Time analysis-based human activity recognition and localization

Abstract

Current human activity recognition methods face many problems, such as the need for multiple sensors, poor implementation, unreliable real-time performance, and lack of temporal localization. In this study, we develop a human activity recognition and localization method based on temporal behavior recognition. In this work, we use a multilayer convolutional neural network (CNN) to extract features. It also uses fine-grained motion grouping to generate precise area suggestions. We then classify candidate regions using structured segmentation network-based activity classifiers and an end-to-end trained cascading design. Compared with previous behavior classification methods, this method increases the temporal boundary and effectively improves detection accuracy.