TRAFFIC AND SECURITY MANAGEMENT SYSTEM: REAL-TIME VIDEO ANOMALY DETECTION USING AI MODELS

Abstract

The traffic and security management systems need effective real-time monitoring to guarantee the safety of the people, especially in the high-risk and crowded areas. Manual monitoring of CCTV video streams is time-consuming, prone to errors, and ineffective, which necessitates intelligent automated solutions. This paper introduces a real-time video anomaly detection system to monitor traffic and security with the help of LSTM-based deep learning. The suggested method uses a hybrid spatio-temporal feature extraction system that involves Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks to extract spatial features and temporal motion patterns in video sequences. The LSTM model is effective in learning sequential dependencies and thus detecting abnormal events accurately with time. The extracted features are then categorized with the help of various machine learning models, such as Histogram-Based Gradient Boosting (HGB), Gradient Boosting (GB), Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LGBM), Support Vector Machine (SVM), and K-Nearest Neighbor (KNN). The UCSD Anomaly Detection Dataset is tested on the framework with performance measures of accuracy, precision, recall, F1-score, Cohen Kappa, Matthews Correlation Coefficient (MCC), and loss. Experimental results show validation accuracies of 76% (SVM), 72% (KNN), 90% (HGB), 98% (XGBoost), 95% (GB), and 98% (LGBM). Among them, XGBoost has the highest performance with 98 percent accuracy, better classification rates, and low loss, which proves the efficiency and strength of the suggested system in the context of real-time traffic and security surveillance applications.

Keywords: Video anomaly detection, abnormality detection, machine learning, deep learning.