Deep Learning-Based Video Anomaly Detection Using Optimised Attention-Enhanced Autoencoders

Abstract

Anomaly detection in video is essential for applications like surveillance, healthcare, and industrial monitoring. Through the reconstruction of normal patterns and the computation of reconstruction error in relation to ground truth, convolutional autoencoders detect anomalies. Frames with errors above a threshold are flagged as abnormal. Existing approaches rely on fixed thresholds, which may not adapt well to varying lighting conditions, leading to false positives or missed anomalies. A novel autoencoder (SESAA) is proposed in this work that combines self-attention with squeeze-and-excitation (SE) blocks and improves video anomaly detection by using a thresholding technique for optimal threshold identification. Our adaptive thresholding technique leverages reconstruction cost, peak signal-to-noise ratio (PSNR) and frame brightness for optimal threshold identification, enhancing adaptability to different scenarios. Comparing with dynamic threshold methods, we assess our model using ROC and AUC metrics. Experiments on three benchmark datasets validate the efficacy of our method in precise anomaly detection through optimal thresholding.