In the last few decades, understanding crowd behavior in videos has attracted researchers from various domains. Understanding human crowd motion can help to develop monitoring and management strategies to avoid anomalies such as stampedes or accidents. Human crowd movements can be classified as structured or unstructured. In this work, we have proposed a method using deep learning technique to characterize crowd behavior in terms of order parameter. The proposed method computes features comprised of motion histogram, entropy, and order parameter of the frames of a given crowd video. The features are fed to a Long Short Term Memory (LSTM) model for characterization. We have tested the proposed method on a dataset comprising of structured and unstructured crowd videos collected from publicly available datasets and our recorded video datasets. Accuracy as high as 91% has been recorded and the method has been compared with some of the recent machine learning algorithms. The proposed method can be used for real-time applications focusing on crowd monitoring and management.