Classification of aggressive behaviour in pigs by activity index and multilayer feed forward neural network

Abstract

The aim of this study was to test a method to automatically detect aggressive behaviour in pigs, by using an activity index and a multilayer feed forward neural network. An experiment was carried out on a mixed group of 11 male pigs weighing on average 23 kg. During the first 3 days after mixing, the pigs were recorded for video analysis. Out of the total video recording time (8 h or 28,800 s), 643 s were labelled as high aggression events and 1253 s as medium aggression events. Activity of the animals was measured on the videos using software that calculated activity index. Five features of activity index were calculated on the recorded videos (average, maximum, minimum, sum and variance) over 14 time intervals. A multilayer feed forward neural network was trained and validated to classify events of high aggression and medium aggression. Seven types of artificial neural network (ANN) architectures were tested in our study. The results revealed that ANNs, fed with 70 features of activity index (5 features × 14 intervals) calculated on 241 s time intervals, classified high aggression events with a sensitivity of 96.1%, specificity of 94.2% and an accuracy of 99.8% whereas medium aggression events were classified with a sensitivity of 86.8%, specificity of 94.5% and an accuracy of 99.2%. These results indicate that a combination of the activity index and multilayer feed forward neural network can be used to classify aggressive pig behaviour.