摘要：Vast studies on action analysis with spatio-temporal features and deep learning algorithms have proposedcountless improvements over the years. This paper presents adomain model with continuous and spatio-temporal featuresbased on object detection for the industrial pipeline scenario.In this way, we can conduct a comprehensive evaluation: actionpose accuracy in the spatial dimension and action efficiency inthe temporal dimension. The method is applicable to scenariosthat require semantic recognition of action in a short time,without any dedicated action capture devices. Firstly, the dis-cretized images are combined into spatio-temporal, structuredand continuous action sequences. Then we apply the modelto the sequences to get the spatial action information throughvideo streams with only two-dimensional information, and thencomplete the analysis of the action specification based on theseaction streams with temporal features. Furthermore, this paperperforms several ablation experiments on training strategiesand hyperparameters to improve accuracy. Experimental per-formances show that it achieves an average recognition accuracyof about 96.45%.
关键词：semantic analysis of action; action specifica- tion; spatio-temporal features; deep learning; object detection
