Transfer learning with inception V4 for behavioral anomaly detection in CCTV surveillance

Heightened security concerns and better affordability have driven the proliferation of video surveillance systems. Developing advanced CCTV cameras capable of real-time analysis of behavioral patterns to find abnormal patterns is highly anticipated and will help avert hazardous incidents. Deep Learning approaches, specifically Convolutional Neural Network (CNN) architecture-based approaches, have been extensively used for anomaly detection tasks in video analytics. The study results from research applying diversified Inception V4 transfer learning methodologies to accurately and efficiently identify abnormal activities in video surveillance systems. This study utilized the Inception V4 framework to classify video keyframes that are representative of normal or abnormal behavior. This paper will elaborate on applying transfer learning techniques of pre-training and fine-tuning, which will employ the required attributes from the input information to build a specialized predictor. The effectiveness of the presented models was evaluated through experimental studies that utilized UCF-Crime data for training and testing. Metrics, such as accuracy, recall, precision, and F1 scores, were employed as evaluation criteria to assess the performance of each model. The Fine-Tuned (F-T) model demonstrated performance metrics of 930%, 91.74%, 88.33%, and 90.01%, whereas the Pre-trained (P-T) model showed 88.70%, 88.93%, 87.12%, and 88.02%, respectively. These findings suggest that Transfer Learning (TL), employing the Inception V4 architecture, can effectively distinguish between normal and abnormal video behaviors. Moreover, adjusting the weights of particular layers through fine-tuning can improve the performance.