Intelligent dual stream CNN and echo state network for anomaly detection

Traditional video surveillance systems detect abnormal events via human involvement, which is exhausting and erroneous, while computer vision-based automated anomaly detection techniques replace human intervention for secure video surveillance applications. Automated anomaly detection in real-world scenarios is challenging due to diverse nature, complex, and infrequent occurrence of anomalous events. Therefore, in this paper, we propose an intelligent dual stream convolution neural network-based framework for accurate anomalous events detection in real-world surveillance scenarios. The proposed framework comprises two phases: in first phase, we develop a 2D CNN as an autoencoder, followed by a 3D visual features extraction machanism in the second phase. Autoencoder extracts spatial optimal features and forward them to echo state network to acquire a single spatial and temporal information-aware feature vector that is fused with 3D convolutional features for events patterns learning. The fused feature vector is used for anomalous events detection via a trained classifier. The proposed dual stream framework achieves significantly enhanced performance on challenging surveillance and non-surveillance anomaly and violence detection datasets.