@InProceedings{Ataya2013,
  author    = {Ataya, A. and Jallon, P. and Bianchi, P. and Doron, M.},
  title     = {Improving Activity Recognition using Temporal Coherence},
  booktitle = {Proceedings of the International Conference of the IEEE Engineering in Medicine and Biology Society},
  year      = {2013},
  pages     = {4215--4218},
  abstract  = {Assessment of daily physical activity using data from wearable sensors has recently become a prominent research area in the biomedical engineering field and a substantial application for pattern recognition. In this paper, we present an accelerometer-based activity recognition scheme on the basis of a hierarchical structured classifier. A first step consists of distinguishing static activities from dynamic ones in order to extract relevant features for each activity type. Next, a separate classifier is applied to detect more specific activities of the same type. On top of our activity recognition system, we introduce a novel approach to take into account the temporal coherence of activities. Inter-activity transition information is modeled by an oriented graph Markov chain. Confidence measures in activity classes are then evaluated from conventional classifier's outputs and coupled with the graph to reinforce activity estimation. Accurate results and significant improvement of activity detection are obtained when applying our system for the recognition of 9 activities for 48 subjects.},
  groups    = {STAT841},
  keywords  = {Biomedical signal classification, Coherence in biomedical signal processing, Markov models in signal pattern classification, EMBC2013},
  review    = {Previous work found that, in general, discriminative approaches work better than generative approaches in classification tasks. However, discriminative classifiers typically assume the data is time-independent, which is not valid for the type of actions we are considering. Hybrid approaches seems promising. 

A graphical method is used, on 1 hip mounted accelerometer
- Various passive filters used. Median filter to remove noise spikes. LPF and HPF used. HPF to remove accel incline info. LPF to remove movement data. The LPF and the HPF data are used separately. 
- Uses a binary tree to determine if the person is static (posture) or dynamic (activity). This way, the system can be selective at which features they use for different circumstances. But later on it says they uses the Signal Magnitude Area on the HPF, which seems to be the sum of all the magnitudes over all 3 axes, divided by t. If it's above some threshold, there is movement. Otherwise, static. 
 - LPF for static postures. Looks at 14 different features: average mean values, temporal energy, average of L-1 norm of acceleration vector,
sensor�s tilt angle from ground, area under the curve and
mean distances between axis
 - HPF for dynamic activities: 18 different features: median frequencies
of the 3 axis, entropy, mean-cross rate, peak-to-peak
distances, mutual correlation between axis, and spectral
energy
- Uses a 2 second sliding window that advances 1 second at a time

Emphasizes that people do not suddenly shift from one exercise to another, This graph method is suppose to connect one logical activity to another, to reduce spurious classifications (sounds kind of like the HMM transmission mtx). Ah they use a Markov chain. So they figure out the initial and transmission probability by prior training. They then use Viterbi to figure out what the most likely next motion is.


Collected 48 healthy subjects, mean age 36 yo. Motions labelled by medical experts. Performed common everyday activities. Among these activities, 3 are postures or static activities (lying down, slouching/sitting, and standing) and the other 6 activities correspond to motion or dynamic activities (stamping, cycling, running, slow walking, fast walking, and using stairs).

So the different discrim classifers are tested, and its output are used to train the overall HMM or graph method. Does better with graph.


Ataya \etal \cite{Ataya2013} assert that, in general. discriminative approaches work better than generative approaches in classification tasks. However, discriminative classifiers typically assume the data is time-independent, which is not valid for all types of data, in particular that of human movement. A hybrid approach is proposed. The paper employs 1 hip-mounted accelerometer. A decision tree is used to determine if the person is static (holding a given posture) or dynamic (performing an activity). This allows the system to select specific features to analyze the movement. For static postures, the data is low-pass filtered (LPF) to remove noise, then 14 different features to calculated in order to determine the type of motion being performed. These features include average mean, temporal energy, as well as the norm of the acceleration vector. For dynamic movement, the data is high-pass filtered (HPF) to remove sensor DC bias, then 18 different features are calculated. These features include median frequencies, entropy, and mean-crossing rate. It can be noted that people generally do not suddenly shift from one exercise to another, so a pre-generated Markov chain can be used to check that the exercise being suggested by the features are logical. This system was assessed against 48 healthy subjects performing common everyday activities, with 88\% recall accuracy. \todo{come back to this. they use Viterbi to segment}

