An Experimental Comparative Analysis of Human Abnormal Action Identification on “SAIAZ” Video Dataset Using SVM, ResNet50, and LSTM Model

In recent days, almost the entire spaces are extended underneath the closed-circuit television (CCTV) (the surveillance system) to protect the public from offenses and provide safety to world heritage monuments and public buildings from damage caused by unauthorized entry. The surveillance camera generates a substantial amount of video stream data. Identifying suspicious activities from this huge amount of video data by manual intervention is impractical. As a result, an automatic action identification system is required to monitor human activities and identify abnormal activities. This processed system establishes a comparative study among the SVM machine learning model, ResNet50 deep learning model, and LSTM recurrent neural network model in terms of human abnormal action identification. To improve the efficiency of the system and adapt to the complex and real-world environment, the system is trained with primary dataset student activities in academic zone (SAIAZ) dataset, classroom violence (CRV), YouTube (cc), and CCTV video dataset. The SAIAZ video dataset was constructed with the help of groups of six to seven student volunteers of our institution. The SAIAZ dataset entails eight different abnormal actions like hair pull, head butt, hand twist, head twist, slap, punch, push, kick, and normal actions like reading, standing, walking, and sitting. The CRV is a classroom indoor violent activity. The CCTV dataset was framed by seven different activities collected from anomaly videos, a publicly available violent video dataset. The system is trained and tested with CCTV, CRV, and SAIAZ datasets on the SVM model, ResNet50 model, and LSTM model, and it produced the higher classification accuracy on the LSTM model at 95.3%, 96%, and 97.3% respectively.