| The development of information technology and the popularity of video cameracontribute to the explosion of video data. Therefore, intelligent analysis is required todeal with massive video data. Video-based human action recognition, a major problemin video understanding, has always been a popular topic in computer vision due to itspromising application intelligent surveillance, video retrieval and other aspects.However, video-based human action recognition confronts the problem of the extraction,learning and classification of high-dimensional features. Recent research indicates thatmany signals are sparse, including natural images. As for now, sparse representation hasbeen successfully applied in areas like image processing.This paper focus on the application of sparse representation in video-based humanaction recognition, main works include:1. We study and summarize the sate-of-art solutions to action recognition. Weconduct a research on the theoretical foundation of sparse representation and analyzemathematical model of the problems to be resolved in practice, we also give ansummarization and comparison of the related algorithms2. In this paper, a human action classification method based on sparserepresentation is proposed. First of all, we utilize bag-of-words to get globalrepresentation based on the local spatio-temporal features extracted from the video.Then we design and implement a classification algorithm with sparse representation andintroduce feature-sign search algorithm to get sparse coefficients. At last we testify thefeasibility of this method.3. We also propose a human action recognition system based on sparse codingframework to take further advantage of the characteristic of sparse representation. Weget a global representation of a video after max pooling the sparse coding on localspatio-temporal features. Online dictionary is used to train an over-complete dictionary.System has achieved satisfying performance on KTH and UCF dataset. |
PDF Full Text Request