Online Recognition with Postural Transition Awareness - Smartphone-Based Human Activity Recognition - page 111

Algorithm 3: HAR Online Algorithm for the PTA-HAR system
Require:
a : Triaxial linear acceleration
ω : Triaxial angular velocity
g : Gravity
H 1 ( · ) : Noise reduction transfer function (Sect. 7.3.1)

H 2 ( · ) : Body acceleration transfer function (Sect. 7.3.1)

ˆ( · ) : Feature extraction function (Sect. 7.3.1)

T : Windows size
m : Number of classes
p : activity probability vector p =

[ p 1 ,..., p m ] T

m × m

P : Buffer of probability vectors P ∈ R

P : Filtered buffer of probability vectors
z : Buffer of discrete activity predictions z

s

∈ R

( · ) : Probability filtering function (Sect. 7.3.3.1)

: Discrete filtering function (Sect. 7.3.3.2)
function ProcessInertialSignals( a r (

( · )

t ) , ω r ( t ) )

a ˄ ( t ) = H 1 ( a r ( t )) , // Noise Filtering

ω ( t ) = H 2 ( H 1 ( ω r ( t )))

) =

a ( t H 2 ( a ˄ ( t )) // Body acceleration Extraction

g (

) = ) −

t a ˄ ( t a ( t ) // Gravity extraction

return a ( t ) , g ( t ) , ω ( t )

end
function OnlinePrediction( t , a ( t ) , g ( t ) , ω ( t ) , B , z )

= a t : t ] ,

t ∈ A [ t − T ,..., // Window sampling

= g t : t ] , t ∈

G [ t − T ,...,

= ω t : t ] t ∈ [ t − T ,...,

x = ˆ ( A , G ,) // Feature Extraction and Normalization

∈ {

}

for c 1 ,..., m do // Multiclass SVM

) = w c T x f c ( x + b c

// FFP

/ 1 e ( c f c ( x ) + c )

p i = 1 +

// Prob. Estimation (Sect. 7.3.2)

end

p T
P ( 1 : end − 1 , : )

P =

// Append probability vector

P = ( P )

// Activity probability filtering

c ∗ = P ( s − 1 , c )

argmax c ∈ [1 ,..., m ]

// MAP

c ∗

z =

// Append last activity prediction

z

( 1 : end − 1 )

c ˆ = ( z )

// Discrete filtering and activity estimation

return c ˆ

end