to the current.
The system has no trouble tracking both fish while the Green fish is stationary and the Red fish is moving and passing the Green fish. There is a loss of track in Figure 7.11ii but the tracking filter recovers without issues.
(i) Ground truth
(ii) MDA data association
(i) Ground truth
(ii) MDA data association
(i) Ground truth
(ii) MDA data association
7.4
Discussion
The use of a target template within the tracking system has the advantage that it utilises the prior knowledge about possible targets, and the use of histograms makes the template shape invariant. The disadvantage of this method is that the template has to be created (manually or automatically) with particular re- gard to different levels of illuminations within a tank. Making illumination levels even throughout the tank is desirable but not always possible. Therefore exper- imental design can impact on the ability to detect fish in a tank. Estimating the background can improve tracking but again care needs to be taken during experimental design to reduce the number of reflections through appropriate po- sitioning of lights. An ideal recording setup would involve a wide lens camera in the middle of the tank just under the surface looking down. This would eliminate the surface movement (and thus reflections).
The results in Section 7.3 demonstrate the ability of the particle filter is in tracking multiple targets, at least in short video sequences. It can track fish during unpredictable movement and change of direction. The data association works successfully when fish pass each other without major changes in speed or direction. However, it is questionable if the filter could maintain its correct data association in longer video sequences. Once an error has been made it would be impossible for the system to recover from it. Therefore an additional correction mechanism would be required which would periodically ensure that fish have correct identifications. One way would be to tag fish using RFID technology. RFID tags are very small and therefore their impact on fish might be minimal. The technology would allow a system to receive periodic readings as fish approach the scanner and readjust identifications if necessary. While using tagging to complement video may be successful, the purpose of this research was to remove the need for tagging fish. Therefore an alternative might be for the tracking system to identify unique features of each fish. This may only be possible with a high-resolution camera used to take periodic shots of fish in order to identify them.
The performance of the coupled partition scheme is also questionable as at times the filter tends to converge into a single mode. It works well when targets move with constant speed and opposite direction but when they manoeuvre while the coupled partition is active this scheme performs poorly. For that reason step 6 and 7 of the pseudo-code, in section 7.2.2, contains a counter variable which is used to allow only 3 iterations of the original coupled partition scheme. After that the importance weight calculation includes distance between partitions favouring those further away from each other. This generally prevents the filter from converging into a single mode and enabling it to remain multi-modal.
Further improvements to the system could involve the use of elliptical par- ticles or the use of multiple bounding boxes to describe each fish. This second possibility could be used to track movements of the body and the tail of each fish.
The next chapter examines the possibility of tracking a slightly larger group of fish in a tank with focus on group dynamics rather than individual behaviours.
CHAPTER 8
TRACKING A SMALL GROUP OF
FISH IN TANKS
8.1
Introduction
The tracking system described in Chapter 7 attempted to track two fish in a tank while always being able to uniquely identify both fish. This proved a difficult task to achieve without invasive methods, including tagging. Another approach to tracking of fish in tanks is to track individual fish as part of a group without unique identification. Instead of observing individual behaviours, the focus is on the group behaviour. This chapter will examine if it is possible to track a small number of fish in a tank for the purpose of identifying group behaviours.
The aim of the chapter is to determine what type of data the tracking system can produce and demonstrate how these data could be used in experiments which require observations of different groups of fish in tanks. Its suitability in feeding experiments will be examined. Similarly the system could also be used in disease challenge experiments to observe differences between diseased fish and a control group.
8.2
Materials and Methods
This chapter examines behaviours of a group of 10 fish (rainbow trout) in a tank and how they interacted as a group in relation to a demand feeder, generating daily activity patterns. While individuals were being tracked, there was no at- tempt made to uniquely identify them. Instead the spatio-temporal activity of fish was examined: preferred areas of occupation, activity at different times of the day, and feeding vs non-feeding activities. The experiment was set up to test the computer vision techniques rather than trying to answer a specific aquaculture question. It was important to find out which parameters of fish behaviour could
be detected by the tracking system. Daily graphs of average swimming speed, distance of fish from the feeding area and location from the centre of the tank are presented in Appendix C.