behavioral analysis system for animal husbandry
The described system addresses the inadequacies of existing health monitoring in pig farms by using image analysis and neural networks to detect swine flu through lying behavior, enabling timely intervention and reducing economic losses.
Patent Information
- Application Number
- FR2021001754
- Authority / Receiving Office
- FR · FR
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2021-02-08
- Filing Date
- 2021-02-23
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2041-02-23
AI Technical Summary
Existing systems for monitoring animal health in pig farms, such as those based on sound recordings and deep learning for respiratory diseases, are inadequate for detecting enzootic diseases like swine flu, which cause recurrent infections and result in significant economic losses and antibiotic use, and there is a lack of effective methods for monitoring clinical signs of such diseases.
A monitoring system using an electronic box with a camera and data processing unit that integrates an image analysis algorithm to detect animals in a lying position, calculating the ratio of inactive animals to total animals, and utilizing a neural network to identify anomalies indicative of respiratory diseases like swine flu.
The system effectively detects the onset of swine flu by identifying increased lying behavior of animals, allowing timely intervention and reducing antibiotic use and economic losses by accurately monitoring the health of pig populations.
Smart Images

Figure 00000012_0000 
Figure 00000013_0000 
Figure 00000013_0001
Abstract
Description
Title of the invention: Behavioral analysis system for animal husbandry
[0001] The present invention relates to the field of monitoring animal behavior in their breeding location, and in particular to monitoring behaviors associated with pathologies in pig farms.
[0002] Traditionally, staff are assigned to breeding sites to ensure the feeding and care of the animals in the facilities. Monitoring systems are also implemented to allow remote control of the building housing the livestock.
[0003] Monitoring the health status of farm animals can, for example, be implemented using a system capable of detecting the onset of respiratory disease during fattening phases by recording and analyzing cough sounds. Such systems allow for monitoring the animals' respiratory health through a platform dedicated to managing data related to defined alert thresholds. However, such systems, based solely on time-based sound recordings, cannot detect all types of diseases associated with pig farming.
[0004] Another known method of monitoring livestock by tracking noise emissions involves analyzing the sounds produced in livestock buildings coupled with artificial intelligence systems using deep neural networks, more commonly known as deep learning. Such a system is effective in detecting pathologies associated with the emission of unusual sounds and is specific to respiratory diseases causing coughing or sneezing.
[0005] Application WO2017 / 161167 describes a method for behavioral analysis of animals using cameras, particularly laboratory mice, to detect an immediate change in animal behavior at a given time. Such a method, while showing some effectiveness for detecting neurodegenerative or anxiety-related diseases in animals, is not well suited for monitoring diseases such as swine flu.
[0006] Influenza A viruses, which cause influenza in pigs, can cause enzootic forms of infection characterized by the persistence of these viruses within the herd, systematically infecting all groups of growing pigs at a fixed age. These "recurrent" influenza infections are common in piglets aged 5 to 8 weeks. Managing the associated clinical presentation is crucial. major for the breeder, but monitoring clinical signs can prove difficult.
[0007] Enzootic lung diseases, such as swine flu, are a major health concern in all pig-producing countries where animals are kept in large groups in buildings. These diseases are responsible for significant economic losses for the pig industry due to reduced zootechnical performance in affected animals, increased condemnation rates at slaughter, and higher production costs related to drug treatments and vaccinations. Beyond their impact on animal health and welfare, lung diseases represent a public health issue for veterinary medicine. Indeed, respiratory diseases are a major reason for antibiotic use in pigs (Chauvin et al., “A survey of group-level antibiotic prescriptions in pig production in France,” Prev. Vet. Med.).55, 109-112, 2002), which can have consequences in terms of the risk of developing antibiotic resistance and poses the problem of the release of antimicrobial residues into the environment (Fablet et al., “Infectious agents as-sociated with respiratory diseases in 125 farrow-to-finish pig herds: A cross-sectional study” Vet. Micro. 157, 152-163, 2012).
[0008] To overcome all or part of the aforementioned drawbacks of the prior art, the present invention relates to a pig farming monitoring system, comprising an electronic unit containing a control means and a data collection means, and a camera connected to said electronic unit. The monitoring system includes a data processing unit connected by a connector to the data collection means of said electronic unit. The system is characterized in that the data processing unit incorporates an image analysis algorithm for detecting the presence of at least one animal lying down in an animal housing area. Such a detection system makes it possible to monitor the number of inactive animals potentially affected by respiratory disease.
