A microwave Doppler sensor can sense wide range of human motion, from minute displacement of body surface caused by respiration to large movement like walking, without contact. To apply the sensor to monitoring human daily activity for safety and health, it is necessary to estimate the current state from the raw signal of the sensor output. In this paper, we propose a method to detect the following three state, “move
”:the target is changing his/her position or pose, “resp
.”:the target sits still and is breathing, and“hold
”:the target sits still and holds his/her breath. The observed sensor signal is classified by three binary classifiers prepared for each state. For input of the classifiers, three kinds of features, energy, frequency-domain entropy, and histogram are extracted from the raw signal. The result of the experiment showed that the proposed detector surpassed that of previous related works especially in respiration detection. We tested three classification methods, least squares, SVM and AdaBoost. There was no significant difference between the performances of classification methods. Cross validation was done by two different ways of data division. In one way of the data division, both training and test data contain data of all the subjects who participated in the experiment. In the other way, training data do not contain the data which belong to the subject of the test data. As the result of this test, the classifiers trained by other persons'data were not inferior to those trained by data which contain the target's own data. This means that the state detection does not depend on the individual target. The parameters can be configured at the factory.