Академический Документы
Профессиональный Документы
Культура Документы
are used in each system. For automatic cattle sensing. The conception of a proxy-detection system is
monitoring/assessment system, for example, it is necessary to motivated by the need of better resolution, accuracy,
determine the shape of the cow’s volume with a mm-level or temporality and lower cost, compared to remote sensing. The
cm-level accuracy, in real-time and in very challenging use of CV techniques allows obtaining this information
environmental and lighting conditions. ToF cameras best fit automatically with objective measurement in contrast with the
such applications as they provide their own light source, they difficulty and subjectivity of visual or manual acquisition.
provide reliable measurements in real-time with the required
precision (as proven by (Salau et al., 2017)). One can classify the wide range of applications where 3D
sensors are used by considering their scenario of application,
On the other hand, for mobile robot navigation, it may only be yielding scene-related tasks, object-related tasks and
necessary to guide an autonomous vehicle across the terrain applications involving interaction with animals. Scene-related
and avoid obstacles which requires decimetre-level accuracy tasks generally involve mobile robots and large displacements,
and low power consumption in less stringent lighting while object-related tasks involve instead robotic arms or
conditions. For these applications, stereovision may be more humanoid-like robots, and small depths. Finally, applications
applicable, being a passive depth-sensing technique it requires involving interaction with animals involve dynamic response
very little power, it provides adequate accuracy for navigation capabilities and safety considerations. In the context of this
(Vázquez-Arellano et al., 2016) and can be incorporated with review; 3D vision for perception in mobile robotics falls under
2D high-definition cameras for the identification of a wide the scene-related category, Automatic cattle monitoring &
variety of different objects. feeding falls under the object-related category and systems for
automatic milking & herding can be categorised as
2.2 3D Vision Data Analysis applications involving interaction with animals.
For the sake of brevity, this section will just outline that 3D 3.1 Mobile Robot perception
data analysis techniques emanate from one of two fields of
research; To enable mobile robots to move autonomously in a variety of
unstructured environments, 3D information about the
1. Traditional 3D processing techniques which consists of environment is essential. Several studies of accuracy and
manually defining 3D descriptors which may be local (e.g. usability of 3D sensors for mobile robots have been published
Point Feature Histogram (PFH), Signature of Histogram as summarised by (Rauscher et al., 2014). Their paper also
of Orientations (SHOT) and Rotational Projection presents an experimentally derived model of the expected
Statistics (RoPS)) or global (e.g. Ensemble of Shape standard deviation which was dependent on the distance to the
Functions (ESF) and Viewpoint Feature Histogram measured object as well as measured intensity and
(VFH)) (Alexandre, 2012) (Ioannidou et al., 2017). environmental illumination for several 3D sensors which is
2. 3D Deep Learning, also known as Geometric Deep important to know for probabilistic approaches in robotics.
Learning, an emerging field which applies deep learning The accuracy of a perception system is important in mapping,
to 3D data. Most of this work involves 3D Convolutional localization and obstacle classification problems. Whilst
Neural Networks (3D-CNNs) and techniques to make obstacle avoidance problems place higher priority on the
them more efficient and more invariant to rigid translation availability and reliability of the measured data. We discuss
and rotation. localisation and object detection techniques and dealing with
A discussion of the pros and cons of these two strands of sources of uncertainty in 3D vision including limited
research may be found in our previous paper, (O’ Mahony et availability due to limited field-of-view and limited accuracy
al., 2019c), and a more in-depth review of 3D deep learning arising from variations in distance & lighting in a previous
can be found in (O’ Mahony et al., 2018b). To summarise, paper (O’ Mahony et al., 2019a).
greater accuracy is achievable with deep learning, given For the most part, much of the work done in mobile robot
training data of sufficient volume and variety. navigation has been involved in crop management applications
Although, 3D deep learning has an inherent edge over its 2D where 3D CV has been used for crop phenotyping (Vázquez-
counterpart, it has not benefited from the recent developments Arellano et al., 2016). These techniques may be extended to
in deep learning due to the unavailability of large training applications in PLF where a robot may have to traverse terrain
datasets as well as large test datasets. Therein lies a gap in the to herd/monitor cattle or carry materials around the farm, e.g.
