Information acquisition of the cow behavior modes and regulars are very important. It is one of judgments for the manual intervention in a particular period. There is certain limitation in the way of obtaining information by using external sensors. So, firstly a modified spatial-temporal local binary pattern for feature presentation was proposed. Secondly through building a visual dictionary, the cow sow-behavior were recognized on the test video. Finally, the basic rule of cow sow-behavior was studied through the statistics occurrence frequency of specific behaviors. For verifying the efficiency of the proposed method, different experimental settings were tested for recognizing the typical sow-activities such as walking, side-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 average accuracy is 94.6% and 88.3%, respectively. The other side, a thirteen-hours video before and after delivery was using to sow-activity recognition and frequency count. The experiment results show that the average occurrence frequencies of look-backing and side-lying are 30 and 21.2, respectively. The occurrence frequencies of above two kinds of behaviors are former-low to after-high. Before delivery, the frequency of look-backing reaches to maximum 8.8 and side-lying reduces to 2.2. The results show that this method proposed reveals the basic principle of cow sow-behavior.