Pig Postures Recognition Based On VGG19-GNN-SVM

Animal health issues are increasingly important in modern society, as an animal's posture can reflect its emotional state and overall well-being. In recent years, machine vision technology has been applied to monitor the psychological and physiological health of animals. This study proposes a pig posture behavior recognition method based on the VGG19-GNN-SVM framework. The research utilizes depth cameras to collect data on pigs, which helps reduce the exposure of captured images. The study is primarily divided into two key steps. Firstly, the collected data set undergoes preprocessing, including the segmentation of target objects and watershed segmentation to obtain pig images. Subsequently, a series of features are extracted from the images, such as convex hull, convex hull perimeter, boundaries, depth distances between shoulder joints and hips, convex hull ratio, perimeter boundary, and convex hull area ratio, as well as the depth distance between the shoulders and hips and the ratio of depth distance between the shoulders. Secondly, the features obtained from the preprocessing stage are input into the improved pig posture recognition model (VGG 19-GNN-SVM), which is then trained and tested. The final results indicate that the model exhibits higher performance in accurately recognizing two specific postures, namely, the standing posture and the lateral lying posture. Compared to the other three models (CNN, SVM, and CNN-SVM), this model achieves a higher posture recognition accuracy of 0.9147. Furthermore, the error rate has been reduced to 12 %. It is particularly noteworthy that this model significantly outperforms the CNN-SVM model in terms of accuracy, demonstrating a noticeable improvement in performance.