Research On Anomaly Detection Algorithm Based On Discontinuous Frames In A Fixed Scene

Traditional video surveillance systems often have high requirements for networks,power supplies and storage.In some special scenarios,such as in the wild,low-power wireless image sensors that use timed shooting or event-triggered shooting can be an effective alternative to scene monitoring.This paper refers to the data collected by this type of sensor as discontinuous frame sequences.Furthermore,traditional surveillance systems often rely on people watching the surveillance video uninterruptedly,so that problems can be found and dealt with in a timely manner.The anomaly detection algorithm designs a discriminator to detect whether the input data is abnormal data by collecting the normal data.This paper studies anomaly detection algorithms based on discontinuous frame sequences.Our work is as follows:1)An anomaly detection algorithm based on autoencoder is proposed.This algorithm uses an adversarial encoder to limit the representation of hidden layers and uses an adversarial framework to move the generated content of the decoder closer to normal samples,thereby suppressing the expressive ability of the autoencoder.When an abnormal sample is sent to the model,the algorithm can determine whether the sample is an abnormal sample by the representation of the abnormal sample in the hidden layer,the feature form of the discriminator,and the reconstruction error.2)Anomaly detection algorithm for fixed-point surveillance image based on light migration is proposed.Considering that the discontinuous frame data set is affected by light and shadow,the algorithm proposes a light migration model based on Cycle GAN.Using the model's encoder,the algorithm can extract content that is not affected by light from the input image.In the test session,the algorithm compares the difference between the image to be detected and the standard image group selected by K Medoids to determine whether the image is abnormal.The algorithm can adapt to slight changes in the scene on discontinuous frame sequences without retraining,and has good performance indicators.