Semantic Sequence Analysis for Human Activity Prediction.

The task of long video analysis is challenging, and it is often the case that many human actions occur but only a few contribute to the semantic topic of the video. However, compared with short video human activity studies, long video analysis has its practical utility especially considering the effort of watching a long video for human. In this paper, we propose to learn semantic symbol sequence patterns of complex videos for activity prediction. The prefix method of semantic stream is designed based on the semantic symbol sequence and their time marks. The prediction phase is implemented via matching semantic sequence of incomplete videos and sequence patterns of different activities. We evaluate various prediction methods depending on low-level features or high-level descriptions. The empirical result suggests that when applied to activity prediction, sequence pattern mining can effectively reduce its reliance upon the low level features and improve predicting performance.