Real time violent actions detection is very important for keeping people safe and stopping possible damage. This task is very challenging due to dynamics of violent activity of human. It is extremely difficult for a human operator to monitor the surveillance footage, which commonly led to error and failure to identify the incidence of odd activities. So, we suggested a deep framework using VGG16 and Bidirectional Long Short-Term Memory (Bi-LSTM) to extract spatial features and simulate temporal dynamics in a useful way. For spatial feature extraction, we use VGG16, pretrained convolutional neural networks. Then we use sequence modeling to find temporal relationships between frames. Testing shows that the Bi-LSTM model is better than the regular LSTM model. It gets 98.8% accuracy, whereas the LSTM model gets 97.6%. These findings show that bidirectional temporal modeling helps us understand how people move during violent events. The suggested technique might be used in intelligent surveillance systems, which would make monitoring in security-sensitive areas faster and easier.