This paper presents the development and evaluation of algorithms for detecting physical activities using data from five small biaxial accelerometers worn simultaneously on different body parts. The data was collected from 20 subjects without researcher supervision, who were asked to perform a sequence of everyday tasks. Various features, including mean, energy, frequency-domain entropy, and correlation of acceleration data, were calculated, and several classifiers were tested. Decision tree classifiers achieved the best performance with an overall accuracy rate of 84%. The study found that while some activities can be recognized well with subject-independent training data, others require subject-specific training. Multiple accelerometers, particularly those placed on the thigh and wrist, significantly improved recognition performance, though the addition of just two accelerometers resulted in only a slight drop in performance. This work is the first to investigate the performance of recognition algorithms using multiple, wire-free accelerometers on 20 activities, with datasets annotated by the subjects themselves. The research highlights the importance of training and testing activity recognition systems on naturalistic data to ensure robust performance in real-world conditions.This paper presents the development and evaluation of algorithms for detecting physical activities using data from five small biaxial accelerometers worn simultaneously on different body parts. The data was collected from 20 subjects without researcher supervision, who were asked to perform a sequence of everyday tasks. Various features, including mean, energy, frequency-domain entropy, and correlation of acceleration data, were calculated, and several classifiers were tested. Decision tree classifiers achieved the best performance with an overall accuracy rate of 84%. The study found that while some activities can be recognized well with subject-independent training data, others require subject-specific training. Multiple accelerometers, particularly those placed on the thigh and wrist, significantly improved recognition performance, though the addition of just two accelerometers resulted in only a slight drop in performance. This work is the first to investigate the performance of recognition algorithms using multiple, wire-free accelerometers on 20 activities, with datasets annotated by the subjects themselves. The research highlights the importance of training and testing activity recognition systems on naturalistic data to ensure robust performance in real-world conditions.