The automated tracking and analysis of sport activities has become increasingly important in the recent years. While it is already very common in endurance sports, in weight training the tracking is still done mainly manually, which is a tedious task. This work aims at exploring the problem of automated tracking and analysing of weight training exercises by the use of low-power smart fitness devices based on inertial measurement units (IMUs), sensors containing accelerometers and gyroscopes. Therefore, basic state-of-the-art signal and data processing approaches, including various filtering techniques, sensor fusion, time series segmentation and classification methods like hidden Markov models (HMMs), support vector machines (SVMs) and nearest neighbours classifiers, are studied and applied to the specific problem domain. A proof-of-concept approach of the proposed methods is implemented on a purpose-built constrained embedded system. Finally, a comprehensive evaluation based on dumbbell exercises is done. The developed prototype achieves a segmentation misdetection rate of 1.5 %, a classification accuracy of 99.7 % and an average response time of about 300 ms. In conclusion, the results show that the initially specified requirements are met and that an accurate and fast tracking of selected weight training exercises is possible.