Cow behaviour pattern recognition using a three-dimensional accelerometer and support vector machines

Automated animal behaviour monitoring systems have become increasingly appealing for research and animal production management purposes. However, many existing systems are suited to measure only one or two behaviour patterns or activity states at a time. We aimed to develop and pilot a method for automatically measuring and recognising several behavioural patterns of dairy cows using a three-dimensional accelerometer and a multi-class support vector machine (SVM). SVM classification models were constructed based on nine features. The models were trained using observations made of the behaviour of 30 cows fitted with a neck collar bearing an accelerometer that recorded horizontal, vertical and lateral acceleration. Measured behaviour patterns included standing, lying, ruminating, feeding, normal and lame walking, lying down, and standing up. Accuracy, sensitivity, precision, and kappa measures were used to evaluate the model performance. The SVM classification models achieved a reasonable recognition of standing (80% sensitivity, 65% precision), lying (80%, 83%), ruminating (75%, 86%), feeding (75%, 81%), walking normally (79%, 79%), and lame walking (65%, 66%). The results were poor for lying down (0%, 0%) and standing up (71%, 29%). The overall performance of the multi-class model was 78% precision with a kappa value of 0.69. Each of the behaviour categories had one or two other behaviour patterns that became confused with them the most. The problematic behaviours were expectedly those that resemble each other in terms of movement. Possible solutions for the problems in classification are presented. In conclusion, accelerometers can be used to easily recognise various behaviour patterns in dairy cows. Support vector machines proved useful in classification of measured behaviour patterns. However, further work is needed to refine the features used in the classification models in order to gain the best possible classification performance. Also the quality of acceleration data needs to be considered to improve the results.