Analysing gait sequences using Latent Dirichlet Allocation for certain human actions

Abstract

Conventional human action recognition algorithm and method generate coarse clusters of input videos approximately 2-4 clusters with less information regarding the cluster generation. This problem is solved by proposing Latent Dirichlet Allocation algorithm that transforms the extracted gait sequences in gait domain into documents-words in text domain. These words are then used to group the input documents into finer clusters approximately 8-9 clusters. In this approach, we have made an attempt to use gait analysis in recognizing human actions, where the gait analysis requires to have some motion in lower parts of the human body like leg. As the videos of Weizmann dataset have some actions that exhibits these movements, we are able use these motion parameters to recognize certain human actions. Experiments on Weizmann dataset suggest that the proposed Latent Dirichlet Allocation algorithm is an efficient method for recognizing human actions from the video streams.