Anomaly Detection in Border Surveillance Using VAE-LSTM Network for Temporal-Spatial Analysis of Sensor Data
B. Haritha, S. P. Chokkalingam, C. Alamelu, M. Vallikkannu · 2025
Continuous monitoring of borders is essential for ensuring safety and peace of a country's people, even under harsh conditions. The main challenge lies in monitoring remote, treacherous, or vast terrains where human soldiers may not be effective in providing continuous surveillance and detecting potential security threats early. There is a growing need for an automated border surveillance framework. The proposed model uses a hybrid Variational Autoencoder and Long Short-Term Memory (VAE-LSTM), as an unsupervised approach for surveillance anomaly detection. The VAE captures latent representation of input data by encoding into Gaussian distribution which consist of mean and variance for model stabilization and enabling training on normal events for better reconstruction. The LSTM temporal relationship enables it to analyze ordered sequence across time frames by storing information to predict unseen events. Subsequently, the model is tested on image sequences containing both normal and abnormal activities. An anomaly is detected if reconstruction error exceeds a certain threshold. Frames with anomaly score above threshold are labeled as anomalous. The model is developed using the UNIRI-TID dataset includes thermal imagery and video data from border areas. The performance evaluation results indicate promising outcomes.