Limits...
Robust Indoor Human Activity Recognition Using Wireless Signals.

Wang Y, Jiang X, Cao R, Wang X - Sensors (Basel) (2015)

Bottom Line: Then, an online filtering method is designed to make actions' CSI curves smooth and allow them to contain enough pattern information.Each primitive action pattern can be segmented from the outliers of its multi-input multi-output (MIMO) signals by a proposed segmentation method.Lastly, in online activities recognition, by selecting proper features and Support Vector Machine (SVM) based multi-classification, activities constituted by primitive actions can be recognized insensitive to the locations, orientations, and speeds.

View Article: PubMed Central - PubMed

Affiliation: School of Software, Dalian University of Technology, Dalian 116620, China. dlutwangyi@dlut.edu.cn.

ABSTRACT
Wireless signals-based activity detection and recognition technology may be complementary to the existing vision-based methods, especially under the circumstance of occlusions, viewpoint change, complex background, lighting condition change, and so on. This paper explores the properties of the channel state information (CSI) of Wi-Fi signals, and presents a robust indoor daily human activity recognition framework with only one pair of transmission points (TP) and access points (AP). First of all, some indoor human actions are selected as primitive actions forming a training set. Then, an online filtering method is designed to make actions' CSI curves smooth and allow them to contain enough pattern information. Each primitive action pattern can be segmented from the outliers of its multi-input multi-output (MIMO) signals by a proposed segmentation method. Lastly, in online activities recognition, by selecting proper features and Support Vector Machine (SVM) based multi-classification, activities constituted by primitive actions can be recognized insensitive to the locations, orientations, and speeds.

Show MeSH

Related in: MedlinePlus

The comparison of similarity matrixes for 13 primitive actions between original features and selected features. The hue demonstrates the similarity between any two actions. (a) The similarity matrix for the original features; (b) The similarity matrix for the selected features.
© Copyright Policy
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC4541929&req=5

sensors-15-17195-f005: The comparison of similarity matrixes for 13 primitive actions between original features and selected features. The hue demonstrates the similarity between any two actions. (a) The similarity matrix for the original features; (b) The similarity matrix for the selected features.

Mentions: As described in Section 3.3, we have 24 features for an action from the statistic data of each MIMO subplot, but they are set artificially and may have some redundancy, so a feature selection process was added. First of all, a forward selection process was performed by adding features one-by-one. The error rate of classification was taken as the evaluation function. Then, a backward selection step was performed by deleting the features one-by-one from the feature sets. The results of the feature selection were 14 features. Figure 5a demonstrates the similarities of the original 24 features for actions AC1 to AC13 and Figure 5b demonstrates the similarities of the selected 14 features. It is obvious that the selected features have more discriminative abilities. For the actions of different periods, DTW [22] was adopted and gave a similarity value instead of that of the feature PA (the period of an action).


Robust Indoor Human Activity Recognition Using Wireless Signals.

Wang Y, Jiang X, Cao R, Wang X - Sensors (Basel) (2015)

The comparison of similarity matrixes for 13 primitive actions between original features and selected features. The hue demonstrates the similarity between any two actions. (a) The similarity matrix for the original features; (b) The similarity matrix for the selected features.
© Copyright Policy
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC4541929&req=5

sensors-15-17195-f005: The comparison of similarity matrixes for 13 primitive actions between original features and selected features. The hue demonstrates the similarity between any two actions. (a) The similarity matrix for the original features; (b) The similarity matrix for the selected features.
Mentions: As described in Section 3.3, we have 24 features for an action from the statistic data of each MIMO subplot, but they are set artificially and may have some redundancy, so a feature selection process was added. First of all, a forward selection process was performed by adding features one-by-one. The error rate of classification was taken as the evaluation function. Then, a backward selection step was performed by deleting the features one-by-one from the feature sets. The results of the feature selection were 14 features. Figure 5a demonstrates the similarities of the original 24 features for actions AC1 to AC13 and Figure 5b demonstrates the similarities of the selected 14 features. It is obvious that the selected features have more discriminative abilities. For the actions of different periods, DTW [22] was adopted and gave a similarity value instead of that of the feature PA (the period of an action).

Bottom Line: Then, an online filtering method is designed to make actions' CSI curves smooth and allow them to contain enough pattern information.Each primitive action pattern can be segmented from the outliers of its multi-input multi-output (MIMO) signals by a proposed segmentation method.Lastly, in online activities recognition, by selecting proper features and Support Vector Machine (SVM) based multi-classification, activities constituted by primitive actions can be recognized insensitive to the locations, orientations, and speeds.

View Article: PubMed Central - PubMed

Affiliation: School of Software, Dalian University of Technology, Dalian 116620, China. dlutwangyi@dlut.edu.cn.

ABSTRACT
Wireless signals-based activity detection and recognition technology may be complementary to the existing vision-based methods, especially under the circumstance of occlusions, viewpoint change, complex background, lighting condition change, and so on. This paper explores the properties of the channel state information (CSI) of Wi-Fi signals, and presents a robust indoor daily human activity recognition framework with only one pair of transmission points (TP) and access points (AP). First of all, some indoor human actions are selected as primitive actions forming a training set. Then, an online filtering method is designed to make actions' CSI curves smooth and allow them to contain enough pattern information. Each primitive action pattern can be segmented from the outliers of its multi-input multi-output (MIMO) signals by a proposed segmentation method. Lastly, in online activities recognition, by selecting proper features and Support Vector Machine (SVM) based multi-classification, activities constituted by primitive actions can be recognized insensitive to the locations, orientations, and speeds.

Show MeSH
Related in: MedlinePlus