Human action recognition based on estimated weak poses

EURASIP Journal on Advances in Signal Processing

Table 4 Comparison of action recognition accuracy ( % ) in HumanEva between our methods and the method presented in [[1]]

Acc.	Box	Jog	Gest	Walk	T/C	All− T/C	All + T/C
[1]	98.9	99.0	63.7	99.6	no	90.3	No
Std-norm	88.4	75.1	87.6	91.0	80.0	85.5	84.4
Ind-norm	97.1	91.8	91.9	94.6	80.0	93.9	91.1

Classification accuracy is defined as correctly labeled samples over total number of samples (refer to Equation (17)). “Std-norm” and “Ind-norm” refer to standard deviation normalizing method and individually normalizing method (refer to Section Weak pose estimation using GPR). The column “All−T/C” shows the average classification accuracy for all actions excluding “Throw/Catch” and and the column “All + T/C” including “Throw/Catch”. Bold values show the best results of action recognition accuracies averaged over all actions.