Abstract
In this paper, a high-level semantic recognition model is used to parse the video content of human sports under engineering management, and the stream shape of the previous layer is embedded in the convolutional operation of the next layer, so that each layer of the convolutional neural network can effectively maintain the stream structure of the previous layer, thus obtaining a video image feature representation that can reflect the image nearest neighbor relationship and association features. The method is applied to image classification, and the experimental results show that the method can extract image features more effectively, thus improving the accuracy of feature classification. Since fine-grained actions usually share a very high similarity in phenotypes and motion patterns, with only minor differences in local regions, inspired by the human visual system, this paper proposes integrating visual attention mechanisms into the fine-grained action feature extraction process to extract features for cues. Taking the problem as the guide, we formulate the athlete's tacit knowledge management strategy and select the distinctive freestyle aerial skills national team as the object of empirical analysis, compose a more scientific and organization-specific tacit knowledge management program, exert influence on the members in the implementation, and revise to form a tacit knowledge management implementation program with certain promotion value. Group behavior can be identified by analyzing the behavior of individuals and the interaction information between individuals. Individual interactions in a group can be represented by individual representations, and the relationship between individual behaviors can be analyzed by modeling the relationship between individual representations. The performance improvement of the method on mismatched datasets is comparable between the long-short time network based on temporal information and the language recognition method with high-level semantic embedding vectors, with the two methods improving about 12.6% and 23.0%, respectively, compared with the method using the original model and with the i-vector baseline system based on the support vector machine classification method with radial basis functions, with performance improvements about 10.10% and 10.88%, respectively.