A Fully Automated Monitoring Tool to Detect Lameness of Broiler Chickens by Image Analysis

Transcription

1 A Fully Automated Monitoring Tool to Detect Lameness of Broiler Chickens by Image Analysis * Arda AYDIN 1, Anıl ÇAY 1, Claudia BAHR 2, Daniel BERCKMANS 2 1 Dept. of Agricultural Machinery, Çanakkale Onsekiz Mart University PO Box 17020, Merkez, Çanakkale, TURKEY 2 Dept. of Biosystems, Division M3-BIORES: Measure, Model & Manage Bioresponses, Catholic University of Leuven PO Box 3001, Kasteelpark Arenberg 30, Heverlee, BELGIUM araydin@comu.edu.tr Abstract: In this study, a new and non-invasive technique is developed to automatically analyse the locomotion and posture behaviours of broiler chickens with different gait score levels. Five experiments were carried out with in total 250 broiler chickens in order to assess the relationship between the gait scores obtained by human experts and the locomotion and posture behaviours quantified by an automatic image monitoring system. For this purpose, video surveillance images of broilers with five different predefined gait scores were recorded and by applied image analysis techniques the bird's locomotion and posture behaviour was classified into two categories: the number of lying event (NoL) and latency to lie down (LTL). A linear regression was performed to define the coefficient of determination between the NoL obtained by the proposed algorithm and the NoL obtained by manual labelling, which resulted in R 2 = 0,993. In addition, 83% of lying behaviour was correctly classified by automatic monitoring system for a total of 250 broiler chickens. The results also showed a clear correlation (R 2 =0.893) between NoL and gait score and a strong negative correlation (R 2 = , P < 0.001) between LTL and gait score level of broilers. Key words: Locomotion, broiler chickens, gait score, lameness, automatic monitoring INTRODUCTION The broiler chicken industry has grown consistently during the last 50 yearsthe most important questions related to broilers welfare quoted in the last two decades are the increasing sensitivity for metabolic and locomotion problems due to fast growth rates and inactivity of the chickens (Bauer et al. 1996, Bessei et al 2006 and Cangar et. al. 2009). Lameness is a broad term used for a series of damages to broiler chickens of infective and noninfective origin ((Riddell, 1991; Thorp. 1996; Sorensen et al., 2000). Losses due to skeleton disorders in broiler chickens are significant (Cook, 2000). In some stables it was observed that at least 90 per cent of the chickens experienced gait problems to some extent at slaughter age (Kestin et al., 1992) and approximately 30 per cent of the chickens were seriously lame (Sanotra et al., 2001). In 1998, the costs of these chickens skeletal disorders for the USA were estimated to be between 80 and 120 million dollars per year (Bradshaw et al., 2002). The occurrence of lameness is thought to be strongly correlated with weight and growth speed (Vestergaard and Sanotra, 1999). Moreover, locomotion problems may be painful to the animal and decrease their mobility while increasing secondary problems, such as hock burns and chest dirtiness. (Weeks et al., 2000). A new method, latency-to lie down test (LTL) for assessing the severity of lameness in broiler chickens was described by Weeks et al., (2002) as the length of time that birds remained standing in shallow water. It was measured and the results were compared with the results of conventional gait scoring. There was a highly significant (P<0-001) relationship between the LTL and birds' gait scores (Weeks et al., 2002). Due to the original testing procedure, in which the birds are tested in groups, involves a certain settling period, which makes the test time-consuming to perform on commercial broiler farms, a new test was designed by Berg and Sanotra, (2003) to record the LTL. The main differences and/or advantage of this new test that the birds were tested individually without visual contact with other birds and 137

2 the experimental setup could be carried between commercial farms. The results of their study also showed a clear negative correlation (r = -0.86, P < 0.001) between time spent standing and gait score. However, these kinds of existing tests are time consuming and the measurements cannot be performed continuously. As consequences, there is no any chance for early detection of lameness before occurred with these manual evaluation methods. Especially, to perform this kind of manual tests in the big commercial farms which has more than chickens in a broiler house is necessary to spent excessively manpower. As an alternative to these manual evaluation methods, the increasing availability of low cost technology currently makes automated monitoring of animal behaviour feasible. Leroy et al. (2005) established a model-based computer vision system to study the behaviour of hens in furnished cages. Individual behaviours such as standing, walking and scratching could be recognised automatically and in real time. Furthermore, investigating