Energy expenditure (EE) estimation is an important factor in tracking personal activity and preventing chronic diseases, such as obesity and diabetes. The challenge is to provide accurate EE estimations in free-living environment through portable and unobtrusive devices. In this paper, we present an experimental study to estimate energy expenditure during sitting, standing and treadmill walking using a smartwatch. We introduce a novel methodology, which aims to improve the EE estimation by first separating sedentary (sitting and standing) and non-sedentary (walking) activities, followed by estimating the walking speeds and then calculating the energy expenditure using advanced machine learning based regression models. Ten young adults participated in the experimental trials. Our results showed that combining activity type and walking speed information with the acceleration counts substantially improved the accuracy of regression models for estimating EE. On average, the activity-based models provided 7% better EE estimation than the traditional acceleration-based models.