Human Action Recognition using Improved Vector of Locally Aggregated Descriptors

摘要

Recently, two high-dimensional encoding techniques for human action recognition, namely, Fisher vector (FV) and vector of locally aggregated descriptors (VLAD), are widely employed. In this study, a new human action recognition approach using improved VLAD with localized soft assignment (LSA) and second-order statistics is proposed. When encoding videos into VLAD, instead of considering only the nearest one, we utilize localized soft assignment, i.e., considering multiple nearest visual words. In general, LSA-VLAD captures only the first-order statistics of descriptors and visual words. In this study, LSA and second-order statistics are encoded into VLAD-like form, namely, LSA2-VLAD. Based on the experimental results obtained in this study, in terms of average accuracy, the performance of the proposed approach combining LSA-VLAD and LSA2-VLAD is better than those of 10 comparison approaches.