Although smoking prevalence is declining in many countries, smoking related health problems still leads the preventable causes of death in the world. Several smoking intervention mechanisms have been introduced to help smoking cessation. However, these methods are inefficient since they lack in providing real time personalized intervention messages to the smoking addicted users. To address this challenge, the first step is to build an automated smoking behavior detection system. In this study, we propose an accelerometer sensor based non-invasive and automated framework for smoking behavior detection. We built a prototype device to collect data from several participants performing smoking and other five confounding activities. We used three different classifiers to compare activity detection performance using the extracted features from accelerometer data. Our evaluation demonstrates that the proposed approach is able to classify smoking activity among the confounding activities with high accuracy. The proposed system shows the potential for developing a real time automated smoking activity detection and intervention framework.