In recent days violence is rampant, and the violence that occurs in public places is considered trembling violence in this study. To avert these anomalous activities, a real-time violence identification model is required to monitoring the behaviour of people and to initiate appropriate measures when the abnormal activities happened. In this study a real time violence detection method is introduced, this evaluates the large amount of streaming information and recognises violence with a human intelligence simulation. The big data is a massive volume of real-time video streams from various sources as input that are analysed using the Apache Spark technique. In this method the frames are divided, and the characteristics of each frame are retrieved by the histogram-of-oriented gradient (HOG) algorithm. Then each frame is classified as violent and non-violent classes depending on their features. This classification is performed by double dueling deep Q-learning network ( in terms of deep reinforcement learning (DRL). The result of the proposed framework attains best and precise fallouts compared with the existing deep neural networks. This model yields the classification accuracy rate of 95.08%for real-time violence detection.