Fish sizing and monitoring using a stereo image analysis system applied to fish farming

Fish sizing and monitoring using a stereo image analysis system applied to fish farming

Engrneenng 14 i 1995 ljS- I73 Q 19Y5 Else&x Science L~mmxl in Great Britain. All rights reserved 0 I-LA-86OY/95/$9.50 + 0.00 Aquaculrural Pruned ELSE...

1MB Sizes 0 Downloads 100 Views

Engrneenng 14 i 1995 ljS- I73 Q 19Y5 Else&x Science L~mmxl in Great Britain. All rights reserved 0 I-LA-86OY/95/$9.50 + 0.00

Aquaculrural Pruned ELSEVIER

Fish Sizing and Monitoring Using a Stereo Image Analysis System Applied to Fish Farming B. P. Ruff, J. A. Marchant & A. R. Frost Image Analysis and Control Laboratory, Bedfordshire

Silsoe Research Institute, Wrest Park, Silsoe. MK45 JHS, UK

(Received 3 August 1993: accepted 1 February 1991)

ABSTRAC7 We describe and discuss a non-invasive 30 measurement technology based on optical stereoscopy and computer image analysisfor continuous monitoring of size, position, shape, and spatialorientation of single fish in cages containing many fish. We provide limited initial results to demonstrate that fish dimensions may be measured to millimetre accuracy and that fish may be tracked over limited time intervals to observe detailed movement. We discuss how this may be developed to monitor whole fish stocks in a Jish farm. Advantages and limitations of this technology are discussed in comparison to other currently available systems. The impact on aquaculture of this type of systemfor accurate and continuous monitoring is discussed.

INTRODUCTION Regular stock inspection and monitoring is a fundamental part of livestock husbandry. Visual inspection of the appearance and behaviout of animals can provide the farmer with a variety of information regarding their health and development. In particular it can give early indications of health problems and enable the farmer to estimate, and possibly regulate, growth rates and so predict how long it wiIl be before the stock is marketable. Visual inspections are usually supplemented by measurements of significant factors such as animal weight and food intake. These inspections and monitoring tasks are very difficult for a fish farmer particularly in the case of sea cage aquaculture. Observation from the surface, particularly during feeding, and the removal of fish with a 1.55

156

B. P. Rufi J. A. Marchant, A. R. Frost

dip net can provide some information, but sampling is infrequent and unlikely to be representative of the population. Furthermore, any operation which involves disturbing and handling fish, e.g. weighing them manually, can cause stress, or physical damage, and so make them more susceptible to disease, or retard growth by reducing appetite (Pickering, 1981). Visual inspection underwater may be made by divers or .by TV camera but this is costly in time and has limited benefit. The difficulty in observing and monitoring fish means that early signs of disease can be missed until a large proportion of the population is affected, and the opportunity for early remedial action is missed. Optimal feeding and harvesting strategies are difficult to implement due to inadequate information about growth rate and feed intake. There is a statutory requirement for terrestrial livestock to be inspected at least once a day for the welfare of the animals. The welfare of fish may not yet cause such significant levels of concern, but it is notable that the development of remote sensors for the diagnosis of disease and ill-thrift in farmed fish is rated as a research priority by an influential animal welfare advisory body (Farm Animal Welfare Council, 1993). Introducing remote observation and automatic monitoring technology into the fish farm could provide at least part of the solution for more effective stock management. In particular, such a technology should be able to provide accurate, detailed and regular dimension and position measurements to allow inference of growth, behaviour and ultimately disease. Computer image analysis coupled to a stereo camera system can provide a non-invasive method for remote monitoring of the size, shape, conformation, and detailed movement of fish. The purpose of this paper is briefly to review this and other remote monitoring technologies and to report on our findings in the application of computer image analysis to aquaculture.

