Abnormal behavior recognition algorithm in small sample scenarios for a utility tunnel project based on dcgan

Complex scenarios and a lack of anomaly data in utility tunnels are key issues hindering the identification of the behavior of construction personnel in utility tunnels. To solve these problems, this paper proposes a multiscene abnormal behavior recognition algorithm based on deep convolutional generative adversarial networks (DCGANs). First, the data are expanded by background and action extraction from a small amount of existing data using DCGAN. Second, a convolutional neural network model based on video segmentation with spatiotemporal dual channels is constructed to extract various abnormal behavioral data features, and the accuracy of the algorithm is improved by various technical methods. Comparison with other mainstream methods indicated that the top-1 accuracy of this paper's method reached 91.75% on the human action recognition data set of real action videos introduced by the University of Central Florida with 101 classes (UCF101). The accuracy of subclass identification after segmentation was greater than than 80%, and the t-distributed stochastic neighbor embedding (t-SNE) visualization of the classification results was more significant. This paper combined various abnormal behaviors that may occur during the inspection process of personnel in a utility tunnel with the internal environmental characteristics of a gas chamber, and constructed a spatial–temporal dual-channel abnormal behavior recognition algorithm based on video segmentation. The algorithm proposed in this paper can effectively solve the problem of insufficient abnormal data in utility tunnel scenarios and ultimately can realize the effective identification of various abnormal behaviors of personnel in a wide range of possible engineering scenarios. The algorithms and findings presented in this paper have important theoretical and practical contributions to the literature on the identification of abnormal personnel behavior.