Publication

Abstract : Ensuring women’s safety in public spaces, especially during nighttime, is one of the most challenging issues today because current surveillance systems are not accurate in low-light conditions and tend to generate high false-positive rates. This paper introduces a novel AI-powered intelligent surveillance system that incorporates state-of-the-art deep learning techniques to address the aforementioned issues. The proposed system works in three stages: (1) Real-time detection of people through YOLOv10, (2) Attribute and Gender recognition using the DeepMAR model enhanced by an Inception-ResNet-v2 backbone, and (3) Proximity analysis through Euclidean distance and Violence detection using YOLOv8 model. A key innovation of our approach is the integration of Zero-DCE for low-light image enhancement, which significantly improves performance in poorly lit environments. The system is set up to trigger real-time alerts to the local authorities using Firebase whenever there is a potential threat detected, hence minimizing false positives and intervention in a timely manner. The evaluation results show that Inception-ResNet-v2 does better compared to ResNet-50 in both genders and attribute classification with 92.8% for PETA and 95.2% for PA100K dataset. The YOLOv8 model also has attained better performance in violence detection, where it has mAP@0.5 at 92.7%. This multi-tiered, real-time surveillance framework enhances safety in poorly lit public spaces, addressing gender-based violence and ensuring women’s security