Automated recognition of postures and drinking behaviour for the detection of compromised health in pigs

Abstract

Changes in pig behaviours are a useful aid in detecting early signs of compromised health and welfare. In commercial settings, automatic detection of pig behaviours through visual imaging remains a challenge due to farm demanding conditions, e.g., occlusion of one pig from another. Here, two deep learning-based detector methods were developed to identify pig postures and drinking behaviours of group-housed pigs. We first tested the system ability to detect changes in these measures at group-level during routine management. We then demonstrated the ability of our automated methods to identify behaviours of individual animals with a mean average precision of , under a variety of settings. When the pig feeding regime was disrupted, we automatically detected the expected deviations from the daily feeding routine in standing, lateral lying and drinking behaviours. These experiments demonstrate that the method is capable of robustly and accurately monitoring individual pig behaviours under commercial conditions, without the need for additional sensors or individual pig identification, hence providing a scalable technology to improve the health and well-being of farm animals. The method has the potential to transform how livestock are monitored and address issues in livestock farming, such as targeted treatment of individuals with medication.

Practices and Applications of Convolutional Neural Network-Based Computer Vision Systems in Animal Farming: A Review

“…However, lighting conditions in farms are complex and diverse, and color-space patterns learned from development datasets may not be matched to those in real applications, which may lead to poor generalization performance [115]. One solution was to convert RGB images (three channels) into grayscale images (one channel) [48,55,70,88,116], so that attention of models can be diverted from object colors to learning patterns of objects. Additionally, hue, saturation, and value (HSV) imaging was not as sensitive to illumination changes as RGB imaging and may be advantageous on detecting characteristics of colorful target objects [114].…”

“…Labels of pose estimation generally consist of a series of key points, and available tools for pose estimation were DeepPoseKit [97] and DeepLabCut [132]. Tracking labels involve assigned IDs or classes for target objects through continuous frames, and the tools for tracking were Kanade-Lucas-Tomasi tracker [88], Interact Software [72], and Video Labeler [68]. It should be noted that references of the aforementioned tools only indicate sources of label tool applications in animal farming rather than developer sources, which can be found in Table 2.…”

“…were utilized to detect and locate animals in images, and then the animals were tracked based on their geometric features in continuous frames using different extra algorithms. These algorithms included popular trackers (e.g., simple online and real-time tracking, SORT; Hungarian algorithm; Munkres variant of the Hungarian assignment algorithm, MVHAA; spatial-aware temporal response filter, STRF; and channel and spatial reliability discriminative correlation filter tracker, CSRDCF) [61,83,88,96,103] and bounding box tracking based on Euclidean distance changes of centroids and probability of detected objects [71,95,118,125,230].…”

“…There are many augmentation strategies available. Some of them augment datasets with physical operations (e.g., adding images with high exposure to sunlight) [88]; and some others use deep learning techniques (e.g., adversarial training, neural style transfer, and generative adversarial network) to augment data [233], which require engineering knowledge. However, these are not the focuses in this section because of complexity.…”

“…Adding noises includes blurring images with filters (e.g., median filters [100]), introducing additional signals (e.g., Gaussian noises [100]) onto images, adding random shapes (e.g., dark polygons [70]) onto images, and padding pixels around boundaries of images [100]. Changing color space contains altering components (e.g., hues, saturation, brightness, and contrast) in HSV space or in RGB space [88,216], switching illumination styles (e.g., simulate active infrared illumination [59]), and manipulating pixel intensities by multiplying them with various factors [121]. All these may create variations during training and make models avoid repeatedly learning homogenous patterns and thus reduced overfitting.…”