[0009] The invention is advantageously implemented according to the embodiments and variants set out below, which are to be considered individually or according to any technically operative combination.
[0010] Advantageously, the treatment unit locates the animal in a supine position and determines its possible location on the periphery of the rearing area. Sick animals most often prefer to lie down on the edge of rearing areas.
[0011] Advantageously, the processing unit calculates the ratio of the number of animals lying down to the total number of animals in the rearing area and compares this ratio to threshold values calculated beforehand. The algorithm is self-adaptive in that it allows the threshold values to be defined by calculation during a preliminary calibration phase. in a particular farming configuration. These threshold values correspond to a ratio of the number of pigs lying down to the total number of pigs present in the farm considered, and are representative of an anomaly linked to a pathology, in particular the presence of a pulmonary disease such as swine flu.
[0012] Advantageously, the processing unit includes a neural network. The present invention also relates to a method for monitoring sick animals, implementing the monitoring system as described above within the scope of the invention, comprising the following steps: a- selection of a sequence of images collected from the camera; b- integration of images into a storage unit preferably including a web platform; c- image analysis by a processing unit integrating an image analysis algorithm implementing a neural network; d- calculation of the barycenters of the squares of the drawings obtained by the algorithm; and e- determination of the number of animals lying down. Such a process makes it possible, thanks to the value obtained in step e-, to determine whether the posture and position in the animal rearing area corresponds to an anomaly.
[0013] Advantageously, the method according to the invention includes a step f- of locating the animals in an area of interest. Such an area of interest is preferably located on the edge of the rearing area; the tests carried out have shown that animals suffering from influenza come to lie down in the spaces located on the periphery of their living area.
[0014] Advantageously, the method according to the invention includes a step g- of calculating the ratio of the number of animals lying down to the total number of animals in the rearing area.
[0015] Advantageously, the method according to the invention includes a step h- of comparing the value obtained at e- or g- with a threshold value. The result obtained allows for the determination of whether a threshold of sick animals has been crossed and enables the generation of alerts.
[0016] Other advantages, purposes and features of the present invention will become apparent from the following description, given for explanatory purposes and in no way as a limitation, with reference to the accompanying drawings, in which:
[0017] [Fig. 1] represents ratio values of the number of pigs lying down to the total number of pigs present in an area for breeding A;
[0018] [Fig.2] represents ratio values of the number of pigs lying down to the total number of pigs present in an area for rearing B; and
[0019] [Fig.3] represents ratio values of the number of pigs lying down to the number of pig present in total in a zone for livestock C.
[0020] [Fig.4] represents a monitoring system according to the invention.
[0021] EXPERIMENTAL SECTION
[0022] Data collection and analysis protocols and tools
[0023] Three pig farms are used to collect data which serve as a reference in the database of a processing unit implemented in a pig farm monitoring system according to the invention.
[0024] Of the three pig farms mentioned above, two are affected by post-weaning (PS) influenza viruses and one farm is a "control" farm, i.e. without influenza viral circulation: Farming “A” (achieved): farrow-to-finish operation of 650 sows managed in 10 batches of 65 sows; piglets are weaned at 21 days of age and are housed in pens of 28 animals; Farming “B” (achieved): farrow-to-finish operation of 220 sows managed in seven batches of 32 sows; piglets are weaned at 28 days of age and are housed post-weaning at 33 days of age in pens of 25 animals; Farm “C” (control): farrow-to-finish operation of 600 sows managed in 20 groups of 30 sows; piglets are weaned at 21 days and are housed in pens of 25 animals. Farm A has a history of influenza A virus infections (H1N1 and H1N2), which recurrently manifest in post-weaning piglets between 6 and 8 weeks of age. Laboratory diagnosis, performed by PCR and multiplex RT-PCR on nasal swab supernatant according to the RESAVIP protocol (Hervé et al., "Virological and epidemiological patterns of swine influenza A virus infections in France: Cumulative data from the RESAVIP surveillance network, 2011-2018," Vet. Micro. 239, 2019), is carried out before enrollment. This diagnosis confirmed the presence of the virus during characteristic clinical episodes. Farm B presents a recurrent respiratory clinical picture in post-weaning between 5 and 8 weeks, of recent onset. Farm C, for which no particular respiratory problems were observed post-weaning during the three years preceding the experimental phase, underwent a pre-inclusion visit to ensure the absence of any significant clinical issues.
