The stabilisation algorithm is based on work by Cai and Walker (2008) described in Section 3.4. Local Motion Vectors (LM Vs) are created by estimating optical flow using the Lucas-Kanade method available in SIMULINK (Mathworks, 2009). The following value of parameters have been used: δ= 20% of the LVM array size,
α = 0.95, and β = 0.8. The β parameter was lower than the range suggested by Cai and Walker (2008) perhaps due to additional rotational motion of the camera which was not accounted for in this stabilisation method. More recent work by Cai and Walker (2009) addresses this issue but has not been part of this research. Section 9.4 (Further Work) will discuss possible improvements to the algorithm mentioned in this section.
5.8
Global Motion Patterns
There are situations where tracking of individual fish may not be possible at all in sea cages. This is most likely to occur when fish are larger, the density is higher and fish swim close to the camera. Also night time recordings, while using
artificial light, produce poor segmentation results and therefore poor tracking performance. However it would still be desirable to detect changes in fish be- haviour. Rather than attempting to track individual fish as described throughout this chapter, an alternative approach may be to look at spatio-temporal changes between a sequence of image frames, taking images as a whole rather than ex- tracting objects of interest.
To achieve this motion flow fields can be used (Hu et al., 2008) as described in Section 4.4. Rather than using Gaussian ART algorithm (Williamson, 1996) to decrease the number of motion vectors in the Point Flow Field, a simpler, yet effective method has been used. The original flow field was a matrix the size of the image frame - 160x120 (19200 motion vectors). Through nearest neighbour interpolation this matrix was resized to 20x15 (300 motion vectors) (Fig. 5.5i) and this is followed by creation of sink paths based on motion vectors. Figure 5.5ii demonstrates how the sink seeking processes traverses through points using the kernel based estimation. Once the sink seeking process is complete, supertracks are generated through sink clustering. Empirically, if Dx ≤17, and Dv ≥0.75 andDp ≤17 (calculated using Eq. 4.43, 4.44, and 4.45) then the sink
is assigned to the cluster, otherwise a new cluster is created. Figure 5.6ii shows five supertracks created as a result of clustering of sinks shown in Figure 5.6i.
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 (i) 0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 (ii)
Figure 5.5: Generation of sink paths: (i) Point Flow Field and (ii) the sink seeking process.
Supertracks from one sequence of images can be compared with another se- quence. This can be used to detect changes between consecutive sequences and the degree of change. Hu et al. (2008) provide metrics to perform supertrack matching and this could be used to detect differences in fish movement.
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 (i) 0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 (ii)
Figure 5.6: Generation of supertracks: (i) sink paths within the field and (ii) the result of sink clustering: supertracks. Short sink paths are discarded as they are likely to originate from points on the edge of the field.
5.9
Summary
The system described in this chapter allows tracking of individual fish within a sea cage for short periods of time. Using a number of fish tracks, an average speed and direction can be calculated for a given sampling period. Estimations of these tracks contain natural variation of fish movement as well as errors resulting from the tracking system. Only preliminary work has been performed to investigate how these system errors can be mitigated, especially the movement of the camera. Further improvements in the area of segmentation and camera stabilisation are required to increase the accuracy of the system. Despite these issues, the system provides useful data on fish movement in sea cages of similar quality to the manual observation but on continuous basis, providing a superior outcome comparing to any current video based method of observation. The next chapter presents several experiments, which provide some insight into how the system could be utilised in a real-time, commercial setting, and what sort of output is expected from it.
CHAPTER 6
ANALYSIS OF FISH BEHAVIOUR IN
SEA CAGES
6.1
Introduction
Video technology is used in aquaculture to observe fish behaviour during feeding (Ang and Petrell, 1998), at different times of day (Kadri et al., 1991), and in different seasons (Smith et al., 1993). Currently it is not possible to sample fish behaviours continuously due to the tedious and intensive nature of video analysis. To facilitate continuous analysis, a computer tracking system was investigated in Chapter 5 to automatically analyses fish movement in sea cages from the underwater video. The system was designed with commercial operations in mind and it was important to test its viability on a commercial aquaculture site.
The aim of this chapter is to demonstrate the results obtained from the track- ing system described in the previous chapter and analyse results in the aquacul- ture context. The chapter examines specific variations in swimming speed and direction within days and between days, aligns data with environmental variables and provide an interpretation of the data in this context. This interpretation is used to validate that the system can produce meaningful data for both aquacul- ture researchers and commercial aquaculture operators.