A hybrid model using 2D and 3D Convolutional Neural Networks for violence detection in a video dataset
Anusha Jayasimhan, P. Pabitha
Abstract
In the era of Internet of Things(IoT), the data from edge devices, wearables and other sensors have become available easily and can be efficiently used for knowledge extraction. One such form of streaming data is in the form of videos which maybe captured from CCTV devices installed in various locations. While deep learning algorithms such as Convolutional Neural Networks(CNN) are not only effective for image classification but their application can also be extended for classifying videos. Two dimensional(2D) CNN classify images by extracting the spatial features from an image. However, when they are applied to videos, the temporal features of a video may get suppressed leading to a poor accuracy. In this work, we implement a hybrid CNN model, by combining a three dimensional convolutional neural network(3D CNN) and 2D CNN for identifying if the video footage contains violence related activities. Results prove that the hybrid CNN model attains a maximum training accuracy upto 99.93% and validation accuracy of 84.5% for violence recognition from video footages.