[0025] Description of data collection: In each of the three farms mentioned above, an electronic unit 1, marketed under the reference Copeeks® by the company Copeeks, comprising a camera 2 with precision optics, is installed in a post-weaning room for the duration of the piglets' presence, on a batch of piglets, and from the moment they enter. images are analyzed by an image analysis algorithm 3 which is programmed on the basis of prior supervised learning, enabling it to identify animals lying down or standing up in a predefined location of the area or breeding place (pen), referred to as the area of interest, which area of interest is, according to an advantageous embodiment, located on the periphery of the breeding place.
[0026] The electronic unit 1 is connected to a data storage unit 4 and transfers data according to the 4G standard. The storage unit 4 is of the "cloud" type and transfers the data in the form of images to a web platform 5.
[0027] The images collected from the electronic device are then analyzed by a data processing unit 6 configured using an image analysis algorithm 3.
[0028] All of these elements together form a monitoring system 7, as illustrated in [Fig.4],
[0029] Image analysis algorithm 3 was tested using images in which piglets were standing or lying down in the area of interest. [Table 1] presents the confusion matrix obtained from the predictions of image analysis algorithm 3, with regard to the detection of piglets in the area of interest.
[0030] This matrix allows the predictions of the image analysis algorithm 3 to be compared with reality. The algorithm correctly identified the posture of the piglets in 7867 cases, and errors were found in 830 cases, out of a total of 8697 cases.
[0031] [Tables 1] Correct prediction of piglet posture: Piglet lying down / standing in the area of interest; Piglet not present in the area of interest. Positive: 7132, 313; Negative: 517, 735
[0032] From these results, three indicators are calculated in order to measure the reliability of the algorithm: the sensitivity, the specificity and the accuracy of the image analysis algorithm 3.
[0033] For this algorithm, the sensitivity is 93.2%, the specificity is 58.7%, and the accuracy is 95.8%. Image analysis algorithm 3 therefore gives satisfactory results for detecting piglets lying down or standing up in the area of interest.
[0034] The photos from the camera are thus temporarily stored in the "cloud" before being analyzed in the data processing unit 6. Modulator-demodulator (modem) type connectivity equipment is connected via USB connectors to a central unit contained in the electronic box 1. The central unit communicates with the web platform 5 to receive orders from the user, in order to modify its behavior according to the context.
[0035] The platform allows a region of interest to be drawn on a photo, to match areas of the image with the edges of the box containing the animals. This region of interest is automatically applied to all new photos.
[0036] For each retrieved photo, a task is created and sent to software known as "workers," distributed across the cloud, whose function is to retrieve the new photo along with its area of interest. A neural network identifies the pigs in an image and discriminates them based on their position. The photo is thus analyzed by the neural network integrated into the TensorFlow learning tool provided by Google LLC™; this tool is configured with a MOBILENET SSD V3™ architecture. The centroid of the squares drawn by the algorithm is retrieved, the animal's posture is determined, and, for each animal, it is checked whether it is in one of the previously retrieved areas of interest. The worker sends the results of the analysis to a platform (copeek.io), including: the number of animals lying down in each area of interest, and the total number of animals.
[0037] The monitoring system 7, comprising all the elements linked and / or connected to each other between the monitoring unit 1 and the data processing unit 6, analyzes each photo and distinguishes them according to the pigs' position. The monitoring system 7 sends the analysis results (number of animals lying down in each area of interest and total number of animals) back to the platform. All of these actions are performed automatically by the software containing the algorithm.
[0038] For each farm (A to C), more than one hundred images are collected per day over a period of several months.
