Latent semantic learning with time-series cross correlation analysis for video scene detection and classification

• Published in: Multimedia tools and applications 2016-10, Vol.75 (20), p.12919-12940
• Main Authors: Cheng, Shyi-Chyi, Su, Jui-Yuan, Hsiao, Kuei-Fang, Rashvand, Habib F.
• Format: Article
• Language: English
• Subjects: Accuracy; Algorithms; Classification; Computer Communication Networks; Computer Science; Computer vision; Correlation analysis; Data Structures and Information Theory; Dynamic programming; Dynamic structural analysis; Dynamics; Learning; Machine learning; Multimedia Information Systems; Recognition; Semantics; Spacetime; Special Purpose and Application-Based Systems; Surveillance; Texture
• ISSN: 1380-7501; 1573-7721
• DOI: 10.1007/s11042-015-2548-y

This paper presents a novel, latent semantic learning method based on the proposed time-series cross correlation analysis for extracting a discriminative dynamic scene model to address the recognition problems of video event recognition and 3D human body gesture. Typical dynamic texture analysis poses the problems of modeling, learning, recognizing and synthesizing the images of dynamic scenes based on the autoregressive moving average (ARMA) model. Instead of applying the ARMA approach to capture the temporal structure of video sequences, this algorithm uses the learned dynamic scene model to semantically transform video sequences into multiple scenes with a lower computational effort. Therefore, to generate a discriminative dynamic scene model with space-time information preserved is crucial for the success of the proposed latent semantic learning. To achieve the goal, the k-medoids clustering with appearance distance metrics first used to partition all frames of training video sequences, regardless of their scene types, to provide an initial key-frame codebook. To discover the temporal structure of the dynamic scene model, we develop a time-series cross correlation analysis (TSCCA) to the latent semantic learning, with an alternating dynamic programing (ADP) to embed the time relationship between the training images into the dynamic scene model. We also tackle the problem of dynamic programming, which is supposed to produce large temporal misalignment for periodic activities. Moreover, the discriminative power of the model is estimated by a deterministic projection-based learning algorithm. Finally, based on the learned dynamic scene model, this paper uses a support vector machine (SVM) with a two-channel string kernel for video scene classification. Two test datasets, one for video event classification and the other for 3D human body gesture recognition, are used to verify the effectiveness of the proposed approach. Experimental results demonstrate that the proposed algorithm obtains good performance in terms of classification accuracy.