Abstarct: Anomaly behavior detection in surveillance videos is one of the important and challenging problems in the field of computer vision and intelligent monitoring systems. The ambiguous nature of anomalies, the wide variety of environmental scenarios, changes in imaging conditions, and the lack of labeled abnormal data have led reconstruction-baxsed unsupervised learning methods to receive significant attention in this domain. This thesis focuses on presenting and improving single-scene video anomaly detection methods using deep learning architectures baxsed on reconstruction and spatio-temporal data analysis. In this research, four proposed methods are presented. In the first method, a weighting mechanism is introduced in the reconstruction stage, which in the fourth proposed method is enhanced by utilizing monocular depth estimation of the scene to compensate for the effect of object distance from the camera in the reconstruction error calculation. In the second method, 3D convolutional networks are employed to extract spatial and temporal features simultaneously. Subsequently, a compactness module is introduced as a complementary criterion alongside reconstruction, which compresses the distribution of normal feature representations and increases their distance from abnormal samples in the feature space. In the third method, a ConvLSTM network is used to model temporal dependencies between consecutive video frxames. The proposed methods are evaluated on the standard Ped2 and Avenue datasets. Experimental results show that applying compactness through depth-wise averaging outperforms spatial-wise averaging, achieving an AUC of 95.27% on the Ped2 dataset. Furthermore, evaluating this configuration on the Avenue dataset results in an AUC of 84.83%, which, considering the higher complexity of this dataset, demonstrates the generalization capability of the proposed approach. In addition, the use of the depth-baxsed weighting mechanism, along with 3DConv and ConvLSTM networks, improves the system’s discriminative power in handling rapid motions and sudden scene changes. The investigation of learning rate effects also reveals that optimization settings significantly influence the convergence of reconstruction and compactness losses and, consequently, the overall system performance. The obtained results indicate that combining spatio-temporal analysis, a depth-baxsed weighting mechanism, and a compactness criterion can provide an effective and generalizable frxamework for anomaly behavior detection in surveillance videos.