Optimized ST-GCN Model for Suspicious Human Activity Recognition using Edge Computing

Introduction: In automated video surveillance applications, detecting abnormal human activity is incredibly difficult to classify them. In these systems, labeling of human activities relies on the visual aspects and motion patterns observed in videos. But, a significant portion of many traditional methods, as well as traditional neural models, either disregard or face challenges in leveraging temporal features for predicting Human Action Recognition (HAR). Objectives: The main objective is identifying different suspicious human activities from videos and achieving precise and effective HAR. Methods: The automatic detection of aberrant human activity in a surveillance system was resolved by the Deep Learning (DL) based edge computing in our proposed work. The videos are first turned into frames and secondly, the spatio temporal features are retrieved from the key frames using a DL model Spatiotemporal-Graph Convolutional Network (ST-GCN). Finally, to recognize anomalous activity from video, the collected features are loaded into an optimized gated recurrent unit (OGRU). Results: The experimentation is carried out on the two benchmark datasets and achieved better accuracies of 98.1% (Dataset 1) and 98% (Dataset 2) respectively. Conclusions: This work effectively recognizes all kinds suspicious human activities from vast set of video files.