Vigilis: A Time-Distributed CNN-ConvLSTM Model for Real-Time Shoplifting Detection in Uncontrolled Retail Environments

Shoplifting remains a critical challenge for retail stores, resulting in substantial annual financial losses. Traditional approaches such as Electronic Article Surveillance (EAS) and Closed Circuit Television (CCTV) monitoring by personnel are ineffective due to tag removal and human observation errors. Several researchers have aimed to address this issue through deep learning-based solutions for video surveillance. However, most have evaluated their approaches against traditional datasets collected under controlled environments. In addition, existing works have ignored the unfavourable effects of background objects like shelves and items on them on the model performance. To this end, the current work introduces “Vigilis,” a time-distributed convolutional neural network (CNN) and convolutional long short-term memory (ConvLSTM) model designed for real-time shoplifting detection in uncontrolled retail environments. The proposed approach employs a two-step methodology, including a video preprocessing technique based on pixel differencing to eliminate static background elements and retain relevant motion patterns and a time-distributed memory network that extracts and learns spatiotemporal features for accurate detection of shoplifting activities. The proposed model architecture integrates four primary components- feature extractor, memory network, learner and classifier, to enable robust detection. Extensive evaluation of datasets from four retail stores operating in uncontrolled environments yielded F1 scores of 0.77, 0.77, 0.79 and 0.77, demonstrating the system's effectiveness in real-world scenarios. This work establishes a scalable framework for addressing shoplifting detection challenges and provides a foundation for in