Deep Learning-Based Forensic Detection of Suspicious Activities in CCTV Systems
TL;DR
A deep learning-based forensic framework for real-time detection of suspicious human activity in CCTV videos, trained without relying on any external sensors is proposed, and incorporates anonymization of personal identities and local edge-based processing to prevent raw data exposure.
Abstract
Advancements in smart surveillance and digital forensics have been significantly influenced by artificial intelligence (AI), machine learning (ML), and cloud computing, enabling real-time forensic analysis of CCTV footage to detect suspicious activities and reduce reliance on manual monitoring. While these technologies enhance public safety and policing efficiency, they also raise important concerns related to privacy, ethics, and regulation. This study proposes a deep learning-based forensic framework for real-time detection of suspicious human activity in CCTV videos, trained without relying on any external sensors. The model utilizes EfficientNet-B5 for spatial feature extraction and a Vision Transformer (ViT) that has been adapted to model both spatial and temporal dependencies between consecutive video frames. Specifically, the ViT processes frame sequences as token embeddings, enabling temporal context learning through self-attention. The EfficientNet-B5, a convolutional neural network (CNN), functions primarily as a high-level feature extractor, capturing geometric and spatial characteristics such as scale variations, aspect ratios, and motion-relevant visual patterns from individual frames. The Binary UCF-Crime dataset is used for training and validation, with class rebalancing achieved through data augmentation and weighted loss adjustment to mitigate class imbalance. Performance is evaluated using accuracy, precision, F1-score, and AUC metrics, each averaged across 5-fold cross-validation. The reported accuracy is obtained through this cross-validation protocol and benchmarked against baseline models, including standalone EfficientNet-B5 and traditional machine learning approaches. The ViT-enhanced model achieves an accuracy of 87.23%, outperforming the EfficientNet-B5 baseline by a significant margin. To address ethical concerns, the framework incorporates anonymization of personal identities and local edge-based processing to prevent raw data exposure.
