Surveillance cameras are used in most of the places, they capture most of the real-world anomalies. By using the deep learning techniques for the analysis of videos captured, varies anomalies can be detected. In this project anomaly is learned by using normal and anomalous videos. Annotating the video frames for training is a time-consuming task, to avoid the annotation of frames, deep multiple instance ranking framework is used, this frame work uses weakly labelled videos for training, i.e. videos are used instead of frames for training. The anomalous and normal videos are taken as bags and videos are taken as instances in the multiple instance learning (MIL), the model for anomaly detection is automatically learned and it produces a high score for anomalous video segments. The Sparsity and temporal smoothness are used to rank the loss function to better localize anomaly during training. The UCF crime dataset is used for training the model, 128 hours long videos are present