Causal relationship between physical activity and prevention of several diseases has been known for some time. Recently, attempts to quantify dose-response relationship between physical activity and health show that automatic tracking and quantification of the exercise efforts not only help in motivating people but improve health conditions as well. However, no commercial devices are available for weight training and calisthenics. This work tries to overcome this limit, exploiting machine learning technique (particularly Linear Discriminant Analysis, LDA) for analyzing data coming from wearable inertial measurement units, (IMUs) and classifying/counting such exercises. Computational requirements are compatible with embedded implementation and reported results confirm the feasibility of the proposed approach, offering an average accuracy in the detection of exercises on the order of 85%.