Video anomaly detection remains a widely investigated area due to its critical importance for applications in video surveillance, security, and traffic monitoring. Deep generative models (DGMs) have made significant advancements in the field, often neglecting crucial interactive dynamics between regions of interest (foreground) and their surroundings (background). The occurrence of an anomalous event significantly alters both the foreground and background; existing methods primarily focus on extracting holistic features and eventually learn redundant features under varying conditions. To address this, we propose CTSNet, a two-stream complementary network that utilizes a dual-contrastive loss for learning disentangled spatiotemporal representations. Our approach processes foreground and background frames separately using a convolutional long short-term memory (ConvLSTM)-based autoencoder, trained exclusively on regular videos. The dual-contrastive loss, combined with the mean squared error (MSE), enforces distinct latent representations for the foreground and background, thereby minimizing redundancy. During inference, anomalies are detected by evaluating both the reconstruction error and the cosine dissimilarity between the foreground and background latent features. The area under the curve (AUC) and accuracy results are obtained on the benchmark Avenue and UCSD datasets to evaluate the performance. Our method achieves an AUC of 96.8% on UCSD Ped2 and 86.6% on Avenue, with accuracies of 95.8% and 81.6% on the respective datasets, demonstrating its effectiveness compared to existing works.