A Hybrid Approach for Detection of Anomalies in Surveillance Videos using ConvLSTM and Mask R-CNN

In recent days automatic detection of anomalies from surveillance scenes have become major research. Though various deep learning algorithms are being developed, it is still difficult for many models to detect unusual activities in surveillance scenes. Anomalies like carrying weapons, fights, accidents and abandoning objects are rapidly increasing. To detect these anomalies and to overcome the challenges in automatic surveillance systems a novel methodology is proposed which offers a computationally effective and efficient way for video anomaly detection. Initially anomaly recognition is performed evaluating visual fidelity abnormalities for each video frame using signal PSNR ratio. Next, a hybrid model using ConvLSTM and MaskRCNN is combined to detect and classify the anomalies. The system can now dynamically focus on important areas of the movie thanks to this integration. This capability is superior to that of conventional analysis techniques. The system with two structures is more susceptible to different irregularities. Furthermore, the system has picture signals installed, which enhances the data feature more accurate extraction. After a comprehensive evaluation of the benchmark dataset, the suggested method has been proven to work better than the state of art methods. It performs exceptionally well in demanding monitoring situations. For instance, we were able to obtain an area under the curve (AUC) value of 87.96% in the PETS dataset, which surpasses the majority of modern techniques.