Real-Time Detection of Crime and Violence in Video Surveillance using Deep Learning
Ali Mansour Al-madani, Vivek H. Mahale, Ashok T. Gaikwad · Advances in intelligent systems research/Advances in Intelligent Systems Research · 2023
Since its widespread application, deep learning is a vital part of the machine learning community's toolbox.With so many crimes and wrongdoings going on without adequate oversight in public spaces, various ways have been developed to identify crime and violence in the camera footage.Automated violence detection has become more important in video surveillance research.However, they have several restrictions, and much of the time, it is based on a certain set of circumstances.This research presents a 3D convolutional neural network-based technique for detecting violence in videos.Accuracy is improved by employing machine and deep learning techniques in a suggested manner.Performance evaluations have shown that the suggested technique effectively identifies violence in video clips.' Experimental data show that the proposed strategy outperforms existing methods for identifying crimes and violence in films.The pre-training models, Inception-V3, InceptionResNetV2, ViolenceNet, and ViolenceNet-OF, were trained on the four datasets.The classification results on the validation data for each model are as follows: ViolenceNet OF 99.40%, InceptionResNetV2 89%, ViolenceNet pseudo 96, and InceptionV3 92%.The DenseNet model was chosen for our application system because it concatenates the feature maps in a simpler method than Inception or InceptionResNet, which are more complicated.It has a more durable design that requires fewer filters and settings to attain high efficiency than other models and achieves the highest accuracy.