[0039] Division of the post-weaning phase into three periods: Based on the literature (Trombani et al., "Monitoring of recurrent influenza clinical signs post-weaning: the value of a real-time image analysis tool", Proceedings of the AFMVP Annual Congress, 2019.) and observations made on pig farms, the post-weaning phase of piglet presence is divided into three periods: -a period 1 lasting 10 days refers to the adaptation phase of animals in post-weaning (Orgeur et al., "Behavioral consequences of ultra-early weaning in the Large-White piglet" Journées de la Recherche Porcine en France 30, 383- 388. 1998.); -a period 2 which is determined by default between the end of period 1 and the beginning of period 3, described below; and -A third period begins when the farmer or the farm's monitoring veterinarian notices animals exhibiting behavior suggestive of the disease. This period ends on the day the pigs enter the fattening pen. For farm A, oral fluid samples were collected on the day of the call in the four monitored pens by placing two ropes in each pen (one rope for every 14 animals). Individual influenza PCR testing was performed on the eight oral fluid samples in pools of two (each pool representing one pen) and yielded a positive result. For farm B, close monitoring of the ratio of piglets lying down to the total number of piglets in the pen allowed the veterinarian to intervene at the appropriate time to collect the samples necessary to detect the respiratory disease. Regarding farm C, the start of period 3 is determined using the literature concerning the frequent age of detection of recurrent influenza clinical signs post-weaning, i.e. between 5 and 8 weeks of age (Cador et al., “Maternally-derived antibodies do not prevent transmission of swine influenza A virus between pigs.” Veterinary Research 47, 86. 2016 a. ET “Matemally Derived Immunity Extends Swine Influenza A Virus Persistence within Farrow-to-Finish Pig Farms: Insights from a Stochastic Event-Driven Metapopulation Model.” PLoS One 11 (9) 2016 b.).
[0040] Statistical analyses: The variable "ratio of piglets lying down to the total number of piglets in the pen" is created by dividing the number of piglets lying down in the predefined area by the total number of animals present in the pen. This number is automatically determined for each image by the analysis software. This value allows for more precise tracking of changes in the number of piglets lying down in the predefined area of interest. For statistical analyses, an average of the "ratio of the number of piglets lying down to the total number of piglets in the pen" is calculated per pen and per period. The pen constitutes the experimental unit. Statistical analyses were performed using R™ software (version 3.6.0). For farms A and C, an analysis of variance (ANOVA) was performed on repeated measures within the farm, using time (the periods) as the fixed effect and the pen as the random effect. For farm B, the data were also analyzed using ANOVA with time (the periods) as the fixed effect. A Tukey test (Cornillon et al., "Statistiques avec R", 2008) was then used to compare the means pairwise between periods. The results are considered significant at P < 0.05. 0.05.
[0041] Results and discussion
[0042] Breeding A:
[0043] For breeding A, the animals exhibited behavior that could suggest the expression of respiratory pathology at 7.5 weeks of age, which is consistent with the literature (Cador et al., 2016 a and 2016 b).
[0044] [Table 2] describes the effect of the period on the ratio "piglets lying in the predefined area to the total number of piglets in the pen" for rearing A(1).
[0045] [Tables2] Period Number of images per group Period duration (days) Raw averages of ratio2 Period 1 2064 10 0.06 a Period 2 2470 22 0.13 b Period 3 986 18 0.17 c ETR3 0.012 Statistical analysis P < 0.001 Tukey P < 0.05
[0046] (1) Crude averages; (2) Ratio of the number of piglets lying down to the total number of piglets in the box ;(3) ETR: Residual standard deviation; values not followed by the same letter are different for the Tukey test.
[0047] Statistical analysis shows a significant difference in the mean ratio of the number of piglets lying down to the total number of animals in the pens between period 2, in the presumed absence of clinical signs, and period 3, which is characterized by the presence of influenza-like symptoms, as deduced from Table 2 and [Fig. 1]. [Fig. 1] is a box plot showing, for the three periods - 1: period 1; 2: period 2; 3: period 3 - plotted on the x-axis, the ratio deduced on the y-axis of the number of piglets lying down to the total number of piglets in the pen as a function of the different periods for farm A.
[0048] Breeding B:
[0049] Regarding farm B, the farmer's late detection of symptoms suggestive of respiratory disease had not yet allowed the farm's monitoring veterinarian to establish a definitive diagnosis of recurrent post-weaning influenza by taking samples. However, thanks to the use of image analysis tools in this study, the veterinarian, through daily monitoring of the number of piglets lying down in the area of interest, was able to intervene at the appropriate time to confirm the presence of influenza in this farm.
[0050] For this breeding, the animals exhibited behavior that could suggest the expression of respiratory pathology at 5.5 weeks of age.
[0051] [Table 3] describes the effect of the period on the ratio “piglets lying down in the area predefined on total number of piglets in the pen" for breeding B1.
