JoWUA

Abstract

Violent incidents in public and private spaces are a growing concern, necessitating more efficient and reliable surveillance systems. Conventional systems, which rely on manual monitoring, are labour-intensive and prone to error, underscoring the need for automated solutions. The proposed research is to use a low-complexity convolutional neural network CNN model for real-time violence detection in videos, which can be used in surveillance settings. Addressing the limitations of existing high-computation methods, which are often unsuitable for real-time detection in resource-constrained environments, a lightweight CNN model incorporating SeparableConv2D layers has been proposed. This architecture reduces computational complexity by decomposing convolutional operations into depthwise and pointwise convolutions, ensuring efficient feature extraction while maintaining accuracy. The model combines a CNN backbone, MaxPooling layers for down-sampling the dimensions, and dense layer with output layer with sigmoid activation using the Keras-TensorFlow framework. Benchmarking against traditional methods like XGBoost, the model achieved a significant accuracy improvement, with computational efficiency suited for deployment on edge devices. Experimental results show the proposed model achieving 95% accuracy in distinguishing violent from non-violent actions, outperforming conventional methods and proving effective for real-time surveillance applications. This highlights the potential of low-complexity CNN architectures in enhancing public safety through timely intervention in high-risk environments, offering an efficient, accurate solution for violence detection.