research, not only for 3D datasets for agricultural applications (Grimstad et al., 2017), (Salah Sukkarieh et al., 2019). Such
but for 3D CV applications in general. an application has been implemented by Lely’s Shrimp robotic
platforms with the use of line scanners for 3D perception
3. AGRICULTURAL APPLICATIONS (Wendel et al., 2017). With respect to safe navigation using
3D vision, (Kragh et al., 2017) have developed a multimodal
In order to optimize herd management and cater for the needs
dataset (including 3D sensors) intended for training/testing of
of every individual cow, precision agriculture uses new
intelligent navigation systems in farm environments. To date,
technologies such as computing, electronics and imaging. Two
mobile robotics has either focused on indoor applications or
types of imagery can thus be used: proxy-detection and remote
outdoor applications, however transitioning from outdoor
313
IFAC AGRICONTROL 2019
December 4-6, 2019. Sydney, Australia
314
IFAC AGRICONTROL 2019
December 4-6, 2019. Sydney, Australia
driven CV techniques which have been trained with sufficient The 3D vision-based animal monitoring implementations
variability across deployment scenarios have fewer rigid demonstrate capability to recognise cow which allows for real-
constraints compared to hard-coded techniques. time insights about animal health to be known and be
incorporated with feeding systems to manage cow nutrition
There have been a lot of studies relating lameness to more effectively and alerting the farmer to ill health promptly.
parameters which can be quantified with CV. Several research
projects have taken advantage of 3D vision to extract the For both Locomotion Scoring and BCS, many of the
curvature of the cow’s spine or to measure between prominent techniques reviewed have used pre-processed depth images as
features such as hook bones and hip bones from the depth input where the 3D spatial relationship between all points
image. These methods follow the similar processing pipelines, measured by the 3D camera is not captured. We propose that
which may include; background removal, height thresholding, techniques which operate on the raw pointcloud data, such as
smoothing to prevent limiting any curvature information, (Qi et al., 2017), would yield better accuracy and reliability as
curvedness data calculation on the high peaks and binary it allows geometric pre-processing and noise reduction
image representation of the curvedness threshold to track the techniques to be applied and the shapes present in the
most distinctly curved/highest convex regions (i.e. spine, pointcloud would be invariant of camera pose, light amplitude
hooks and pins). In an effort to identify temporal signatures of and background light.
lameness, (Gardenier et al., 2018b) propose the use of object
detection CNNs to accurately detect the cow’s hoofs and Table 1: Past 3D Vision-based Body Condition Scoring
thereby track hoof placement which has been demonstrated to Research
be an effective indicator of lameness (Bahr et al., 2008). Such
an approach based on deep learning has proven, at least in Reference Size of Camera
Accuracy
other fields of research, to be more robust compared to the test used
Features (% within
dataset
hand-crafted classical CV approaches which contain 0.25)
(images)
parameters fine-tuned by CV engineers which do not tend to
generalise well. Height (Spoliansky 2650 Kinect 74
measurements et al., 2016) v1
Locomotion scoring is very complex because a cow might not around cow’s
exhibit these traits consistently or at all in certain scenarios back + age +
(e.g. if a camera is deployed in an area where cattle are eager weight
to get to feed or are being herded. This makes deployment very Sphere fitted (Hansen et 95 Asus 66
difficult as a setup which may be effective in one location may to surface of al., 2018b) Xtion
not work in another location, even on the same farm. Some cow’s back Pro
works have tracked traits over time, establishing a sliding
window over which each cow’s characteristic traits are Features (Rodríguez 503 Kinect 78
determined by Alvarez et v2
determined which then allows deviations from the norm to be
CNN al., 2018)
detected. Another approach is to use an assessment
methodology which is flexible enough to be employed in a De Laval’s (Mullins et 344 De 76
wide range of scenarios where the increased frequency of proprietary al., 2019) Laval
observations increases the chances of detection. This is where system camera
geometric deep learning may play a vital role with techniques
which learn from shape alone and are robust to noise and rigid
translation and rotation of the point cloud. 3.3 Applications involving Interaction with Animals
3.2.3 Body Condition Scoring 3.3.1 Robotic Milking
Body Condition Score (BCS) is an indirect estimation of the There are several offerings on the market for Automatic
level of body reserves, and its variation reflects cumulative Milking Systems (AMS). Recent activity in the area has
variation in energy balance. It interacts with reproductive and included the integration of more modern sensors/vision
health performance, which are important to consider in dairy systems, the addition of animal monitoring functionalities such
production but not easy to monitor. Manual visual BCS is as those discussed in Section 3.2, and the integration of robots
subjective, time-consuming and requires experienced into rotary milking parlours.