Ataya \etal \cite{Ataya2013} employed 1 hip-mounted accelerometer and examined numerous features and assessed the classification strengths with several common classifiers, where random forest was shown to perform the best. 9 different postures and activities, such as lying down, slouching, standing, cycling and running, were examined. Data was collected from 48 different subjects, with 55 minutes of accelerometer data each. The outputs of these classifiers are compared against a pre-trained Markov chain and the Viterbi algorithm is used to verify the activity labels proposed by the classifiers. $Ver_{AllPoints}$ verification scheme is used, and $Acc_{Class}$ is 89\%},
  timestamp = {2013.07.29},
}

{"_id":"dig4N3ewtXxwKpN5d","bibbaseid":"ataya-jallon-bianchi-doron-improvingactivityrecognitionusingtemporalcoherence-2013","downloads":0,"creationDate":"2017-09-14T16:34:36.177Z","title":"Improving Activity Recognition using Temporal Coherence","author_short":["Ataya, A.","Jallon, P.","Bianchi, P.","Doron, M."],"year":2013,"bibtype":"inproceedings","biburl":"https://raw.githubusercontent.com/jfslin/jfslin.github.io/master/jf2lin.bib","bibdata":{"bibtype":"inproceedings","type":"inproceedings","author":[{"propositions":[],"lastnames":["Ataya"],"firstnames":["A."],"suffixes":[]},{"propositions":[],"lastnames":["Jallon"],"firstnames":["P."],"suffixes":[]},{"propositions":[],"lastnames":["Bianchi"],"firstnames":["P."],"suffixes":[]},{"propositions":[],"lastnames":["Doron"],"firstnames":["M."],"suffixes":[]}],"title":"Improving Activity Recognition using Temporal Coherence","booktitle":"Proceedings of the International Conference of the IEEE Engineering in Medicine and Biology Society","year":"2013","pages":"4215--4218","abstract":"Assessment of daily physical activity using data from wearable sensors has recently become a prominent research area in the biomedical engineering field and a substantial application for pattern recognition. In this paper, we present an accelerometer-based activity recognition scheme on the basis of a hierarchical structured classifier. A first step consists of distinguishing static activities from dynamic ones in order to extract relevant features for each activity type. Next, a separate classifier is applied to detect more specific activities of the same type. On top of our activity recognition system, we introduce a novel approach to take into account the temporal coherence of activities. Inter-activity transition information is modeled by an oriented graph Markov chain. Confidence measures in activity classes are then evaluated from conventional classifier's outputs and coupled with the graph to reinforce activity estimation. Accurate results and significant improvement of activity detection are obtained when applying our system for the recognition of 9 activities for 48 subjects.","groups":"STAT841","keywords":"Biomedical signal classification, Coherence in biomedical signal processing, Markov models in signal pattern classification, EMBC2013","review":"Previous work found that, in general, discriminative approaches work better than generative approaches in classification tasks. However, discriminative classifiers typically assume the data is time-independent, which is not valid for the type of actions we are considering. Hybrid approaches seems promising. A graphical method is used, on 1 hip mounted accelerometer - Various passive filters used. Median filter to remove noise spikes. LPF and HPF used. HPF to remove accel incline info. LPF to remove movement data. The LPF and the HPF data are used separately. - Uses a binary tree to determine if the person is static (posture) or dynamic (activity). This way, the system can be selective at which features they use for different circumstances. But later on it says they uses the Signal Magnitude Area on the HPF, which seems to be the sum of all the magnitudes over all 3 axes, divided by t. If it's above some threshold, there is movement. Otherwise, static. - LPF for static postures. Looks at 14 different features: average mean values, temporal energy, average of L-1 norm of acceleration vector, sensor�s tilt angle from ground, area under the curve and mean distances between axis - HPF for dynamic activities: 18 different features: median frequencies of the 3 axis, entropy, mean-cross rate, peak-to-peak distances, mutual correlation between axis, and spectral energy - Uses a 2 second sliding window that advances 1 second at a time Emphasizes that people do not suddenly shift from one exercise to another, This graph method is suppose to connect one logical activity to another, to reduce spurious classifications (sounds kind of like the HMM transmission mtx). Ah they use a Markov chain. So they figure out the initial and transmission probability by prior training. They