Bag of Visual Words for Cows' Basic Activity Recognition

The manual intervention before the birth dates is very important to reduce the dystocia rate and enhance the survival rate of calves effectively. Information acquisition of the cows' basic sow-activity modes and regulars is one of judgments of manual intervention critical points. Now, the mainstream way has limitations that external sensors are used to obtain information by attaching to the specific positions. So, we proposed a method for recognizing the cows' basic sow-activities by using bag of visual words under the video monitoring in this paper. Firstly, this method detected the remarkable areas of the cow activities in the videos by using spatial-temporal interest points. Secondly, it clustered the quantitative features into visual words and then constructed the visual dictionary for describing the activities. Finally, it used the nearest neighbor classification for the activity classification and recognition. This method was tested in different experimental setting for recognizing the typical basic sow-activities such as walking, lying and look-backing. One was tested on 90 groups of videos under given visual angles and the other on 30 groups of videos under the random perspective. The experiment results show that the method proposed in this paper achieve satisfactory accuracy of cows' basic sow-activity recognition.