Vision-based violence detection through deep learning

• Main Author: Koh, Wei Zhe
• Format: Final Year Project / Dissertation / Thesis
• Published: 2024
• Subjects: Q Science (General); QA75 Electronic computers. Computer science; T Technology (General)
• Online Access: http://eprints.utar.edu.my/6826/1/2004757_KOH_WEI_ZHE.pdf; http://eprints.utar.edu.my/6826/

In the present society, video surveillance systems have continued to develop and incorporate more sophisticated video analysis to enhance security and public safety. With increasing demand, the need for accurate and efficient violence detection in video footage has become more critical. However, detecting violence in video footage remains challenging due to varying lighting conditions and data quality. While advancements in deep learning techniques can improve the accuracy and robustness of violence detection, they often require extensive datasets, leading to overloaded training processes. This research focuses on advancing and utilizing deep learning models for violence detection in surveillance videos, with particular emphasis on varying lighting conditions. A dataset of 2,000 videos mostly in normal lighting conditions is used to train a hybrid deep learning model combining MobileNet-v2, a lightweight Convolutional Neural Network (CNN), with BiLSTM (Bidirectional Long Short-Term Memory). This hybrid model seeks to employ MobileNet-v2 for feature extraction and BiLSTM for temporal analysis in video datasets. To enhance detection accuracy under different lighting conditions, histogram equalization is integrated into the video prediction process alongside the trained base model. The approach is designed to optimize video-based violence detection without overwhelming the model with large datasets and excessive training times. The base model (MobileNet-v2 and BiLSTM) performed well in normal light conditions (96.33%). While the base model with histogram equalization achieved higher accuracy (98.91%) and the model trained on varying lighting conditions further improved to (99.15%). On the other hand, the base model performed poorly in very dark conditions (24.89%) but showed significant improvement with histogram equalization (92.21%), nearly matching the performance of the base model trained on varying lighting conditions (99.97%). This result highlights the benefit of the proposed histogram equalization method, which achieves high detection accuracy without relying on extensive datasets and overloaded training resources, making it a potential solution for real-time violence detection in diverse lighting scenarios.