<p>Video anomaly detection aims to detect incidental and sudden patterns of events in complex scenes, which plays an essential role for maintaining public security and averting potential risks. Unsupervised learning is particularly effective in solving the problem of insufficient annotation of video abnormal behaviour in low resource scenarios. However, due to the complex entangled factors contained in the appearance/motion features of videos, it remains a formidable task for unsupervised video anomaly detection to overcome the barrier of spatio-temporal coupling. Most related methods focus on constructing a compact normal manifold within a normality training dataset or tracking objects, which often neglect the distinct visual characteristics of various anomalies and lead to the static coupling bias. To unravel the spatio-temporal coupling mechanism, we present an unsupervised approach STVAD that focuses on efficient context modelling and elimination of the spatio-temporal factor gap in a novel encoder-decoder network. Additionally, we propose a new paradigm that integrates spatio-temporal dependencies using self-attention mechanisms to boost the discriminative capacity to identify human-related irregular events from surveillance video sequences. Extensive experiments validate the accuracy and computational effectiveness of our method on three challenging benchmarks: UCSD Pedestrian(Ped1 and Ped2), CUHK Avenue and ShanghaiTech datasets respectively. Notably, our method significantly exploits the spatio-temporal clues to estimate association and improves the detection accuracy in detecting anomaly behaviours.</p>