the locomotion behaviours of broiler chickens in relation to gait score can serve as a measure for lameness (Bokkers et al. 2006, Aydin et al. 2010). A major advantage of this type of automated behaviour monitoring is that the measurements can be made continuously throughout the life of a flock, they are fully automated, a completely non-invasive and non-intrusive way and do not involve the biosecurity risk of having people visiting different farms to perform gait scoring (Dawkins et al 2009). The first objective of this study is to investigate locomotion and posture behaviours of broiler chickens (the total number of lying events and the duration of the latency to lie down of broilers) in relation to their gait scores using an image based monitoring system under laboratory conditions. By the use of this system it is possible to automatically classify behaviours relevant for lameness assessment in broiler chickens. The second objective of this study is to serve as an additional method for developing an automatic lameness monitoring tool for chickens with different gait scores. By the combine of this method with other systems it is possible to develop an automated lameness monitoring tool with higher accuracy. MATERIALS and METHOD Experimental Design, Video Recordings and Birds The experimental setup consisted of a wooden test corridor, with dimensions 2,40m (length) x 1,00m (width) x 0,50m (height). A digital video camera, Guppy F036C equipped with a C30811KP 8,5mm lens (Pentax) was mounted 2.0 m above the ground with its lens pointing downwards and directly above the center of the corridor to get a top view of the walking area in the camera image. The camera was connected to a PC with a built-in frame grabber (E U, AWM 20276, VW-1) using an IEEE 1394 fire wire cable. Images were captured with a resolution of 1024 x 768 pixels at a sample rate of 3.5 frames per second. Video recordings were made during 5 experiments. Five experiments were carried out with in total 250 broiler chickens (Ross-308) which were obtained from the Provincial Centre for Applied Poultry Research, Province Antwerp (located in Geel, Belgium). The chickens were scored and selected according to their degree of lameness by experts using to the method of Kestin et al. (1992). Based on Kestin et al. (1992), lameness of the chickens was ranked by experts in increasing order from gait score zero (GS0) to gait score four (GS4) where GS0 is being the healthiest. GS5 chickens were not used in the experiments due to the severity of lameness; these birds were unable to walk. Per experiment, 39 day-old broiler chickens were chosen in such a way that there were 10 samples from each gait score. Chickens were taken from a compartment of 1500 for each of the five experiments. During each experiment, a chicken was placed at the start point in the test corridor and video images of the walking area were recorded while the chicken was walking from the start point till the end point of the corridor during 2.4 meter. This procedure was repeated with all 250 chickens. Image analysis As basic image analysis technique of background subtraction for the segmentation of the shape was used. This technique was applied because the camera setup was fixed and hence the background remained constant over time. Segmentation was performed by 138

3 subtracting a background image of the empty corridor from each recorded image of the corridor containing a chicken. The pixel for which the difference is above a certain threshold was defined as belonging to the shape of the animal (Leroy et al. 2005). After this process, the shape of the animal could be characterized with a set of measurable parameters, such as the centre of the shape mask, the area of the shape mask (Minagawaha et al., 1994, De Wet et al., 2003). Furthermore an image processing algorithm was used to extract a chicken from a sequence of video images and by fitting an elliptical shape around animal the centre point and orientation, length and width of the animal in the image could be defined Leroy et al. (2003a, 2003b, 2005). Elliptical shapes are simple but widely applicable as an approximation of natural shapes (Birchfield, 1998) and its shape can be altered by varying only five parameters: (xc, yc, α, a, b) respectively, xc background subtraction were calculated in the first image of each video sequence and given to the programme. The optimal value of the shape parameters (xc, yc, α, r1, r2) for each image was labelled as posture parameters and stored for further processing (Leroy et al. 2005). When a certain type of behaviour occurred in the camera image, this caused a distinctive pattern in a number of successive values of the posture parameters. The posture parameters for each image were computed and the past values within a certain time window were considered, so that the window could hold the entire pattern (Leroy et al. 2005). To model the dynamic trajectories of the posture parameters within the time window, a first order transfer function (TF) model was used (Young, 1984, Aerts et al., 