To address insufficient spatio-temporal dynamic modeling and inadequate scene semantic adaptability in video anomaly detection under complex scenarios, this study proposes a skeleton-guided spatio-temporal behavior-scene alignment network (ST-BSAN). Robust anomaly detection is achieved through the collaborative optimization of two core modules. The dynamic spatio-temporal attention module (DSTAM) integrates spatial self-attention and temporal self-attention, overcoming the limitations of traditional fixed graph structures and local temporal modeling, and adaptively capturing dynamic joint correlations as well as long-range temporal variations. The behavior-scene alignment module (BSAM) constructs a dynamic memory bank, measures behavior-scene semantic alignment via cosine similarity, and suppresses false detections across different scenes. In addition, a diffusion probabilistic model is introduced to generate diversified normal behavior hypotheses. The generation process is conditioned on features extracted by DSTAM, thereby addressing the limitation of unimodal prediction in covering the diversity of normal behaviors. Experiments on the HR-STC and UBnormal datasets show that ST-BSAN achieves frame-level AUCs of 79.9% and 70.1% respectively, outperforming baseline methods by 2.3% and 1.8%. Ablation studies further verify the synergistic effect of DSTAM and BSAM.