This paper presents a novel framework by training a real dataset taken from a specific area, and this data is of the video type. We have used several clips from surveillance cameras that we want to train and extract natural and abnormal events. This stage starts by collecting real-time data from multiple surveillance cameras. The video data is then converted into frame images, which are processed by a deep learning network consisting of three modern lightweight CNN models to generate prediction values for each class, namely Inceptionv3 and VGG16 and Efficientnetv2 model. This involves converting each video clip into a set of frames. The video file is opened using the OpenCV library. The video is then read frame by frame. Each frame is then saved as an image file, which is then trained using pre-trained models Inceptionv3 and VGG16 and Efficientnetv2 model, which classify the images as natural or abnormal events.