Activity detection of suckling piglets based on motion area analysis using frame differences in combination with convolution neural network

• Published in: Computers and electronics in agriculture 2022-03, Vol.194, p.106741, Article 106741
• Main Authors: Ding, Qi-an, Chen, Jia, Shen, Ming-xia, Liu, Long-shen
• Format: Article
• Language: English
• Subjects: Activity value; Animal husbandry; Anomalies; Artificial neural networks; Automation; Frame difference; Light sources; Machine vision; Monitoring; Neural networks; Suckling piglets; Swine; Technical services; YOLOv5
• ISSN: 0168-1699; 1872-7107
• DOI: 10.1016/j.compag.2022.106741

•The object detection model of suckling piglets based on YOLOv5 was established.•A machine vision method (FD-CNN) for activity detection of suckling piglets was proposed.•The activity of the piglets was quantified as a value.•The proposed method could continuously observe the piglets’ activity. The lactation period is the first stage in a piglet’s life cycle, and a piglet’s activity during this time is an important indicator of the growth stage, better activity curve reflecting higher growth state. Detecting changes in physical activity may help in the early detection of unhealthy conditions and help a breeder carry out targeted treatment. However, both light interferences caused by devices such as heat preservation lamps and the difficulty of identifying piglets pose significant challenges for automated monitoring of piglet activity. In this study, we proposed an automated method for monitoring piglet activity. This method, named Frame Differences in combination with Convolution Neural Network (FD-CNN), was used to detect the regions of the active piglets by combining the frame difference method with the YOLOv5s network model. To estimate the whole average activity of piglets during lactation, the ratio of the area with active piglets to the area of all piglets was determined. The changing activity of day-old piglets was analyzed, and their status under different light source conditions was also evaluated. Video traceability was performed to detect the abnormal activity points, and the causes of these anomalies were studied. The results showed that our method could detect motion piglets (precision = 0.936) and quantify the activity status of piglets. When the detection frequencies were 6 s and 1 h, the similarity of activity value between the FD-CNN and manual detection were 58.36% and 78.90%, respectively. After the conversion of the activity metrics to values ranging from 0 to 1, the average activity value of day-old piglets with light was higher than 0.25 and that under the no-light condition was lower than 0.2. The traceability results of the abnormal activity points (activity value above 1) show that excess activity was mainly caused by sows attacking piglets or bumping into limits. The FD-CNN could replace the manual detection of piglet activity during lactation to a certain extent. The backtracking of abnormal activity points is beneficial for the timely detection of abnormal interaction behaviors between sows and piglets and provides technical sup