METHODS OF REMOTE FISH MONITORING The most commonly used methods of remote monitoring of fish involve acoustic techniques. Echo sounders have been used at sea for at least 50 years to locate shoals of fish. During that time the quality of the information that can be extracted from acoustic signals has improved significantly (Rose, 1992). Echo integration techniques allow estimates of the total biomass of groups of fish to be made. Individual fish counting and

Fish monitoring using a stereo image analysis system

157

size estimation is possible, using multiple-beam echo sounding techniques, but requires greater spatial separation between fish than is found at the high stocking densities of farmed fish (I-ovik, 1986). A method of estimating the size distribution of a population of fish has been proposed based on a knowledge of the relationship between fish size and resonant frequency, but the accuracy of the technique was not proven (bvik, 1986). A recent commercial development (VAKI, 91) for stock assessment consists of a rectangular (approximately 05 m square}, planar matrix of light beams which are directed across a frame and, if they are not obstructed, are detected by sensors in the opposite side of the frame. This effectively produces a curtain of light beams which can detect objects passing through it. The frame, suspended for example in a fish cage, can therefore count fish passing through it. Also, if the velocity of the fish is known, with a knowledge of which beams were obstructed, and for how long, a three-dimensional (3D) representation of the fish can be constructed, from which fish dimensions and hence weight estimates can be made. The device could be used to sample a population by suspending the frame freely within a fish cage, and allowing fish to swim through voluntarily, or it could be used on a whole population by positioning it so that fish are obliged to swim through it as they are transferred from one cage to another. Equipment based on this principle should provide good dimensional information, although the resolution is limited by the spacing of the light beams and only single fish may swim through the frame at any one time. The behaviour of individual fish has been monitored using ultrasonic telemetry. Implanted transmitters have enabled fish to be tracked continuously within a dense population of salmon in a sea cage (JuelI & Westerberg, 1993) to an accuracy of 10 cm. Implanted transmitters have also been used to monitor heart rate which is an important physiological indicator (Bjordal et al., 1986; Holand, 1986). These techniques are valuable for behavioural and physiological research but have limited value for commercial fish monitoring because the need to implant transmitters restricts the number of fish that can be monitored. Conventional underwater video cameras can be used to visually observe a fish stock. They have been used, for example, in systematic studies of fish behaviour, in which video tapes were analysed manually to provide information on the frequency of particular behaviour patterns (Bjordal et al., 1986). However, a significant limitation of a conventional video system is that it provides only 2D information. This means that, for example, it is not possible to deduce the size of a fish from its image (a

B. P. Ruff; /. A. Marchant, A. R. Frost

158

given size of image may be due to a small fish close to the camera, or to a larger fish further away). Similarly, the spatial distribution of fish is not available from a conventional 2D image. The potential of automatic image analysis for remote monitoring of fish has already been identified (Balchen, 1986). Stereo image analysis offers the possibility of observing fish in three dimensions and measuring the numbers and detailed shape of fish at the high number densities encountered in commercial fish farms.

STEREO IMAGE ANALYSIS Stereo image analysis (Mayhew, 1991) requires two views of an object (a fish here) so that a point on the object in one 2D image may be matched with the corresponding point in the second image. Given a calibration of the optical arrangement of the system, then the x-y 2D Cartesian coordinates of these two points may be used to directly estimate the 3D Cartesian coordinates (x,y,z) of that point in a given world coordinate system. 3D scene measurement is thus a problem of analysing the stereo image pair to identify and accurately locate pairs of corresponding 2D points on the object’s surface. Figure 1 shows how a fish’s length may be measured by identifying its head and tail in a stereo image pair. The tip of the head and the point-where the body of the fish meek its tail fin have been identified as characteristic points on the fish for length

OkYU \

-.

Optic centre

Right Camera Plane

Feature Point at Head of Fish Fig. 1.

Head and tail points of fish projected onto camera planes.

