USING A CONVOLUTIONAL NEURAL NETWORK TO DETERMINE THE TEAT LIMITS OF A COW’S UDDER

Abstract

In dairy farming, there is an active use of systems for the accurate recognizing of cow’s teats, which is necessary for robotic milking. Object detection consists both in assessing the exact position of the object (localization of the object) and in determining the category to which it belongs (classification of the object). (Research purpose) The research purpose is developing a method for searching for clusters of points to determine the limits of the teat in three-dimensional maps obtained with a 3D ToF camera. (Materials and methods) An IFM O3D303 camera with a resolution of 352 × 264 pixels was used. Authors have developed a structure for analyzing static and dynamic scenes, the model consists of five blocks of convolutional layers, three fully connected layers of neurons and a classifier layer. The first two blocks contain two layers, the other three blocks contain three convolution layers. (Results and discussion) The software highlights the detected teats in color. The method is considered suitable if at 10 repetitions of the experiment the limits of four teats are recognized and there are no artifacts, the experiment can be considered successful, the permissible error of the experiment is no more than 10 percent. During the search for limits of teats, the algorithm repeatedly searches for points in the bounding cylinder of the refined cluster of points. (Conclusions) A method for searching for clusters of points to identify teats limits in three-dimensional maps processing data from a 3D ToF camera using a neural network with «anchor rectangles». Data from 36 animals and 6000 images were collected for training the neural network, field experiments confirmed the accuracy of the neural network in 90 percent. The proposed method allows determining length X, diameter Y and inclination angle Z of the teat in automatic mode.