AP—Animal Production Technology: Predicting Sensor Placement Targets on Pigs using Image Analysis

• Published in: Biosystems engineering 2002-04, Vol.81 (4), p.453-463
• Main Authors: Tillett, R.D., Frost, A.R., Welch, S.K.
• Format: Article
• Language: English
• ISSN: 1537-5110; 1537-5129
• DOI: 10.1006/bioe.2001.0018

This work forms part of a longer-term project aimed at developing a robot to bring a sensor into contact with an animal. The example application is to place an ultrasonic sensor onto the back of a pig as it uses a feeding stall. This paper presents the accuracy required in placing the sensor, and the accuracy of target location achieved using a linear function of landmark points on the pig outline, located by image analysis. The P2 standard position is close to a local minimum in fat thickness. When the target is chosen to be 25 mm ahead of the last rib and 50 mm from the mid-line of the pig, then a positional error of ±10 mm laterally and ±25 mm longitudinally will give a backfat measurement within 5% of the minimum. A linear model predicting the target position from six points on the outline was trained on 7549 images and tested on 2978 images. On the test data the root means square (r.m.s.) errors in the X and Y directions were 16 and 38 mm, respectively. A linear model with separate offset terms in X and Y coordinates for each sequence gave r.m.s. errors of 6 and 8 mm. However, this requires the offset value to be calculated for each sequence. An active system is proposed where multiple measurements from each pig will be used to improve the backfat estimate to within 5% of the minimum value.