The security of people during public events is one of the significant concerns of authorities. The authorities have to monitor the entire crowd continuously, and they must be capable of preventing all the abnormal activities in the crowd. They are responsible for avoiding these kinds of situations. To prevent such abnormal behaviours, first, they need to detect the abnormal crowd behaviour from the high-density crowd under observation. Detecting abnormal crowd behaviour from the crowd video has been one of the most critical research areas in the intelligent video surveillance system field over the past decade. Numerous strategies for crowd abnormality detection with the assistance of computer vision algorithms and machine learning methods have been proposed in recent years. Many of those traditional approaches are using hand-crafted features like optical flow, HoG, SIFT, and SURF. Even though most of these methods were able to produce a considerably good performance, these methods will take a lot of computational time to extract features, and that enhances the whole computational time. Especially the high-level features like SIFT and SURF are computationally complex, and this will affect the real-time performance of the system. In this paper, we propose a novel deep learning strategy for abnormal crowd behaviour detection. Rather than utilising hand-crafted features, deep neural networks naturally learn feature representations of the crowd video and which will help the system to detect abnormal behaviours. The learned feature representations help the system to differentiate normal and abnormal crowd behaviours. This method uses convolution neural networks, which have been utilised as an integral tool for feature learning in computer vision algorithms to extract the features from the videos. Instead of using traditional single-stream convolution neural networks, we have used pre-trained two-stream convolution neural networks to detect the crowd abnormality, which can consider both spatial and temporal information in the video. Our method is tested with available standard datasets and compared with state-of-the-art methods.