Abnormal Behavior Detection in Surveillance Video via Multi-Input Feature Clustering with GAN-Augmented Autoencoders

Abstract

With the booming development of the global tourism industry, the increase in tourists has gradually made the safety management of tourist attractions more important. Monitoring abnormal behavior in tourist attractions is crucial in the safety management. To improve the accuracy of monitoring abnormal behavior in tourist attractions, this study combines convolutional neural networks with autoencoder network structures to reduce the learning generalization ability of convolutional neural networks. Attention mechanism is incorporated to improve sensitivity and recognition accuracy of abnormal behavior in complex environments. The method was experimentally validated using the CUHK Avenue and UCSD datasets, and compared with existing baseline methods. The results showed that the mixed multi-input feature clustering algorithm based on deep convolutional autoencoder had better detection performance than traditional methods on these two datasets. On the CUHK Avenue dataset, the AUC value was 91.9%, which was 27.1%, 10.6%, 15.0%, and 2.8% higher than that of the Adam, MDT, SF, and SRC methods, respectively. On the UCSD dataset, the AUC value reached 94.7%, which was 31.0% higher than that of the other four methods. In addition, the precision on the CUHK Avenue dataset was 94.5%, the recall rate was 95.6%, and the error rate was 12.6%. On the UCSD dataset, the precision was 95.2%, the recall rate was 94.8%, and the error rate was 10.9%. Overall, the research on the detection method of abnormal behavior in tourist attraction monitoring videos based on mixed multi-input feature clustering algorithm has high detection accuracy and can provide more effective technical support for the safety management of tourist attractions.