2000). Fitting this function to the data within each time window resulted in a set of two dynamic parameters a, b for each posture parameter (Leroy et al. 2005).Table 1 summarises the Once for initialisation, the centre point, position, orientation and sizes of the chicken mask obtained from dynamic variables that were extracted. Table 1. Dynamic variables extracted from the video sequence of the chickens and their description. Variable name Description Units x-y coordinates x-y coordinates of the centre of the animal in the pen as a function of time m Walking trajectory Subsequent positions of the animal s mass centre in x- and y-coordinates m Orientation Subsequent angles of the chicken with respect to the horizontal axis in the image as a function of time α (degrees) Back surface Top view area of the chicken m 2 Classification of locomotion and posture behaviours The classification procedure involved the variables: maximum radon, orientation change, x y coordinates and back surface area of the chicken. These variables were analysed applying a sliding window approach. The chicken s behaviour was classified as lying (lying) at a period moment if during the past window size (3.5 frames per second): 1) The slope of the cumulative distance walked is less than a certain threshold 2) The x-y coordinates of the geometric centre of the animal is stable meaning that the fluctuations stayed in a certain stability range expressed in percentage. 3) The filtered back surface variable of the animal (m2) exceeded a certain threshold (Cangar et al. 2008) If these conditions were fulfilled the chicken s behaviour was classified as lying. The resulting output of the method consisted of the animal s position, orientation and body configuration as a function of time. Using these outputs, a distinction between lying and standing was done automatically. Furthermore, latency to lie down (LTL) of broiler chickens were calculated. As opposite to previous studies, in this study, any kind of disturbing factor such as water was not used to measure LTL of broiler chickens. Experiments were performed in a commercial farm and manual labelling of lie down events and the duration of the latency to lie down was performed by an 139

4 expert during the experiments. Afterwards, the results of the proposed algorithm were compared with manual labelling results. Statistical Analysis The statistical analysis was carried out on 50 video data sets per experiment with 10 data sets belonging to each of the gait score groups. A total of 250 data set was used to investigate the differences in locomotion and posture behaviours between the gait score groups. The Friedman test, which is a non-parametric test that compares the columns without the row effects, was used to analyse the effects of gait score on birds lying behaviours. In the test sample, size and dependencies do not affect the test results. After the Friedman Test, the Dunn test was applied to define the statistical differences between the gait scores. The Dunn post test compares the difference in the sum of ranks between two columns with the expected average difference (based on the number of groups and their size).the calculations were performed using the Statistics Toolbox of Matlab (The Math Works, Massachusetts, USA). RESULTS and DISCUSSION Classification of locomotion and posture behaviours This automated monitoring tool enabled to measure the body variables back surface, centre point and body contour (Table 1). A change in orientation (rotation angle around the horizontal axis) was a clear indication of the activity of the animal. In some moments although the centre point of the animal did not change, its orientation changed which signified a clear movement but no displacement (See figure 1) the chicken are plotted as a function of time. X-y coordinates indicated the specific position of the chicken in the corridor at a specific time. Less variations in x-y coordinates showed that the movement of the chicken was limited. During those periods the chicken was either standing and not moving or lying down. The acceleration (m/s 2 ) was another representation of the chicken s movements. Not surprisingly, the speed was approximately zero during lying periods. Variations over time in x-y coordinates, and speed greater than zero, signified that the chicken was moving. This movement could be an indication of walking or could be indicative of lateral movements while in the lying position. The percentage of correctly classified lying behaviour for 250 chickens can be seen in table 2. Especially x-y position, back surface and acceleration had a strong correlation with the manual labelling of the lying and standing behaviour. When the slope of the cumulative distance was high, the animal was standing and moving. On the other hand, when the slope was close to zero, the chicken was lying or standing still. While lying the back area was greater than while standing or walking. Compared with manual labelling, the image analysis method correctly classified lying behaviour of 250 chickens