employees. The state-of-the-art in automatic BCS estimation
Time of Flight (ToF) depth sensing cameras have been used in
is summarised in Table 2 under categories such as the size of
many recent developments in AMSs. Acceptable accuracy
the test dataset, the degree of automatization, the type of
(teat location detection within 5mm) has been achieved when
camera used and the accuracy of scores within 0.25 of manual
the search space is limited to region of interest 150mm wide
BCS reference.
by 80mm deep (Duffy et al., 2006). Teat detection and tracking
using algorithmic solutions from depth images and point-
cloud data was also achieved by (Van Der Zwan et al., 2015).
315
IFAC AGRICONTROL 2019
December 4-6, 2019. Sydney, Australia
The method can robustly handle occlusions, variable poses, Extensive testing is necessary to ascertain the automatability
and geometries of the tracked shape, and yields a correct of current research. No two farms are the same and to deploy
tracking rate for over 90% for tests involving real-world vision-guided robotic systems to the end user will require
images obtained from an industrial robot. Other vision systems which can operate in unconstrained environments and
technologies have also been investigated, for example in adapt to variations in herd characteristics, weather conditions,
(Rastogi et al., 2017) using a Kinect structured light camera farmyard layout and different scenarios in animal-robot
and a haar-cascade classifier, in (Ben Azouz et al., 2015) using interaction.
a combination of thermal imaging and stereovision techniques.
This project intends to address this concern through the use 3D
The task of attaching the milking clusters to cows’ teats is a deep learning techniques to derive features which are robust to
challenging one as the shape of an udder is variable between rigid transformation (e.g. rotation and translation) and
cows and between different stages of lactation. The udder is therefore allows more flexible installation options for the
also a moving target if cows are restless in which case there is animal monitoring vision system and real-time situation
a risk for the robot to harm the cow and vice versa for the cow awareness in robotics applications. Secondly the learning of
to throw the robot out of calibration. Cycle time of the milking characteristics specific to individual cattle (e.g. the appearance
operation must also be minimal so that both cow and farmer of each cow for cow identification) has also been identified as
can be more productive with their time. Therefore, an a research gap which this project intends to fill.
intelligent control of the robot is required using visual
feedback to navigate the cluster onto the cow accurately, safely
and quickly. A system which can milk cows with unusually
shaped udders or that do not take to a robot milker which ACKNOWLEDGEMENTS
would otherwise have to be culled as is the case with some
This work was supported, in part, by Science Foundation
existing AMS, would also be a considerable advantage.
Ireland grant 13/RC/2094 and co-funded under the European
This functionality could be achieved through the Regional Development Fund through the Southern & Eastern
implementation of AI that can generalise well to a wider Regional Operational Programme to Lero - the Irish Software
variety of udder shapes in challenging sensing conditions and Research Centre ( www.lero.ie )
dynamically assess the status and viability of teat cup
placement during the procedure so as to avoid stressing the REFERENCES
cow. Such a perception system must be capable of coping with Abdul Jabbar, K. et al. (2017) ‘Early and non-intrusive
obstructions, dynamic targets and avoiding collisions and lameness detection in dairy cows using 3-dimensional video’,
therefore requires a number of components including teat, tail Biosystems Engineering, 153. doi:
and leg identification, target tracking and situation awareness. 10.1016/j.biosystemseng.2016.09.017.
3D point cloud processing techniques are an ideal candidate as
they are less subject to pose dependent ambiguity and allow Alexandre, L. A. (2012) 3D Descriptors for Object and
for geometric reasoning of segmented 3D shapes. Category Recognition: a Comparative Evaluation.
