Rapid Measurements of Pig Body Size Based on DeepLabCut Algorithm

Abstract:

At present, the computer vision-based pig body measurement shows a high dependence on pig posture and low measurement efficiency. To solve these problems, a rapid and non-contact pig body size measurement method based on DeepLabCut was proposed. The top view RGB-D images of landrace pigs were captured by a RealSense L515 camera. The training effects of 10 backbone networks of ReNet, MobileNet-V2, and EfficientNet series were compared and analyzed, and then the EfficientNet-b6 model was selected as the optimal backbone network of DeepLabCut algorithm for feature point detection of pig body size. In order to achieve accurate calculation of pig body size data, SVM model was used to identify the standing stance of pigs and screen the natural standing stance of pigs. Based on this, the depthvalued proximity region replacement algorithm was used to optimize the outlier feature points and calculate the five body size indexes of pig body length, body width, body height, rump width and rump height by Euclidean distance. This method was tested on 140 groups of standing images of pigs, and it was found that the algorithm could achieve real-time and accurate measurement of body size in the natural standing posture of pigs, with maximum root mean square error of 1.79cm and computation time of 0.27s per frame.