One of the essential and challenging research areas in proactive and ubiquitous computing is recognizing of human activities via sensors. In this paper, we present some preliminary results of recognizing human physical activities using a traditionally time domain features with periodic calculation from the activity signals measured by using a multi-sensor model and artificial neural nets. The features mean, standard deviation, energy and periodicity of a signals are calculated after using a moving average filter for five daily activities such as power walking, jogging, speed running standing and walking. Further, we have considered two static positions through locating module in pant''s pocket and holding in hand to collect more accurate data. Likewise, the experiment results for the three different approaches are compared. The experimental results through the proposed technique show accuracy of 98.0% activity recognition.