Arslan, A. C., Akar, M. and Alagoz, F. (2014) ‘3D cow
4. CONCLUSIONS identification in cattle farms’, in 2014 22nd Signal Processing
This paper has identified the applicability of 3D vision in farm and Communications Applications Conference (SIU). IEEE,
automation tasks. The diversity of applications will require the pp. 1347–1350. doi: 10.1109/SIU.2014.6830487.
implementation of CV algorithms and deep learning models Ben Azouz, A. et al. (2015) ‘Development of a teat sensing
which deliver accuracy, generalisation, low computational system for robotic milking by combining thermal imaging and
cost and customisability. For example, cow health assessment stereovision technique’, Computers and Electronics in
systems must make accurate measurements of subtle features Agriculture. Elsevier B.V., 110, pp. 162–170. doi:
such as posture and body fat thickness while also being 10.1016/j.compag.2014.11.004.
customisable to individual cow and herd characteristics.
Bahr, C. et al. (2008) ‘Automatic detection of lameness in
The use of 3D vision in the implementation of automated dairy cattle by vision analysis of cow’s gait’, in International
systems for Precision Dairy Farming has been prevalent in Conference on Agricultural Engineering,. Hersonissos, Crete,
recent years. This has enabled real-time accurate 3D Greece: European Society of Agricultural Engineers (AgEng).
measurements which are so important to the applications in
Duffy, A. H. et al. (2006) ‘Teat Detection for an Automated
question. A further trend has been the development of
Milking System’.
algorithms which make use of AI for improved accuracy and
reliability in a diverse range of conditions considering that Gardenier, J., Underwood, J. and Clark, C. (2018a) ‘Object
physical traits can vary greatly depending on stage of lactation Detection for Cattle Gait Tracking’, 2018 IEEE International
and cattle breed and also that poor lighting and lens fouling are Conference on Robotics and Automation (ICRA), pp. 2206–
likely to occur in farmyard environments. 2213. doi: 10.1109/ICRA.2018.8460523.
Gardenier, J., Underwood, J. and Clark, C. (2018b) ‘Object
316
IFAC AGRICONTROL 2019
December 4-6, 2019. Sydney, Australia
Detection for Cattle Gait Tracking’, 2018 IEEE International Systems for Precision Livestock Farming (unpublished)’, in
Conference on Robotics and Automation (ICRA), (May), pp. 9th European Conference on Precision Livestock Farming
2206–2213. doi: 10.1109/ICRA.2018.8460523. (ECPLF).
Grimstad, L. and From, P. J. (2017) ‘Thorvald II - a Modular O’ Mahony, N. et al. (2019c) ‘Deep Learning vs. Traditional
and Re-configurable Agricultural Robot’, IFAC- Computer Vision’, in Advances in Computer Vision. Springer
PapersOnLine. Elsevier, 50(1), pp. 4588–4593. doi: Nature Switzerland AG, pp. 128–144.
10.1016/J.IFACOL.2017.08.1005.
Qi, C. R. et al. (2017) ‘PointNet++: Deep Hierarchical Feature
Halachmi, I. and Guarino, M. (2016) ‘Editorial: Precision Learning on Point Sets in a Metric Space’, arXiv preprint
livestock farming: a “per animal” approach using advanced arXiv:1706.02413v1.
monitoring technologies’, animal, 10(09), pp. 1482–1483. doi:
Rastogi, A. et al. (2017) ‘Teat detection mechanism using
10.1017/S1751731116001142.
machine learning based vision for smart Automatic Milking
Hansen, M. F. et al. (2018a) ‘Automated monitoring of dairy Systems’, in 2017 14th International Conference on
cow body condition, mobility and weight using a single 3D Ubiquitous Robots and Ambient Intelligence (URAI). IEEE,
video capture device’, Computers in Industry, 98, pp. 14–22. pp. 947–949. doi: 10.1109/URAI.2017.7992872.
doi: 10.1016/j.compind.2018.02.011.
Rauscher, G., Dube, D. and Zell, A. (2014) ‘A Comparison of
Hansen, M. F. et al. (2018b) ‘Automated monitoring of dairy 3D Sensors for Wheeled Mobile Robots’.
cow body condition, mobility and weight using a single 3D
Rind Thomasen, J. et al. (2018) ‘Individual cow identification
video capture device’, Computers in Industry. Elsevier, 98, pp.
in a commercial herd using 3D camera technology’, in World
14–22. doi: 10.1016/J.COMPIND.2018.02.011.
