Temporal frame redundancy for GEI-based gait recognition

In gait scenes, the speeds of subjects (people) are slower comparatively to the camera. This may cause a redundancy in frame. The objective of this paper is to study the temporal frame redundancy in GEI-based gait recognition and to explain the use of down-sampling strategy as a means for reducing the number of redundant frames. It is argued that the strategy, with reasonable sample rate to provide necessary information, yields superior performance in cases the performance is measured by learning efficiency. Despite its simplicity, the resulting temporal frames contain enough information to perform well on human identification task. The hypothesis is empirically validated by examining the performance of a nearest neighbor classifier on training samples drawn from the standard CASIA Gait Databases (Dataset B) with 128 distinct subjects. To utilize the frame redundancy more efficiently, we propose a down-sampling strategy after gait cycle estimation. The learning curves of the classifier are analyzed with respect to the choice of the down-sampling factor. A nearly-optimal classification performance of the classifier is achieved using a relatively small training sample, showing that gait recognition can be successfully applied to low frame rate human identification problems with affordable computation.