Innovative Video Anomaly Detection: TCN-AnoDetect With Self-Supervised Feature Learning

Abstract

Video anomaly detection is a critical task in surveillance, industrial quality control, and anomaly monitoring systems. Recognizing anomalous events or behaviors within video sequences is challenging due to the diverse and often vague nature of anomalies. A novel temporal convolutional network-based anomaly detection (TCN-AnoDetect) is proposed that leverages TCNs and self-supervised learning. In this, TCNs are employed to model the temporal context within video sequences effectively, capturing short and long-term dependencies. The algorithm integrates TCNs with pretrained models to encode rich spatiotemporal features. The core of TCN-AnoDetect lies in self-supervised feature learning, where a neural network is pretrained on unlabeled video data to capture high-level spatiotemporal features. The anomaly detection module combines reconstruction-based and temporal context–aware approaches, using reconstruction errors and temporal context deviations for anomaly scoring and classification. To enhance model robustness, TCN-AnoDetect incorporates domain adaptation technique to handle domain shifts and evolving anomalies. The proposed algorithm is evaluated on three different benchmark datasets and ShanghaiTech Campus, demonstrating its superior performance. The extensive experiments performed in terms of different evaluation measures show the efficiency of the TCN-AnoDetect algorithm. The TCN-AnoDetect, an innovative approach, thereby provides promising solutions in video anomaly detection and in various applications.