Congress on Genetics Applied to Livestock Production, p. 613.
van Hertem, T. et al. (2018) ‘Implementation of an automatic
Rodríguez Alvarez, J. et al. (2018) ‘Body condition estimation
3D vision monitor for dairy cow locomotion in a commercial
on cows from depth images using Convolutional Neural
farm’, Biosystems Engineering. Academic Press, 173, pp.
Networks’, Computers and Electronics in Agriculture, 155,
166–175. doi: 10.1016/J.BIOSYSTEMSENG.2017.08.011.
pp. 12–22. doi: 10.1016/j.compag.2018.09.039.
Van Hertem, T. et al. (2014) ‘Automatic lameness detection
Salah Sukkarieh and Matthew Truman (2019) Agriculture and
based on consecutive 3D-video recordings’, Biosystems
the environment - Faculty of Engineering. Available at:
Engineering, 119, pp. 108–116. doi:
https://sydney.edu.au/engineering/our-research/robotics-and-
10.1016/j.biosystemseng.2014.01.009.
intelligent-systems/australian-centre-for-field-
Ioannidou, A. et al. (2017) ‘Deep learning advances in robotics/agriculture-and-the-environment.html (Accessed: 15
computer vision with 3D data: A survey’, ACM Computing July 2019).
Surveys, 50(2). doi: 10.1145/3042064.
Salau, J. et al. (2017) ‘A multi-Kinect cow scanning system:
Kragh, M. et al. (2017) ‘FieldSAFE: Dataset for Obstacle Calculating linear traits from manually marked recordings of
Detection in Agriculture’, Sensors. Multidisciplinary Digital Holstein-Friesian dairy cows’, Biosystems Engineering, 157.
Publishing Institute, 17(11), p. 2579. doi: 10.3390/s17112579. doi: 10.1016/j.biosystemseng.2017.03.001.
Mullins, I. L. et al. (2019) ‘Validation of a Commercial Spoliansky, R. et al. (2016) ‘Development of automatic body
Automated Body Condition Scoring System on a Commercial condition scoring using a low-cost 3-dimensional Kinect
Dairy Farm’, Animals, 9(6), p. 287. doi: 10.3390/ani9060287. camera’, Journal of Dairy Science. Elsevier, 99(9), pp. 7714–
7725. doi: 10.3168/JDS.2015-10607.
Van Nuffel, A. et al. (2015) ‘Lameness detection in dairy
cows: Part 1. How to distinguish between non-lame and lame Vázquez-Arellano, M. et al. (2016) ‘3-D Imaging Systems for
cows based on differences in locomotion or behavior’, Agricultural Applications-A Review.’, Sensors (Basel,
Animals. MDPI AG, pp. 838–860. doi: 10.3390/ani5030387. Switzerland). Multidisciplinary Digital Publishing Institute
(MDPI), 16(5). doi: 10.3390/s16050618.
O’ Mahony, N. et al. (2018a) ‘Computer Vision for 3D
Perception A review’, in Intelligent Systems Conference Viazzi, S. et al. (2014) ‘Comparison of a three-dimensional
(IntelliSys) 2018. Springer, pp. 1–9. and two-dimensional camera system for automated
measurement of back posture in dairy cows’, Computers and
O’ Mahony, N. et al. (2018b) ‘Convolutional Neural Networks
Electronics in Agriculture. Elsevier, 100, pp. 139–147.
for 3D Vision System Data A review’, in 12th International
Conference On Sensing Technology 2018. Wendel, A. and Underwood, J. (2017) ‘Extrinsic Parameter
Calibration for Line Scanning Cameras on Ground Vehicles
O’ Mahony, N. et al. (2019a) ‘Adaptive Multimodal
with Navigation Systems Using a Calibration Pattern’. doi:
Localisation Techniques for Mobile Robots in Unstructured
10.3390/s17112491.
Environments A Review’, in IEEE 5th World Forum on
Internet of Things (WF-IoT). Van Der Zwan, M. and Telea, A. (2015) ‘Robust and Fast Teat
Detection and Tracking in Low-Resolution Videos for
O’ Mahony, N. et al. (2019b) ‘Automatic Cattle Monitoring
Automatic Milking Devices’.
317