with an average of 83%. A linear regression test was performed to define the coefficient of determination between the number of lying events obtained by the proposed algorithm and the number of lying events obtained by manual labelling, which resulted in R 2 = 0,993. Afterwards, the relation between the latency to lie down (LTL) obtained by algorithm and LTL obtained by manual labelling was investigated and the coefficient of determination (R 2 ) was found to be Figure 1. The ellipse and the center point of roiler chickens. The x-y coordinates and the speed of the centre point, together with the orientation of the main axis of Assessment of lying behaviour in relation to gait score The results of the algorithm were statistically analysed for differences between the different gait score levels. As shown in Table 3, the number of lying events (NoL) of GS3 and GS4 (mean ± standard deviation) were significantly (P<0.001) higher than the ones of GS0, GS1 and GS2. 140

5 Table 2: Correctly classified lying behaviour using image analysis No Exp ) NoL (Algorithm NoL (Manual Labeling) TP FP FN acy. Accur (%) Avg Besides that there are no significant differences in the number of lying events (NoL) among GS0, GS1 and GS2. Additionally, the LTL was evaluated and results are displayed in Table 3. Lame chickens of GS3 and GS4 (mean ± standard deviation) sit down significantly (P<0.001) earlier than the ones of GS0, GS1 and GS2. Considering the relation between NoL and LTL, the results show a high correlation (R 2 = and P<0.001) between the two. (Figure 2). Table 3. NoL and LTL of broiler chickens with different gait scores. Gait Scores NoL (Mean+Std) LTL (s) (Mean+Std) 0 16±2 a 28.69±11.30 a 1 17±1 a 25.27±09.43 a 2 16±1 a 23.56±07.86 a 3 34±2 b 11.15±05.46 b 4 43±3 c 03.33±01.75 c * a,b,c Mean ranks, within a column, with no superscript in common differ significantly (P < 0.05). The minimum and maximum LTL values recorded were s for gait score 0, s for gait score 1, s for gait score 2, s for gait score 3, and s for gait score 4 NoS-LTL GS0 GS1 GS2 GS3 GS4 Gait Scores NoS (Algorithm) NoS (Manual Labelling) Figure 2. Correlation between gait scores, lying (nummuer of sitting) and LTL behaviours of broilers and comparison of algorithm output against manual labelling. The relation between lying and gait score is shown in the correlation between NoL, LTL and gait score (Figure 2). The analysis revealed a strong positive correlation (R 2 =0.893) between NoL and gait score and a strong negative correlation (R 2 = , P < 0.001) between LTL and gait score. A novel technique was developed using computer vision to automatically monitor the gait variables and coordinates from top view images of broilers. It provided broad information based on the centre point, body contour, walking trajectory and back surface of broilers. The obtained variables were then used for classification of locomotion and posture behaviours such as lying events and latency to lie down. These classified behaviours were then compared with manual labelling of experts. It was found that on average 83% of the lying behaviour of the 250 chickens during the experiments could be correctly classified. On the other hand, the results of this study also showed a clear correlation (R 2 =0.0893) between gait 141

6 scores and lying behaviour of broiler chickens similar to the results of Weeks et al., (2000). As concluded in the study of Weeks et al., (2000) sound broilers averaged 76% of 23 hours lying and this increased significantly to 86% of 23 hours in lame birds. Based on the well-known latency to lie test, a lame chick would spend a longer part of the day lying and it tends to sit much faster than the sound one. The automatically extracted LTL was evaluated and the results showed a similarity with the results of Weeks et al., (2002); the lame birds (GS3 and GS4) sat down significantly (P<0.001) earlier (03.33±01.75 sec.) than the sound birds. This study also indicates that lame chickens tend to have a lower latency to lie down than non-lame chickens. Berg and Sanotra, (2003) found a clear negative correlation (r = -0.86, P < 0.001) between LTL and gait score. Similarly, in this study, a strong negative correlation (R 2 =-0.954) was found between the LTL and gait score level of broiler chickens. Comparable results were also found by (Dawkins et al., 2009) with gait scores highly negatively correlated with the percentage of time chickens spent walking. Additionally to the previous studies, the NoL of broiler chickens was investigated in this study and the analysis revealed a strong positive correlation (R 2 =0.893) between NoL and the gait score levels of broiler chickens. Further, the results of the algorithm were statistically analysed for differences between the different gait score levels. The NoL of lame chickens (GS3 and GS4) were significantly (P<0.001) higher (34±2 and 43±3) than the NoL of sound chickens (GS0, GS1 and GS2). Although, there were strong correlations among NoL, gait score and LTL, there were no significant differences between the sound