1. Introduction
Recently, there has been increased interest in detecting abnormal behavior among domestic animals. If abnormal activity is not detected accurately and in a timely manner, efficient reproductive performance may be limited. Therefore, some recent studies have applied information technology to a livestock management system to minimize the damage resulting from such anomalies [
1,
2,
3,
4,
5,
6].
In this study, we aimed to detect and classify aggressive behaviors among weaning pigs in an intensive commercial pigsty. When unfamiliar pigs first meet after mixing group-housed weaning pigs, social conflict involving excessive aggression can occur. Aggression may be part of the behavioral repertoire that pigs exhibit to solve social conflicts [
7]. When introduced to unfamiliar conspecifics, pigs naturally interact aggressively for social hierarchy status access to resources, such as space and feed [
8,
9,
10,
11]. These interactions may hamper animal welfare and increase wounding, leading to infections which may be lethal in extreme cases [
12]. In addition, such aggression results in economic losses as weaker animals dominated by more aggressive ones may not have sufficient access to food so that their grow rates decrease and weight variability within the pen increases [
13]. Therefore, aggression among pigs is one of the most important health, welfare, and economic problems in intensive farming [
14,
15].
Recently, two interesting analyses of pig aggression have been reported. First, Viazzi
et al. [
16] developed a method of detecting pigs’ aggressive behavior continuously and automatically through image processing, which enables obtaining information on the motions of pigs from historical images to find out aggressive interactions. Two features, the mean intensity of motion and the space occupation index, are derived from the segmented region of the motion history images and are used to classify aggressive interactions during episodes through linear discriminant analysis. This method was the first attempt to use image analysis to automatically detect aggressive behaviors among pigs. Second, Oczak
et al. [
17] tested a method for automatically detecting aggressive behavior in pigs using an activity index and a multilayer feed-forward neural network. In that method, the activities of the animals are measured using videos, and software is used to calculate an activity index. Five features (average, maximum, minimum, summation, and variance of the activity index) are calculated based on the video recordings, and a multilayer feed-forward neural network is trained and validated to classify events involving high and medium aggression. Their results suggest that combining the activity index with a multilayer feed-forward neural network can be used to classify aggressive pig behavior. Recently, some advances have been made in pig monitoring using red, green, and blue (RGB)-based video data; however, to the best of our knowledge, no automated analysis of anomalies using a Kinect depth sensor has been reported yet.
In contrast to current methods, in this study, we developed a non-invasive, inexpensive, and automatic monitoring prototype system that uses a Kinect depth sensor to monitor animal activity in a commercial pig facility. This proposed system notifies the farmer of an aggressive situation when it occurs in a hog barn. The method begins by extracting activity features from the Kinect depth information obtained in a pigsty. The detection and classification module, which employs two binary-classifier support vector machines (SVMs) in a hierarchical manner, detects aggressive behavior, and classifies it into aggressive sub-types such as head-to-head (or body) knocking and chasing. The results of our experiments indicate that the accuracy of aggression detection approached 95.7%, and the aggression classification approach (90.2% accuracy) was validated, where the recall and precision measures were satisfactory. As far as we know, this is the first report of aggression detection in weaning pigs by a pig monitoring system using Kinect depth data and SVMs. The results of this study suggest that Kinect depth sensors can be used to monitor the behavior of pigs. Furthermore, given the continuous and large stream of data coming from a pig monitoring system, the application of our data mining method is appropriate.
The remainder of this paper is composed as follows.
Section 2 describes the proposed aggressive behavior recognition system for pigs using a Kinect depth sensor, and it provides some information on the background concepts.
Section 3 presents the simulation results and
Section 4 presents the conclusions.
4. Conclusions
In the management of group-housed livestock, detecting anomalies early is very important. In particular, failure to detect aggression in a timely and accurate manner in intensive commercial pigsties could seriously limit efficient reproductive performance. In this study, we developed a low-cost, non-invasive, and automatic prototype system to monitor animal activity in a commercial pig farm, which notifies the farmer of aggression situations in the pigsty. The proposed system preprocesses an activity-feature subset by analyzing the pig activities acquired using a Kinect depth sensor. The recognition module detects aggressive behaviors and classifies them hierarchically based on aggression sub-types, such as head-to-head (or body) knocking and chasing. In our experiments, we found that the accuracy of the aggressive behavior detection obtained using the proposed system was 95.7%, and the aggressive behavior classification measures were satisfactory.