Fig 1.
An overview of the gait recognition based on MSM.
From gallery images of class c and input images of a probe subject, similarities between the class c and the subject are calculated based on MSM. This process is repeated for all classes in the database, and the subject is classified based on the similarities.
Fig 2.
(a) Examples of gait images (2 [km/h]) in OU-ISIR Treadmill Dataset A [13], (b) GEI of (a), (c) examples of gait images (2 [km/h]), and (d) GEI of (c).
Fig 3.
Visualizations of eigenvectors ϕc (1st, 4th, 8th, and 16th) of MSM.
The image of the first principal component is similar to GEI.
Fig 4.
Visualization of two canonical vectors (2 [km/h] and 7 [km/h]) for the 1st canonical angle.
Fig 5.
Overview of 2D-PCA MSM (smaller dotted rectangle) and 2D-PCA-R MSM (larger dotted rectangle).
Fig 6.
(a) Correct classification rate (CCR) with respect to the change of the number of eigenvalues of 2D PCA and (b) CCR of the PCA.
Fig 7.
Examples of reconstructed images with K = 1, 4, 8, 32 by 2D PCA.
Blue and red colors min positive and negative values.
Fig 8.
Examples of reconstructed images with K = 1, 4, 8, 32 by the PCA.
Fig 9.
(a) An example original image, (b) image with scanning incident lines at 20 degree, and (c) 20 degree rotated image. (b) and (c) are equivalent from the point of view of resulting intersections of lines with body.
Fig 10.
Examples of rotated images (0, 10, 30, 50, 70, 90, -10, -30, -50, -70, -90 degree).
Fig 11.
The output similarity is used in the classification process in Fig 5.
Fig 12.
(a) Selected rotation angle at each iteration by the boosting method, (b) correct classification rate (CCR) at each rotation angle (-90 ≤ θ ≤ -20 and 20 ≤ θ ≤ 90), and (c) CCR at each rotation angle (-20 ≤ θ ≤ 20) (the dotted line shows the CCR by fusing all rotation angles by the boosting method).
Table 1.
Comparison of correct classification rate (CCR) [%] of each of MSM [12], the proposed 2D-PCA MSM, the proposed 2D-PCA-R MSM (θ = −10), and the proposed B-2D-PCA-R MSM on CASIA-C [14] (slow walk fs).
Fig 13.
(a) Selected canonical angle r at each iteration by the boosting method, and (b) selected dG and dP at each iteration by the boosting method.
Fig 14.
(a) Correct classification rate (CCR) at each iteration by applying the boosting method to similarities of all parameters (rotation angles, canonical angles, and dimensionalities), (b) selected canonical angle r and rotation angle θ, and (c) selected dG and dP.
Table 2.
Comparison of the Correct Classification Rate (CCR) [%] of each of the B-2D-PCA-R MSM, the RSM-based gait recognition [17], the motion-based method [19], the MSM-based method with divided areas [12], and single-support GEI (SSGEI) [18] on CASIA-C [14] (slow walk fs and quick walk fq).
Table 3.
Comparison of average correct classification rate (CCR) [%] of each the proposed B-2D-PCA-R MSM, MSM-based method with divided areas [12], and SSGEI-based method [18].
Table 4.
CCR [%] of the B-2D-PCA-R MSM and the MSM-based method with divided areas [12] (numbers in round brackets) in the cross-speed walking gait recognition.
Table 5.
Averaged equal error rate (EER) [%] at each absolute speed difference between gallery and probe data for the proposed B-2D-PCA-R MSM and the MSM-based method with divided areas [12].
Fig 15.
ROC curves of the MSM-based method with divided areas [12] and the B-2D-PCA-R MSM.
Gallery speed and probe speed are 3 km/h and 7 km/h, respectively.
Fig 16.
ROC curve by the B-2D-PCA-R MSM and the MSM-based method [11] for dataset 1.
Table 6.
Equal error rate (EER) and correct classification ratio (CCR) of MSM-based method [11] and the B-2D-PCA-R MSM, for dataset 1.
Table 7.
EER and CCR of Mansur’s method [8], MSM-based method [11] and the B-2D-PCA-R MSM, for dataset 2.
Fig 17.
Specificities and sensitivities (defined as TN / (TN + FP) and FN / (FN + TP), respectively) by changing thresholds of 2D-PCA MSM and B-2D-PCA-R MSM.