Fish monitoring uing a stereo image analysis system

159

measurement. The two points corresponding to the head in the left and right images give a pair of coordinates (XL, yL ) and (xR,yR) respectively, from which the 3D position of the fish’s head is located. The tail’s 3D position is located in the same manner. The 3D separation of these two points provides a simple length measurement for the fish. The accuracy of the stereo analysis (Ruff, 1993~) is largely dependent on three features of the system: the spatial sampling frequency of the CCD sensors in the TV cameras used. which is typically 5 12 columns by 256 rows in each field of the interlaced image; the accuracy to which the image features can be found in the images; and the spatial separation of the cameras (i.e. the stereo baseline). Since the fish are moving then only one field from each TV frame is useful as the scene changes enough between the 0.02 s field sampling interval to introduce significant movement of image features between successive fields. In a simple stereo system, a human operator manually selects point features within a stereo image pair of a test object and the computer calculates the corresponding 3D positions and dimensions. A human operator is competent in selecting features for measurement but is liable to make random and biased errors in the exact position for those features. For monitoring a whole fish stock it is preferable to automate the point selection process so that dimension and position measurements may be continuously calculated and recorded to build up statistics on the stock. It is also important that these measurements are not reliant upon the random or biased error of human selection but should be made according to well-defined criteria in a computer program. However, automating the point selection process is far from simple. The image analysis techniques under development at SiIsoe to automate this process will be described elsewhere.

OBJECTIVES The objective of the research in this paper is to establish the practical accuracy of stereo image analysis in underwater measurement of live fish. In this paper we: (1) demonstrate techniques using optical stereoscopy and automatic computer image analysis for measuring individual fish dimensions and position; (2) investigate and develop experimental techniques for data acquisition, and calibration of stereo systems underwater; (3) track individual fish over short periods to give detailed descriptions of movement in three dimensions.

B. P. Rufi J. A. Marchant, A.

160

R.Frost

EXPERIMENTAL EQUIPMENT Stereo camera arrangement

The physical arrangement of the cameras for the stereo system is depicted in Fig. 2. The arrangement consists of two CCD cameras separated by a baseline B and verged to a point at a distance D from the cameras to give a working volume in which the subject is visible from the field of view of both cameras. The plan view of the working volume for stereo analysis is highlighted by cross-hatching in the figure. This area is delineated by the field of view of both cameras and the limit of visibility in the water. The focal length of the cameras, the stereo baseline, and the vergence angle define the geometry of the working volume for observation. In general, decreasing the focal length increases the working volume but reduces stereo accuracy. Increasing the baseline increases stereo accuracy by increasing parallax but reduces the working volume and can cause problems in locating and matching corresponding measurement points.

Right eamem emsof

Fig. 2.

Stereo optical arrangement showing section through the observation volume.

Fish monitoring rcsinga stereo image analysb system

161

Data acquisition

Images from the cameras are digitized and stored in memory on two frame-grabbing boards at rates up to 25 Hz. However, since these images cannot be processed at real time in our present system then they must be stored elsewhere and analysed off-line. Storage on conventional video tape is not an acceptable solution because this would cause a serious loss of pixel timing accuracy which would lead to unacceptable 3D position errors. In order to perform accurate stereo image analysis it is thus vital that the data produced by the cameras is recorded digitally. In our system this is accomplished by transferring one digitised field per image into an 8 MByte system memory area. This allows 32 stereo pairs to be recorded at a maximum rate of 5 Hz with our present system, or 60 images from a single camera to be stored at 8 Hz. Computer hard disk storage is possible and is desirable using commercial fast disk access electronics to directly drive the disk(s) but is not used in our system due to costs. Calibration

Accurate calibration of the stereo optical system is necessary before 3D measurements may be made. The method of Tsai (1986) is used to measure the intrinsic and extrinsic camera parameters that uniquely define the optical system. The calibration procedure (Ruff, 19936) requires that a target, whose dimensions are accurately known, is imaged by both cameras. The target used is a chequer board of 25 squares, each of dimensions 100.0 X 100.0 mm. This results in 16 internal comer points (see Fig. 3) with a guard perimeter of one square width to provide a controlled background for the boundary points. The well-defined geometry of the calibration target leads to a very robust and accurate method of automatic calibration in software. A stereo image pair of the calibration target underwater is shown in Fig, 4. The crosses show calibration points graphically overlaid onto the original image. Referring to the figure, it can be seen that the target shows strong perspective variation. This is necessary as the calibration algorithm requires a good variation in distance to target points. Laboratory arrrangement

The accuracy of stereo is known to depend on distance to the measurement point and the errors in measurement position are distributed unequally along the coordinate axes. To establish the practical

162

B. P. Rufi 1. A. Marchant, A. R. Frost

_ 1OOmm WV_

Fig. 3.

