Using millimetre-wave radar for monitoring sow postural activity in individual pen: first results
Résumé
A millimetre-wave radar is tested in an INRAE experimental unit to monitor sow postural activity in presence of her
piglets. A total of 16 sows of the Large White breed with piglets aging from 7 to 15 days are monitored inside farrowing
pens with measurement sequences lasting 2 to 3 hours at different dates. The radar is attached to the farrowing pen
entrance at a distance to ground of 1.8m, and the sow’s position is remotely estimated from the backscattering of
electromagnetic waves. The radar-based detection technique does not require equipping animals with radiofrequency
tags. The automatized system records 3D images are built from the simultaneous azimuth (digital) and elevation
(mechanical) radar beam scannings with a time resolution of 3 seconds. By applying an algorithm based on Constant
False Alarm Rate, undesirable radar echoes from the pen are mitigated and only radar detections of the sow are recorded
over time. A clustering algorithm is applied to the detections to obtain the 3D position of the sow. A classification of
sow postures is performed from a Quadratic Discriminant Analysis of the 3D positions. Ground-truth postures of sow
are annotated manually from video recordings. The two following classes of postures are finally analysed: the ‘standing
and transition’ and the ‘lying’ classes (‘transition’ refers here to both ‘sitting’ and ‘kneeling’ postures). The training
data is composed of radar detections of 4 sows for the total monitoring duration of around 8 hours, or equivalently
for 7,097 detections. The tested data is composed of 34,356 radar detections of 12 other sows for a total monitoring
duration of 38 hours. The precision and sensitivity are 88.3 and 90% for the ‘standing and transition’ class, and 97.9
and 98% for the ‘lying’ class. Precision and sensitivity of the classification may vary from one sow to another due
to the difficulty to classify correctly ‘transition’ postures for some of them. Based on these first encouraging results,
future work will be devoted to further develop the radar detection method, and detect the motion of the sow according
to changes in her location in the pen.