Physical activity recognition represents a new frontier of improvement for context-aware applications, and for several other applications related to public health. Activity recognition requires the monitoring of physical activity in unconfined environments, using automatic systems supporting prolonged observation periods, and providing minimal discomfort to the user. Accelerometers reasonably satisfy these requirements and have therefore often been employed to identify physical activity types. This chapter will describe how the different applications of activity recognition would influence the choice of the on-body placement and the number of accelerometers. After that it will be analyzed which sampling frequency is necessary to record an acceleration signal for the purpose of activity pattern recognition, and which is the optimal strategy to segment the recorded signal to improve the recognition performance in daily life. In conclusion, it will be discussed how the user friendliness of accelerometers is influenced by the classification algorithm and by the data processing required for activity recognition.