_

Diagram of calibration target.

upper limit of accuracy for length dimension measurement in a controlled environment, two sets of experiments were devised. In the first (see Fig. S), the calibration target was used not only for calibration but also as a stereo benchmark by measuring various dimensions on its surface using the stereo technique since points on the target could be located automatically and accurately in software. The calibration target was mounted on an optic bench with 3 degrees of movement measurable to O-1 mm and roughly aligned to the cyclopean (X, Y,2) coordinate axes whose orientation is shown by the dotted axes in the figure. The origin for the coordinate system is at the optic centre of the cyclopean camera position. The target lines were aligned with the X-Y axes to allow measurement of dimensions parallel and normal to the stereo baseline. The plane of the target lay within X-Y to isolate errors in X- Y from errors in 2. In the second set of experiments a model fish of known dimensions was measured using stereo. The model was mounted on the optic bench again with 3 degrees of translational freedom. A high contrast background was provided to facilitate manual selection of points on the images of the fish (to within + 1 pixel). Underwater arrangement

Underwater experiments were conducted in a 2 m diameter circular land-based tank containing sea-water. Two salmon of measured length

Fish monitoring using a stereo image analysissystem

163

Left image of underwater target

(b)

Right image of underwater target Fig. 4.

(a) Left and (b) right images of underwater target.

were placed in the tank and then observed using a stereo camera arrangement. The calibration target was also placed in the tank to allow in&l calibration of the optical system. Referring to Fig. 6, the cameras are mounted on a rigid platform which maintains the camera baseline, vergence angles, and relative

B. P. RUB 1. A. Marchant, A. R. Frost

164

z ____

Fig. 5.

Fi8h

Fig. 6.

cydopeall

Laboratory stereo equipment.

V-point

Arrangement of cameras in tank.

position. Currently the camera platform is a simple pole to which the cameras are rigidly connected. The pole is connected to a surface structure so that the cameras may be suspended underwater. The two CCD video cameras are connected to a video image acquisition system at the surface. The cameras are arranged so that the stereo baseline is approximately vertical. This results in a top-bottom camera configura-

Fish monitoring using a stereo image analysissystem

165

tion in which visual parallax is vertical, Fish are predominantly horizontally oriented, producing images with horizontally biased boundaries which are approximately normal to the stereo baseline in the top-bottom configuration. This arrangement is vital, from theoretical considerations, if best system accuracy is to be achieved. The lenses (6 mm focal length in air), stereo baselines (500 mm), and vergence angles (verged to 1 m) were chosen as a compromise between 3D accuracy and size of observation volume ( - 1 m3) in the limited dimensions of the experimental tank (2 m diameter by 1 m deep).

EXPERIMENTS All stereo experiments have four stages: (1) Calibrate the stereo system (performed once only for a particular optical arrangement). (2) Acquire stereo images of the subject. (3) Identify corresponding measurement points on the object(s). (4) Calculate 3D position of points using stereo analysis and infer dimensions, movement, etc., of object. Experiment 1 (establishing best practical stereo accuracy in laboratory)

Using the apparatus of Fig. 5, a series of experiments was performed to establish the accuracy of stereo dimension measurement parallel and normal to the stereo baseline. All measurements were performed using automatic analysis software which was capable of locating test points to sub-pixel resolution. The calibration target was maintained in approximate fronto-planar orientation to the conceptual cyclopean camera’s optic plane. This allows movements in X, Y and 2 (with reference to the cyclopean camera’s coordinate system) to be isolated for the purpose of establishing errors along each axis separately. Three experiments were then performed in which the target was moved along the X-axis, the Yaxis, and the Z-axis, and length measurements along X and Y separately were calculated. For each measurement position a total of 12 Xdimensions and 12 Y-dimensions were calculated from which the mean dimension and the variation about the mean in X and Y respectively could be determined. Experiment 2 (establishing best practical accuracy of fish length)

