Ant-based clustering of visual-words for unsupervised human action recognition

Abstract

Ant-based clustering is a biologically-inspired computational heuristic that has been used in various domains for general clustering tasks. In this paper we propose its use as the tool for clustering high-dimensional vectors (visual words) which are descriptive features for human actions extracted from video sequences. This codebook generation stage is critical in the popular ‘Bag-of-Words’ framework in which a visual codebook is constructed on the statistics of various features in images or videos. K-means algorithm is widely used in this process but this has two major shortcomings namely: it requires user specification of input parameter k which can bias the algorithm and make it converge at a sub-optimal number of clusters. Also, optimal value of k needs to be determined empirically. Our method generates a codebook of highly descriptive spatio-temporal ‘words’ using ant-based clustering to determine the optimal number of clusters in the dataset. The number of clusters generated was set as the number of codewords for the vocabulary. The limitations of k-means were overcome with the robustness of ant-based clustering heuristic. This idea when applied to the benchmark KTH database produced codewords that produced a compact representation of human actions which gave the desired recognition result when compared to similar approach based on k-means clustering.