HierarchicalTemporalandSpatialMemoryforGaitPatternRecognition Open Access
Downloadable ContentDownload PDF
Many pattern recognition problems are inherently related to space and time. Forexample, when we want to recognize someone by observing his walking patterns,every time step there is a spatial structure of his body parts’ positions, and thosepositions tend to vary with time. Thus, a model that can exploit both the spatial andtemporal structure of the visual world for classification is necessary for this kind oftask. This research extends a new model - Hierarchical Temporal Memory, which isa biologically - inspired algorithm that can perform spatial and temporal learningsimultaneously. By claiming that patterns that come close by in time are likely to bevariations of the same thing, HTM is able to learn the invariant representation of anobject by exploiting the temporal proximity of patterns; by claiming that the worldhas a hierarchical structure, HTM is able to learn complex object in terms of simplerbuilding blocks. The contribution of this research is adding the hierarchical temporalinference mechanism to the existing HTM model, so that the extended model is ableto better exploit the temporal structure of the visual event and performing sequentialinference. We evaluated the algorithm with the task of gait recognition, and the resultshows that our algorithm performs better than some of the current methods and noworse than others on a gait dataset of 151 individuals.