Few Shot Spatio-Temporal Anomaly Detection Model For Suspicious Activities

Convolutional Neural Network (CNN) has performed better for recent application of object recognition and object detection especially for image data but problem with CNN is that they require labels as learning signals. It is quite impossible to label all types of anomalies in a particular environment. Unsupervised methods used for video anomaly detection has drawback that they require too much data so that accurate results should be produced using unlabeled data which in turn increases computational cost. For this research a Few shot anomaly detection method is introduced using spatio-temporal autoencoder model for detecting suspicious activities in videos is proposed which doesn’t require any labels during training and also has very less computational cost then traditional unsupervised deep learning methods. Spatiotemporal autoencoder model has two components. Spatial autoencoder is used for spatial feature representation while temporal autoencoder extracts features from temporal dimensions. Few shot anomaly detection technique comprises the fact that it takes few images in each batch of training loop and trains the model on those images. At last averages the learning of all images and compute the loss for reconstruction by taking average loss of all batches. Experimental results on Avenue Dataset gives better results and achieves much lesser computational cost then other unsupervised anomaly detection methods.