Using apparatus similar to that of Fig. 5, two experiments were performed using manuahy selected feature points on a model fish. A

166

B. P. Rug J. A. Marchant, A. R. Frost

model fish of known dimensions was mounted on the optic bench in the figure to allow movement in X, Y and Z to an accuracy of 0.1 mm. The axes of movement of the optic bench were aligned to the coordinate system of the virtual cyclopean camera. In one experiment the model’s long axis was oriented approximately parallel to the stereo baseline; in the other the model’s long axis was oriented approximately normal to the baseline. The position .of the model was varied in the Z direction. The boundary of the fish is difficult to locate so measurement points were calculated as the average 3D position for points chosen in the images just inside the model boundary and just outside the boundary at the nose and tail of the model. The length of the fish was then calculated as the 3D separation of the nose and tail feature points. Experiment 3 (measuring length of two real control fish in an image sequence)

Using the equipment shown in Fig. 6 a sequence of images was recorded showing a single fish swimming approximately 1 m away from the stereo camera arrangement. Image points at the nose and tail of the fish were located manually so that the 3D separation between them could be found. Maximum and minimum lengths were also calculated based on the measurement technique used in experiment 1 for model fish. These measurements were performed for each image pair in a sequence of images showing the fish performing a swimming motion of two to three tail movements. Two different fish were measured using this technique.

RESULTS Experiment 1

The results of the experiment show that errors in measurement vary with movement of the target in Z but not in X-Y. Errors in mean dimension measured parallel to the stereo baseline are greater than errors normal to the baseline but show a similar variation about the mean. The graph of Fig. 7 shows the mean measured length of the target square sides oriented in the normal direction to the stereo baseline (100.0 mm actual length) as target-camera distance is increased. At a distance of Z= 180.0 cm the X-dimension is measured at between 98.5 mm and 98.6 rmn and the Y-dimension at between 99.5 mm and 996 mm. The measurement technique clearly has an accuracy that varies with the orientation relative to the stereo baseline of the dimension to be measured.

Fish monitoring using a stereo image analysissystem

167

// I tftff XII

250

300

350

al

432

Z, mm

Fig. 7.

Mean normal target dimension

varied with distance Z.

The shape of the graph also shows that there is a small systematic error with distance in 2. The total error is similar in variation for parallel and normal measurements but the magnitude of the systematic error is greater for the parallel measurement, i.e. about 1.5 mm on 100.0 mm dimension at Z= 180.0 cm, as compared to the normal measurement error of 05 mm at 180-O cm. Computer models of the experimental arrangement were used to confirm and explore the systematic error. The source of the systematic error is not clear but is repeatable and suggests the possibility of introducing a correction factor to the results. Experiment 2

The graph of Fig. 8 shows the variation with distance of the length of the long axis of a model fish where the long axis is oriented normal to the stereo baseline. The actual length of the model was 232 mm f 0.5 mm. The graph shows a mean length of between 232-O mm and 234.0 mm, increasing slightly with depth with errors about the mean of about 1 mm. The vertical bars represent the maximum/minimum measurements made from which a length estimate was calculated. Maximum/minimum variation increases slightly with distance from 4 mm to 7 mm. A theoretical discussion of errors will not be given here. However, it should be

B. P. Rufi J. A. Marchant, A. R. Frost

168

240-

pe-

z-26261

180

XII

al

243

2%l

an 2,

Fig. 8.

III

PO

34)

360

:

cm

Model fish moved in Z showing length normal to stereo baseline.

noted that random errors in depth are mostly due to errors in point position parallel to the stereo baseline and that errors are magnified with increasing depth. Experiment 3 Experiment 3 uses the experimental setup of Fig. 6 to perform 3D measurements on two real fish whose dimensions have been established by manual measurement. The fish were observed over a sequence of images during l-2 s at an image sample rate of 5 Hz. This is sufficient to observe the fish during slow swimming motions. Points at the head and tail of the fish were selected manually on each image to allow 3D measurement of length and position to be made using the technique described in experiment 2. Length measurement The graph of Fig. 9 shows the variation length over time for one fish. In this case the fBh was observed over 18 successive frames to show a number of swimming motions starting with one large tail beat followed by an accelerating sequence of shorter beats. The vertical bars on the graph show the maximum and minimum length measurements for each

169

Fish monitoring using a stereo image analysissystem

, 0

02

114

06

1

W

,

I

/

,

I

,

1

12

14

16

18

2

I

22

I

24

,

26

2’e

3

3.2

34

Time base, seconds

Fig. 9.

