Understanding the behavior of the crowd in densely populated regions is vital in enhancing public safety and security. This paper combines CNN with LSTM networks and an attention mechanism in proposing a deep learning-based technique for understanding behaviors and competent in the identification of anomalies in crowded places. In order to learn the normal and abnormal behaviors across time, the CNN classifies stills while the LSTM stores the sequences. As attention is directed to the most salient features of the video spans, the attention mechanism also enhances the model further. This method has been tested using publicly available datasets and has demonstrated successful detection of abnormal behaviors for instance panic attacks and hostile actions. The proposed method aims at increasing the reliability of intelligent surveillance systems.