then use Viterbi to figure out what the most likely next motion is. Collected 48 healthy subjects, mean age 36 yo. Motions labelled by medical experts. Performed common everyday activities. Among these activities, 3 are postures or static activities (lying down, slouching/sitting, and standing) and the other 6 activities correspond to motion or dynamic activities (stamping, cycling, running, slow walking, fast walking, and using stairs). So the different discrim classifers are tested, and its output are used to train the overall HMM or graph method. Does better with graph. Ataya \\etal i̧teAtaya2013 assert that, in general. discriminative approaches work better than generative approaches in classification tasks. However, discriminative classifiers typically assume the data is time-independent, which is not valid for all types of data, in particular that of human movement. A hybrid approach is proposed. The paper employs 1 hip-mounted accelerometer. A decision tree is used to determine if the person is static (holding a given posture) or dynamic (performing an activity). This allows the system to select specific features to analyze the movement. For static postures, the data is low-pass filtered (LPF) to remove noise, then 14 different features to calculated in order to determine the type of motion being performed. These features include average mean, temporal energy, as well as the norm of the acceleration vector. For dynamic movement, the data is high-pass filtered (HPF) to remove sensor DC bias, then 18 different features are calculated. These features include median frequencies, entropy, and mean-crossing rate. It can be noted that people generally do not suddenly shift from one exercise to another, so a pre-generated Markov chain can be used to check that the exercise being suggested by the features are logical. This system was assessed against 48 healthy subjects performing common everyday activities, with 88% recall accuracy. \\todocome back to this. they use Viterbi to segment Ataya \\etal i̧teAtaya2013 employed 1 hip-mounted accelerometer and examined numerous features and assessed the classification strengths with several common classifiers, where random forest was shown to perform the best. 9 different postures and activities, such as lying down, slouching, standing, cycling and running, were examined. Data was collected from 48 different subjects, with 55 minutes of accelerometer data each. The outputs of these classifiers are compared against a pre-trained Markov chain and the Viterbi algorithm is used to verify the activity labels proposed by the classifiers. $Ver_{AllPoints}$ verification scheme is used, and $Acc_{Class}$ is 89%","timestamp":"2013.07.29","bibtex":"@InProceedings{Ataya2013,\n author = {Ataya, A. and Jallon, P. and Bianchi, P. and Doron, M.},\n title = {Improving Activity Recognition using Temporal Coherence},\n booktitle = {Proceedings of the International Conference of the IEEE Engineering in Medicine and Biology Society},\n year = {2013},\n pages = {4215--4218},\n abstract = {Assessment of daily physical activity using data from wearable sensors has recently become a prominent research area in the biomedical engineering field and a substantial application for pattern recognition. In this paper, we present an accelerometer-based activity recognition scheme on the basis of a hierarchical structured classifier. A first step consists of distinguishing static activities from dynamic ones in order to extract relevant features for each activity type. Next, a separate classifier is applied to detect more specific activities of the same type. On top of our activity recognition system, we introduce a novel approach to take into account the temporal coherence of activities. Inter-activity transition information is modeled by an oriented graph Markov chain. Confidence measures in activity classes are then evaluated from conventional classifier's outputs and coupled with the graph to reinforce activity estimation. Accurate results and significant improvement of activity detection are obtained when applying our system for the recognition of 9 activities for 48 subjects.