chickens (GS0, GS1 and GS2) regarding the NoL and LTL. In this study, the broiler breed Ross 308 was used exclusively to produce comparable data. The results and conclusions of this research apply to the behaviour of Ross 308 chickens as the most common breed in Europe. The lying behaviour may be different in other breeds or genetic lines. The classified behaviours were compared with manual labelling of experts. Strong correlations were found between the outcome of the algorithm and manual labelling, leading to the conclusion that the algorithm produces reliable results. However, the correct classification of lying down reached on average 83% indicating room for improvements. CONCLUSIONS The followings were concluded from the study: This research focused mainly to investigate the relation between automatically classified locomotion and posture behaviours and the gait scores of broiler chickens to assess the lameness of broilers. 83% of lying behaviour was correctly classified by this automatic monitoring system for a total of 250 broiler chickens. The system has potential but needs further optimisation to improve the classification and it needs validation in different field conditions, on different types of chickens and on a larger sample size of broilers. As also concluded by Rushen et al., (2012), to more accurately identify the effects of gait score on broiler behaviour, this automatically obtained locomotion and posture information can be combined with other automatic behaviour analysis systems such as measuring the activity levels of chickens to detect the degree of lameness (Aydin et al 2010) and/or detecting the optical flow patterns in broiler chicken flocks as suggested by Dawkins et al. (2009 and 2012). REFERENCES Bauer, m., heissenhuber, k., damme, k. and köbler, m. (1996) Welche roilerherkunft eignet sich? DGS Magazin 44: Berg, C., Sanotra, G.S., Can a modified latency-tolie test be used to validate gait-scoring results in commercial broiler flocks? Animal Welfare 12, Bessei, W., Welfare of broilers: a review. Worlds Poultry Science Journal 62. Blokhuis, H.J., Van Der Haar, J.W., The effect of the stocking density on the behaviour of broilers. ZADI (Germany, Federal Republic of) v. 54(2) p Bokkers, E.A.M., Koene, P., Behaviour of fast- and slow-growing broilers to 12 weeks of age and the physical consequences. Appl. Anim. Behav. Sci. 81, Bokkers, E.A.M., Zimmerman, P.H., Bas Rodenburg, T., Koene, P Walking behaviour of heavy and light broilers in an operant runway test with varying durations of feed deprivation and feed access. Ethology Group, Department of Animal Sciences, Wageningen University, P.O. Box 338, 6700 AH Wageningen, the Netherlands Applied Animal Behaviour Science 108, Cangar, O., Cardinaels, S., Everaert, N., De Keteleare, B., Bahr, C., Zoons, J., Decuypere, E., Berckmans, D A study on the cause and effect of Lameness on broiler 142

7 chickens. Precision Livestock Farming Wageningen Academic Publishers, Netherlands, Cangar, O., Leroy, T., Guarino, M., Vranken, E., Fallon, R., Lenehan, J., Meed, J., Berckmans, D Automatic real-time monitoring of locomotion and posture behaviour of pregnant cows prior to calving using online image analysis. Computers and Electronics in Agriculture 64, Dawkins, M.S., Lee, H,J,, Waitt, C.D., Roberts, S.J Optical flow patterns in broiler chicken flocks as automated measures of behaviour and gait. Applied Animal Behaviour Science 119(3-4), Dawkins, M.S., Cain, R., Roberts, S.J., Optical flow, flock behaviour and chicken welfare. Animal Behaviour 84, Kestin, S.C., Knowles, T.G., Tinch, A.E., Gregory, N.G., Prevalence of leg weakness in broiler chickens and its relationship with genotype. Vet. Rec. 131, Kestin, S. C., Gordon, S., Su, G. and Sorensen, P Relationships in Broiler Chickens between Lameness, Live weight, Growth Rate and Age. Vet. Rec. 148(7): Leroy, T., Vranken, E., Van Brecht, A., Struelens, E., Janssen, A., Tuyttens, F., De Baere, K., Zoons, J., Sonck, B., Berckmans, D., A quantitative computer vision method for on-line classification of poultry behavior in furnished cages. Trans. ASAE 49 (3), Leroy T., Vranken E., Van Brecht A., Struelens E., Sonck B., Berckmans D. 2006a. A computer vision method for online behavioral quantification of individually caged poultry. Transactions of the ASABE 49 (3): Riddel, C Ascites and right ventricular failure in broiler chickens and ducks. In: Diseases of Poultry, 9th Edition Iowa State University Press: 839. Thorp, B.H. and Dduff, S.R.I. (1988) Effect of exercise on the vascular pattern in the bone extremities of broiler fowl. Research Veterinary Science 45: Weeks, C.A., Danbury, T.D., Davies, H.C., Hunt, P., Kestin, S.C The behaviour of broiler chickens and its modification by lameness. Department of Clinical Veterinary Science, University of Bristol, Langford, Bristol BS40 5DU, UK Applied Animal Behaviour Science 67, Weeks, C.A., Knowles, T.G., Gordon, R.G., Kerr, A.E., Peyton, S.T., Tilbrook, N.T., New method for objectively assessing lameness in broiler chickens. Veterinary Record 151,