Sensors embedded in smartphones are an essential component for activity recognition. Even though the accelerometer is the most widely used sensor, the highest recognition accuracies are obtained when using data collected from multiple sensors. However, the use of multiple sensors has an adverse impact on the energy consumption of power-limited devices such as smartphones. In this paper, we present a new method to improve the recognition accuracy of physical activities by using only the accelerometer. We utilize a low-pass filter to split the acceleration data into a low- and a high-frequency component. These components provide a new set of features, which can be used as a complement to the raw acceleration to reduce the number of sensors needed to recognize physical activities. After evaluating our method for a public dataset, we found that our approach represents an average of up to 16% increase in the recognition accuracy over the raw acceleration data, outperforming even widely used combinations such as the raw acceleration plus the gyroscope. The highest accuracies are obtained when using a cut-off frequency in the interval [0:001--0:05] Hz as well as a combination of the acceleration with its low-frequency component.