Real fish 1 showing length during a swimming sequence.

frame from which the length of the fish was estimated (the solid curve). The maximum observed length during the swimming sequence, when the fish was straight, from the graph is 286 mm at tune= O-4 s, and the minimum length, when the fish was maximally curved, is 265 mm at time= 2.6 s. The actual length was 276 + 1 mm from manual measurements giving a measurement error of 10 mm or 3.6% in excess of the actual length. The second fish was measured to have a maximal length of 267 mm compared to a real length of 258 f 1 mm, an error of 9 mm or 3.5% above the true length. Further work is necessary to determine how much of these errors is due to a systematic source and how much is due to random error. However, the two length estimates have similar errors which are greater than is expected from the accuracy established in the laboratory. This leads to a tentative conclusion that systematic error accounts for most of the experimental error. Further trials will be needed to establish this, together with additional calibration procedures to correct for this error. Position measurement

The positions of the fish during their recorded swimming motion were determined by the 3D point half-way between the 3D points calculated

170

B. P. Rufi J. A. Marchant, A. R. Frost

for the head and tail of the fish. The movement of this mid-point has two components, conveniently orthogonal to each other. First, the fish has a translational component of motion which determines its gross movement. Secondly, the swimming action of the fish results in repetitive posture variations which cause the measurement point to oscillate normal to the direction of gross movement. This produces a component of motion that gives supplementary information about how actively the fish is swimming. The graph of Fig. 10 shows the variation over time of Z-position with X-position for a fish in left camera coordinates (effectively a plan view). This graph shows that the fish performed several tail movements which is shown in the undulation of the locus of movement. Each extreme point on the undulating locus indicates a point of maximal curvature of the fish’s body in its swimming action. As the fish accelerates during the sequence its flexing increases in frequency but with reduced amplitude. Figure 11(a) and (b) show the same fish at its straightest position (t= 0.4 s) and at its most curved (t= 2.6 s). DISCUSSION AND CONCLUSIONS Experiment 1 shows that, in a practical situation, stereo image analysis is capable of providing a non-contact mechanism for measuring 3D dimenm s3l

910

i

sE

En-

ii!

2

WJ-

N

870-

2 .u

am X displacement,

Fig. 10.

4m

@la

a

0

mm

Fish 2 locus of fish centre viewed in X-Z plane of left camera during a swimming sequence.

Fish monitoring using a stereo image analysissystem

171

Straight fish

(b)

Curved fish Fig. 11. (a) Straight and (b) curved fish.

sions of test objects with an error of less than 2 mm if image points can be accurately located. Mathematical modelling of the measurement technique and the calibration procedure of Tsai (a well-accepted calibration method) shows a systematic error in dimension measurement. The variation of this error with distance and orientation both in real and simulated experiments agrees closely and offers a mechanism for correcting this systematic error if it is first characterized during a calibration process.