},\n groups = {STAT841},\n keywords = {Biomedical signal classification, Coherence in biomedical signal processing, Markov models in signal pattern classification, EMBC2013},\n review = {Previous work found that, in general, discriminative approaches work better than generative approaches in classification tasks. However, discriminative classifiers typically assume the data is time-independent, which is not valid for the type of actions we are considering. Hybrid approaches seems promising. \n\nA graphical method is used, on 1 hip mounted accelerometer\n- Various passive filters used. Median filter to remove noise spikes. LPF and HPF used. HPF to remove accel incline info. LPF to remove movement data. The LPF and the HPF data are used separately. \n- Uses a binary tree to determine if the person is static (posture) or dynamic (activity). This way, the system can be selective at which features they use for different circumstances. But later on it says they uses the Signal Magnitude Area on the HPF, which seems to be the sum of all the magnitudes over all 3 axes, divided by t. If it's above some threshold, there is movement. Otherwise, static. \n - LPF for static postures. Looks at 14 different features: average mean values, temporal energy, average of L-1 norm of acceleration vector,\nsensor�s tilt angle from ground, area under the curve and\nmean distances between axis\n - HPF for dynamic activities: 18 different features: median frequencies\nof the 3 axis, entropy, mean-cross rate, peak-to-peak\ndistances, mutual correlation between axis, and spectral\nenergy\n- Uses a 2 second sliding window that advances 1 second at a time\n\nEmphasizes that people do not suddenly shift from one exercise to another, This graph method is suppose to connect one logical activity to another, to reduce spurious classifications (sounds kind of like the HMM transmission mtx). Ah they use a Markov chain. So they figure out the initial and transmission probability by prior training. They then use Viterbi to figure out what the most likely next motion is.\n\n\nCollected 48 healthy subjects, mean age 36 yo. Motions labelled by medical experts. Performed common everyday activities. Among these activities, 3 are postures or static activities (lying down, slouching/sitting, and standing) and the other 6 activities correspond to motion or dynamic activities (stamping, cycling, running, slow walking, fast walking, and using stairs).\n\nSo the different discrim classifers are tested, and its output are used to train the overall HMM or graph method. Does better with graph.\n\n\nAtaya \\etal \\cite{Ataya2013} assert that, in general. discriminative approaches work better than generative approaches in classification tasks. However, discriminative classifiers typically assume the data is time-independent, which is not valid for all types of data, in particular that of human movement. A hybrid approach is proposed. The paper employs 1 hip-mounted accelerometer. A decision tree is used to determine if the person is static (holding a given posture) or dynamic (performing an activity). This allows the system to select specific features to analyze the movement. For static postures, the data is low-pass filtered (LPF) to remove noise, then 14 different features to calculated in order to determine the type of motion being performed. These features include average mean, temporal energy, as well as the norm of the acceleration vector. For dynamic movement, the data is high-pass filtered (HPF) to remove sensor DC bias, then 18 different features are calculated. These features include median frequencies, entropy, and mean-crossing rate. It can be noted that people generally do not suddenly shift from one exercise to another, so a pre-generated Markov chain can be used to check that the exercise being suggested by the features are logical. This system was assessed against 48 healthy subjects performing common everyday activities, with 88\\% recall accuracy. \\todo{come back to this. they use Viterbi to segment}\n\nAtaya \\etal \\cite{Ataya2013} employed 1 hip-mounted accelerometer and examined numerous features and assessed the classification strengths with several common classifiers, where random forest was shown to perform the best. 9 different postures and activities, such as lying down, slouching, standing, cycling and running, were examined. Data was collected from 48 different subjects, with 55 minutes of accelerometer data each. The outputs of these classifiers are compared against a pre-trained Markov chain and the Viterbi algorithm is used to verify the activity labels proposed by the classifiers. $Ver_{AllPoints}$ verification scheme is used, and $Acc_{Class}$ is 89\\%},\n timestamp = {2013.07.29},\n}\n\n","author_short":["Ataya, A.","Jallon, P.","Bianchi, P.","Doron, M."],"key":"Ataya2013","id":"Ataya2013","bibbaseid":"ataya-jallon-bianchi-doron-improvingactivityrecognitionusingtemporalcoherence-2013","role":"author","urls":{},"keyword":["Biomedical signal classification","Coherence in biomedical signal processing","Markov models in signal pattern classification","EMBC2013"],"downloads":0},"search_terms":["improving","activity","recognition","using","temporal","coherence","ataya","jallon","bianchi","doron"],"keywords":["biomedical signal classification","coherence in biomedical signal processing","markov models in signal pattern classification","embc2013"],"authorIDs":[],"dataSources":["iCsmKnycRmHPxmhBd"]}