[0052] [Tables3] Period Number of images per group Duration of the period (days) Raw means1 of the ratio2 Period 1 336 10 0.13 a Period 2 475 7 0.32 b Period 3 3315 35 0.40 c ETR3 0.059 Statistical analysis P <0.001 Tukey P <0.05
[0053] (1) Crude averages; (2) Ratio of the number of piglets lying down to the total number of piglets in the box ;(3) ETR: Residual standard deviation; values not followed by the same letter are different for the Tukey test.
[0054] The statistical model also highlighted an effect of the period on the "ratio of the number of piglets lying down to the total number of piglets in the pen." A significant increase in this ratio was observed between period 1, period 2, and period 3 (Table 3, [Fig. 2]). [Fig. 2] is a box plot showing, for the three periods—1: period 1; 2: period 2; 3: period 3—ported on the x-axis, the ratio deduced on the y-axis of the number of piglets lying down to the total number of piglets in the pen as a function of the different periods for farm B.
[0055] The monitoring system according to the invention made it possible to detect the increase in the number of pigs lying down over the days during the three periods with a particular increase in this number of piglets lying down at the time when the veterinarian observed the respiratory clinic.
[0056] In farms A and B, affected by the Influenza virus, the animals observed are therefore less active during the activity phase of period 3 and tend to lie down more at the ends of the pen, a possible consequence of the influenza clinical signs.
[0057] Breeding C:
[0058] For farm C, not affected by post-weaning respiratory clinic, the statistical model did not show an increase in the "ratio of the number of piglets lying down to the total number of piglets" between periods 1 and 3 and between periods 1 and 2, contrary to what was observed at farms A and B, affected by the Influenza virus.
[0059] [Table 4] describes the effect of the period on the ratio "piglets lying in the predefined area to the total number of piglets in the pen" for farm C1.
[0060] [Tables4] Number of images per group Period duration (days) Crude averages¹ of the ratio² Period 1 532 10 0.28 ab Period 2 698 18 0.3 a Period 3 1225 24 0.26 b ETR³ 0.032 Statistical analysis P < 0.01 Tukey P < 0.05
[0061] (1) Crude averages; (2) Ratio of the number of piglets lying down to the total number of piglets in the box ;(3) ETR: Residual standard deviation; values not followed by the same letter are different for the Tukey test.
[0062] Thus, from the data collected and reported in Table 4 and [Fig.3], in the case of farm C, i.e. farm without identified pathology, no significant increase in the number of pigs lying down or decrease in the activity of the animals could be shown.
[0063] Alert thresholds are set according to the ratio of the number of piglets lying down to the total number of piglets in a pen. When the monitoring system detects that one of these thresholds has been exceeded, it alerts the farmers at a specified time via a dedicated computer interface (laptop, tablet, smartphone, etc.).
Claims
Claims
1. System (7) for monitoring behaviors associated with pathologies, for pig farming, comprising an electronic box (1) containing a control means and a data collection means and a camera (2) connected to said electronic box, said monitoring system comprising a data processing unit (6) connected by a connector to the data collection means of said electronic box, characterized in that the data processing unit (6) integrates an image analysis algorithm (3) for detecting the presence of at least one animal in a lying position in a breeding area of the animal, and in which the processing unit locates the animal in a lying position and determines its possible location on the periphery of the breeding area.
2. Monitoring system (7) for pig farming according to claim 1, in which the processing unit calculates the ratio of the number of animals lying down to the total number of animals in the breeding area and compares said ratio to threshold values calculated upstream.
3. Method for monitoring sick animals, implementing the monitoring system according to any one of claims 1 to 2, comprising the following steps: a- selection of a sequence of images collected from the camera (2); b- integration of the images into a storage unit; c- analysis of the images by a processing unit (6) integrating an image analysis algorithm implementing a neural network; d- calculation of the barycenters of the squares of the drawings obtained by the algorithm; and e- determination of the number of animals lying down.
4. Monitoring method according to claim 3, comprising a step f- of locating the animals in an area of interest.
5. Monitoring method according to claim 4, comprising a step g- of calculating the ratio of the number of animals lying down to the total number of animals in the breeding area.
6. Monitoring method according to claim 5, comprising a step h- of comparing the value obtained at e- or g- with a threshold value.
