Deep Learning-Based Violence Detection: A YOLO V7 Approach for Real-World Security Applications
Pabitra Mohan Sethi, Hitesh Mohapatra, Asish Kumar Dalai, Pravinkumar Bhujangrao Landge, Soumya Ranjan Mishra School · 2025
The field of safety for everyone and monitoring has made the use of cutting-edge technologies to identify and stop violent acts more and more necessary. With the use of a deep learning algorithm and the YOLO (You Only Look Once) 7th version architecture, this research proposes a novel method for improving the detection of human violence. The system under consideration utilizes video footage captured by security cameras as input, utilizing YOLO V7 to effectively and precisely detect instances of fighting. The real-time processing capabilities of the model facilitate prompt detection and response, hence augmenting the overall efficacy of security systems. Our approach is to train YOLO V7 using a heterogeneous dataset that includes several violent contact scenarios. As a result, the model performs reliably in a variety of surveillance scenarios and is able to adapt well to varying environmental conditions. The paper describes the details of the training procedure, including the methods used for data augmentation to enhance the generalization of the model. To further minimize false positives, the incorporation of behavioral analysis techniques improves the system's capacity to distinguish fleeting hostile behaviors from typical activity. Numerous tests using real-world surveillance film are used to assess the performance of the proposed system, confirming its effectiveness in accurately identifying and categorizing violent episodes. Our solution is superior to existing violence detection techniques in terms of both accuracy and recollection, as demonstrated by comparative analyses. The article also addresses the practical ramifications of implementing a system like this, including scalability for large-scale monitoring networks and possible incorporation with current security infrastructure.