CAPÍTULO IV: MARCO REAL
4.5. ESPACIOS ABIERTOS
4.9.3. H OTEL V ELA B ARCELONA
4.9.3.3. Características
A gait period starts with the heel strike of either foot and ends with the subsequent heel strike of the same foot comprising of two steps [7] (presented in Chapter 3 and summarised in this chapter). Each foot in a gait period transits between two phases: a stance phase when the foot remains in contact with the ground and a swing phase when the foot does not touch the ground. The stance phase has the following components: (a) initial contact; (b) double support; (c) mid-stance; and (d) propulsion. The swing phase has the following components: (a) pre-swing; (b) mid-swing; and (c) ending swing as illustrated in Fig. 7.3. In GR-LSTR, the seven key frames corresponding to each component of the stance and swing phases of a gait period as shown in Fig. 7.3 are extracted from a gait sequence. These key frames are characterized by distinguishable shape characteristics from each other in a gait period, and therefore effective for shape-based gait recognition.
Since it is not possible to determine gait period from a low frame-rate gait sequence due to unavailability of adequate number of frames, GR-LSTR extracts the seven key frames (i.e., initial contact, ending swing, mid-swing, pre-swing, propulsion, midstance and dou- ble support) from a gait sequence using ROI based silhouette matching by Krawtchouk mo-
Figure 7.3: Selected key frames of a gait period (a)-(h) of a subject: stance phase (a)-(d); and swing phase (e)-(g).
ments. Krawtchouk moments have better image reconstruction capability than the Zernike and Hu moments in both noisy and noise-free conditions, and the orthogonal property of weighted Krawtchouk moments ensures minimal information redundancy [108]. They are also useful when dealing with partially distorted frames of a gait period, as they can be used to extract local features from any ROI of an image by varying the parametersN and
M. The Krawtchouk moments of order (n+m) of aN×Msilhouette with intensity function
f(x,y) are computed using the sets of weighted Krawtchouk polynomialsKn(x;p,N) and
Kn(x;p,M) as in Eq. (4.1), Eq. (4.2), Eq. (4.3) and Eq. (4.4). The silhouettes of the seven
key frames in Fig. 7.3(a)-(g) are manually extracted. The bottom segment of the bounding rectangle, i.e., the region bounded by bottom of the bounding rectangle and the anatomical position of just before the subject’s hand measured from bottom of the bounding rectan- gle (i.e., 0.377HwhereHis height of the bounding rectangle) is set as the Rf-ROI as this foreground region is not distorted by self-occlusions due to arm-swing. To obtain the key frames from any gait sequence automatically, the Rf-ROIs are compared with the same sil- houette segments of all frames of a subject’s gait sequence (each referred to as a (Tr-ROI)) using silhouette comparison based on weighted Krawtchouk moments to obtain similarity scores, i.e., Sscore = [ (Rf-ROIkmn −Tr-ROIkmn) 2]12 , (7.3) where Rf-ROIkmn and Tr-ROIkmn respectively denote the (n+m) order weighted Krawtchouk
moments of the Rf-ROI and Tr-ROI. The frame whose Tr-ROI results in the lowestSscore
with the corresponding Rf-ROI is extracted as one of the seven key frames, and stored in an array. The process continues by comparing the next Rf-ROI with the remaining Tr-ROIs to obtain all key frames one at a time based on the lowestSscore. As a result, an array of seven
key frames are generated. Unlike the silhouette-based gait recognition method STM-SPP in [7] which uses contour matching based on Hu moments for the detection of ten phases of gait period, GR-LSTR computes (n+m) order weighted Krawtchouk moments of each
of the Rf-ROIs and Tr-ROIs using Eq. 4.1 by suitably choosing the values ofMandN for detecting seven key frames from a gait sequence.
The preprocessed binary silhouette imagesBp(x,y) at the pth key frame are com-
bined to form grey-level GKI using
GKI= 1 N N ∑ p=1 Bp(x,y), (7.4)
whereN=7, andx,ydenote the coordinates of a binary silhouette. Similar to GEI in [25] which is robust against silhouette noise in individual frames of a gait period, GKI is also less sensitive to silhouette noise in the constituent key frames of a gait sequence as GKI is formed by averaging the key frames of a gait sequence. Gait period detection is essential prior to GEI formation as GEI is formed by averaging all the frames of a gait period of a subject. Since it is impractical to determine the gait period of a subject from low frame- rate gait sequences, GEI is not effective for low temporal resolution gait recognition. GKI preserves significant gait characteristics of a gait sequence in reduced storage by averaging its key frames and hence, it effectively addresses the challenge of low frame-rate gait se- quences. Fig. 7.4(a) and (b) respectively show the GKIs of two subjects of OU-ISIR gait dataset D.
(a) (b)
Figure 7.4: GKIs constructed from two gait sequences of OU-ISIR gait dataset: (a) subject from DBhigh subset; and (b) subject from DBlow subset (see Section 7.4 for DBhigh and
DBlow).