[Studies in Computational Intelligence 481] Artur Babiarz, Robert Bieda, Karol Jędrasiak, Aleksander Nawrat (auth.), Aleksander Nawrat, Zygmunt Kuś (eds.) - Vision Based Systemsfor UAV Applications (2013, Sprin

Feature Extraction and HMM-Based Classification of Gait Video Sequences 235
Methods based on the concept of the hidden Markov model [16] play also an
important role in research on gait-based human identification, starting from
application of the generic HMM [17], [18], where the width of the outer contour
of the binarized silhouette of a walking person was chosen as the feature vector.
Different variants of the HMM, such as, inter alia, population HMM [19],
factorial and parallel HMMs [20] found also applications in the considered
research area. The method proposed in [21] is very close to the topic of the present
paper – it transforms sequences of silhouettes into low-dimensional embedding by
manifold learning using the Gaussian Process Latent Variable Model (GP-LVM)
before modeling their dynamics by means of the HMM.
The organization of the paper is as follows. Important components of the
research procedure, including a brief description of the methods used for the
purpose of dimensionality reduction as well as aspects of HMM adaptation to gaitbased
recognition are discussed in section 2. Results of the conducted experiments
are presented and analyzed in section 3. The conclusions are formulated in
section 4.
2 Research Procedure
A collection of silhouettes, extracted from the consecutive frames of a single gait
sequence, was used as a feature vector on which linear (PCA) or non-linear
method (isometric features mapping (Isomap) or locally linear embedding (LLE))
was applied in order to reduce data dimensionality. Subsequently, HMM was
utilized as a classifier of the reduced data.
2.1 Data Source
Similar to the research described in the chapter “Selection of individual gait
features extracted by MPCA applied to video recordings data” 80 sequences from
the dataset A of the CASIA Gait Database (only those recorded from the parallel
view, in case of need reflected in order to unify the direction of motion) were used
for the purpose of gait-based human recognition. Contrary to the requirements of
the MPCA algorithm, the sequences didn’t need to be normalized in spite of their
different length. The average length of a single sequence is about 90 frames and
the bounding box of every video frame was determined with a fixed resolution
100×180. Hence, average number of features characterizing a single gait sequence
comes to 1620000 (100×180×90).
2.2 Dimensionality Reduction Methods
Both linear and non-linear dimensionality reduction methods were applied to
prepare gait data for classification.