172

B. P. Rufl J. A. Marchant, A. R. Frost

Without removing the systematic error, the length of real fish can be measured with an error of 35%. It is hoped that, with improved calibration, future experiments on real fish will achieve the 2 mm accuracy predicted with laboratory experiments. A more extensive series of experiments is planned to explore this using a population of 30-60 tagged fish. The correlation of fish length (and perhaps other dimensions) to mass will be investigated. The working volume’for observation by the stereo system used in the experiments was about 3 m deep by 1 m in cross-section, though this could be increased to some extent with corresponding decrease in stereo accuracy. This volume is insufficient to observe a complete fish stock at once though multiple cameras placed throughout a cage could be used to extend the observation volume. It is intended in future experiments to determine whether continuous observations of fish size in one or more parts of a cage can be used to estimate total biomass in the cage by applying statistical techniques on the number, density and sizes of fish observed in the limited experimental observation volume. The stereo method has been shown to provide highly detailed position and movement information in 3D at a high sampling rate. All fish visible to the cameras may be analysed simultaneously. This technology has the potential to accurately monitor biomass and enable investigation of the relationship between fish size and behaviour (e.g. feed uptake) in a fish farm which could lead to an improved management of the stock. Compared to other technologies the stereo method provides improved accuracy and detail of information at much higher data rates, allowing activities such as the swimming motion to be observed. This new improved accuracy and detail of information promises to support research into more detailed behaviour observation, particularly diseaserelated behaviour. This may eventually lead to automatic detection of disease in a fish stock.

FUTURE WORK The current research for image analysis applied to aquaculture at Silsoe is directed towards completely automating the 3D measurement process by developing techniques to locate individual fish in natural images taken in fish farms. Automating the point location has two important benefits: first, it allows continuous monitoring of the subject, and secondly, it allows interpolation methods to be implemented to improve point location accuracy which leads to improved accuracy in 3D measurement. Extensive experimentation is planned to investigate the practicality of the technology.

Fish monitoringusinga stereo image analysissystem

173

ACKNOWLEDGEMENTS The authors would like to thank Dr L. Ross and Dr W. Roy of the Institute for Aquaculture at the University of Stirling for supplying facilities for observing live fish at the University’s Machrihanish research station. This research is funded by the Agricultural and Food Research Council. REFERENCES Balchen, J. G. ( 1986). Bridging the gap between aquaculture and the information sciences. In Automation and Data Processing in Aquaculture. ed. J. G. Balchen. Pergamon Press, Oxford, pp. l-4. Bjordal, A., Floen, S., Fosseidengen, J. E., Totland, B., 0vreda1, J. T., Ferns, A. & Huse, I. ( 1986). Monitoring biological and environmental parameters in aquaculture. In Automation and Data Processing in Aquaculture, ed. J. G. Balchen. Pergamon Press, Oxford, pp. 1.5l-5. Farm Animal Welfare Council (1993). Report on Priorities for Animal Welfare Research and Development. FAWC, Surbiton, UK. Holand, B. A. (1986). Underwater telemetry as a tool in aquaculture research and development. In Automation and Data Processing in Aquaculture, ed. J. G. Balchen. Pergarnon Press, Oxford, pp. 177-80. Juell, J.-E. & Westerberg, H. (1993). An ultrasonic telemetric system for automatic positioning of individual fish used to track Atlantic Salmon (Salmo salar L.) in a sea cage. Aquacultural Engineering, 12 ( 1), l- 18. LBvik, A. (1986). Biomass estimation in aquaculture facilities. Modelling, Identification, and Control, 8 (l), l-9. Mayhew, J. E. ( 1991). 30 Model Recognition from Stereoscopic Cues. MIT Press, Cambridge, MA. Pickering, A. D. (ed.) ( 198 1). Stress and Fish. Academic Press, London. Rose, G. A. ( 1992). A review of problems and new directions in the application of fisheries acoustics on the Canadian East Coast. Fisheries Research, 14, 105-28.

Ruff, B. P. (1993a). Achievable accuracy using stereo image analysis for measuring fish dimensions. Internal document, SiIsoe Research Institute, Bedford. Ruff, B. P. (19936). Automatic calibration of a stereo optical system to high accuracy in a noisy environment. Internal document, SiIsoe Research Institute, Bedford. Tsai, R Y. ( 1986). An efficient and accurate camera calibration technique for 3D machine vision. Conf. Proc. IEEE Computer Society Conf. on Computer Vision and Pattern Recognition, 22-26 June 1986, Miami Beach, FL. VAKI-Aquaculture Systems Ltd (1991). The submersible biomass counter. Technical advertisement, VAKI-Aquaculture Systems Ltd, FAXAFENI 10, IS- 108 Reykjavik, Iceland.