Human Activity Recognition (HAR) is an essential process for human safety and social well-being. This makes to develop a cost-effective and accurate HAR classification method for surveillance applications. The proposed work consists of three steps such as, pre-processing, feature extraction, and classification. First step, apply the pre-processing technique to reduce the dimensionality of the input data by using Temporal Singular Value Decomposition (TSVD) technique. In the second step, features are extracted from the preprocessed data with the help of spatio-temporal feature extraction technique. Finally, the extracted features are classified using the Spatio-Temporal Neural Network (STNN) model and an alert message will be sent to the police by a newly created application. The proposed technique is implemented by using the KTH video dataset and the results are compared with existing algorithms to find the capability of the model for real-time applications. The proposed method was found to have an efficiency of about 98%, which is far more efficient than existing deep learning-based classification methods.