This paper presents algorithms for detecting physical activities from data collected using five small biaxial accelerometers worn on different body parts. The data was collected from 20 subjects without researcher supervision. The subjects performed a sequence of everyday tasks without specific instructions. Features such as mean, energy, frequency-domain entropy, and correlation of acceleration data were calculated, and several classifiers were tested. Decision tree classifiers achieved an overall accuracy of 84% in recognizing everyday activities. The results indicate that some activities can be recognized with subject-independent training data, while others require subject-specific training. Multiple accelerometers help in recognition because they allow for the discrimination of many activities through conjunctions in acceleration feature values. Using only two accelerometers (thigh and wrist) resulted in only a slight drop in recognition performance. This study is the first to investigate the performance of recognition algorithms with multiple, wire-free accelerometers on 20 activities using datasets annotated by the subjects themselves. The paper addresses the challenge of detecting context from noisy and ambiguous sensor data, particularly in recognizing physical activities. Prior work has focused on acceleration data or its fusion with other data modalities, but it is unclear how these systems perform in real-world conditions. Most prior systems use data collected in laboratory settings, which may not reflect real-world conditions. This study evaluates activity recognition algorithms under conditions similar to real-world settings, using data collected from 20 subjects under laboratory and semi-naturalistic conditions. Supervised learning classifiers were trained on data collected without researcher supervision. Algorithms trained using only user-labeled data could potentially allow users to train algorithms to recognize their own behaviors. Researchers have already developed wearable systems that use acceleration, audio, video, and other sensors to recognize user activities. Advances in miniaturization allow accelerometers to be embedded in wearable devices and wirelessly send data to mobile computing devices. Training and testing activity recognition systems on naturalistic data is important because laboratory environments may artificially constrain subject activity patterns. This study shows that algorithms trained on lab data may not perform well on naturalistic data due to variations in gait patterns. Past works have demonstrated recognition rates of 85% to 95% for ambulation, posture, and other activities using acceleration data. Activity recognition has been performed on acceleration data collected from the hip and multiple body locations. Related work using activity counts and computer vision supports the potential for activity recognition using acceleration. The energy of a subject's acceleration can discriminate sedentary activities from moderate and vigorous activities.This paper presents algorithms for detecting physical activities from data collected using five small biaxial accelerometers worn on different body parts. The data was collected from 20 subjects without researcher supervision. The subjects performed a sequence of everyday tasks without specific instructions. Features such as mean, energy, frequency-domain entropy, and correlation of acceleration data were calculated, and several classifiers were tested. Decision tree classifiers achieved an overall accuracy of 84% in recognizing everyday activities. The results indicate that some activities can be recognized with subject-independent training data, while others require subject-specific training. Multiple accelerometers help in recognition because they allow for the discrimination of many activities through conjunctions in acceleration feature values. Using only two accelerometers (thigh and wrist) resulted in only a slight drop in recognition performance. This study is the first to investigate the performance of recognition algorithms with multiple, wire-free accelerometers on 20 activities using datasets annotated by the subjects themselves. The paper addresses the challenge of detecting context from noisy and ambiguous sensor data, particularly in recognizing physical activities. Prior work has focused on acceleration data or its fusion with other data modalities, but it is unclear how these systems perform in real-world conditions. Most prior systems use data collected in laboratory settings, which may not reflect real-world conditions. This study evaluates activity recognition algorithms under conditions similar to real-world settings, using data collected from 20 subjects under laboratory and semi-naturalistic conditions. Supervised learning classifiers were trained on data collected without researcher supervision. Algorithms trained using only user-labeled data could potentially allow users to train algorithms to recognize their own behaviors. Researchers have already developed wearable systems that use acceleration, audio, video, and other sensors to recognize user activities. Advances in miniaturization allow accelerometers to be embedded in wearable devices and wirelessly send data to mobile computing devices. Training and testing activity recognition systems on naturalistic data is important because laboratory environments may artificially constrain subject activity patterns. This study shows that algorithms trained on lab data may not perform well on naturalistic data due to variations in gait patterns. Past works have demonstrated recognition rates of 85% to 95% for ambulation, posture, and other activities using acceleration data. Activity recognition has been performed on acceleration data collected from the hip and multiple body locations. Related work using activity counts and computer vision supports the potential for activity recognition using acceleration. The energy of a subject's acceleration can discriminate sedentary activities from moderate and vigorous activities.