Research On The Abnormal Action Detection In Key Facility Areas Based On Video Sequences

With the country’s economic and social development,a large number of key facilities have been built to meet the needs of people’s lives and to promote national industrialization.Due to the complexity and diversity of the scenarios in the key facility areas,new challenges are posed for security prevention and control.In order to effectively deal with all kinds of accidents caused by abnormal action of people,the abnormal action detection algorithm has become the primary means to solve this problem.It is able to make use of the intelligent features of computers and greatly improves the efficiency and accuracy of detection.This dissertation carries out research on video sequence-based anomalous action detection methods in key facility areas,mainly including key frame sequence extraction using video sequence features and temporal sequence action recognition using key frame sequences.1)Video frame feature extraction and result sequence optimization in the key frame extraction method:In order to solve the problems of inadequate feature extraction and lack of generalization caused by a single means of feature fusion in the existing key frame extraction method based on video frame features,which cannot cope with complex environments,this paper proposes an adaptive fusion key frame extraction algorithm based on video frame features,with the aim of selecting a suitable feature fusion method based on the dynamic and static feature distribution of video sequences.At the same time,in order to effectively reduce the missing and redundant key frame sequences,this dissertation proposes a representative and difference loss function based on the high representativeness and high conciseness of key frames,which helps the algorithm to improve the completeness of key sequence representation of abnormal action features and the conciseness of key frame sequences;experiments show that the feature fusion method in the key frame extraction algorithm proposed in this paper can effectively The experiments show that the feature fusion method in the key frame extraction algorithm proposed in this paper can effectively improve the detection accuracy in complex scenes,while the proposed optimisation function can further improve the results and enhance the completeness and correctness of the detection.2)Anomalous action feature extraction techniques in temporal anomalous action recognition methods:To address the problems of inefficient model inference and loss of positive samples in the output decoding stage of existing methods resulting in reduced accuracy.In this dissertation,we propose a backbone feature extraction network based on reparameterization,using sparse matrix parameter fusion to reduce the complexity of the residual network structure and improve the inference efficiency.At the same time,the complete spatio-temporal feature vector decoding using the action instance distribution method and the optimization of action instance features by means of a loss function based on action continuity constraints help the model to improve its learning ability;experiments show that the backbone feature extraction network method proposed in this dissertation can effectively improve the efficiency of abnormal action detection,while the decoding method based on the action instance vector effectively increases the number of learnable data and improves the detection accuracy.