Abstract:
This research presents a comprehensive framework for detecting and analyzing
anomalies in video surveillance systems through an integrated hybrid model called
Adaptive Hybrid Integration Detection (AHID). This model was developed to detect
anomalies using an Autoencoder mechanism to construct the AHID model. It was
evaluated on three benchmark datasets: UCSD Ped2, CUHK Avenue, and ShanghaiTech, achieving detection accuracies of 98.31%, 91.03%, and 94.02%, respectively, along with AUC-ROC scores of 96.7896, 94.2396, and 89.5196. For the Ped2 dataset. the model achieved a Precision of 91.83%, Recall of 94.50%, and F1-Score of 92.12%. For the Avenue dataset, it achieved a Precision of 88.52%, Recall of 90.56%, and F1-Score of 92.03%. In the more challenging ShanghaiTech dataset, it achieved a Precision of 88.25%, Recall of 86.43%, and F1-Score of 84.92%. These metrics demonstrate the robust and adaptive performance of the AHID model in detecting anomalies across diverse environments.