Anomaly Detection Based On Deep Neural Network In Surveillance Videos

The monitoring system plays an important role in preventing dangerous activities and maintaining the public order.As the scale of surveillance in urban streets continues to increase,relying solely on manual management is inefficient and there are potential safety hazards such as risk omissions.The development of automated and intelligent surveillance video analysis systems is imminent.Based on the street video surveillance,the thesis studies the anomalies in the scene,and proposes an anomaly detection algorithm.The algorithm divides the anomalies into two types:abandoned objects and pedestrian abnormal behavior.For abandoned objects detection,the thesis mainly studies foreground extraction and abandoned detection algorithms in complex situations such as environmental interference and scene changes.For abnormal behavior detection,the thesis mainly studies how to determine the location of video frames which contain the abnormal behavior in the surveillance videos.In addition,the thesis studies the classification of abandoned objects and abnormal behavior.The main work includes the following three aspects:(1)The thesis proposes an abandoned objects detection algorithm.The algorithm is divided into three modules: foreground extraction,abandoned judgment and misdetection removal.For the foreground extraction module,the algorithm uses fully convolutional siamese neural network,and the foreground can be accurately extracted only by two frames.Firstly,a contrast loss function is proposed,which can generate an accurate similarity metric map between pair of input images,and then the encoder-decoder network is used to perform end-to-end foreground extraction.For the abandoned judgment module,the algorithm is based on pixel time accumulation,and the abandoned objects detection can be accurately performed according to the foreground extracted by the foreground extraction module.For the misdetection removal model,the algorithm use siamese neural network to compare the similarity between the foreground and the background block at the location of the abandoned object.The situation where the foreground and the background are similar is defined as a false detection and the situation where the foreground and the background are not similar is defined as a correct alarm.The experiments show that the proposed foreground extraction module performs well in CDNet2014 dataset and gets 99% accuracy.The module can extract the foreground in various complex environments.The proposed abandoned objects detection algorithm gets 99%accuracy in the labeled PETS2006 dataset.(2)The thesis proposes an abnormal behavior detection algorithm.The algorithm is divided into two steps.First,the attention map is calculated according to the input video sequence.The attention map contains the movement information of the foreground area,so that the neural network can pay more attention to characteristics of abnormal behavior.Then,abnormal behavior detection is performed through a neural network based on Conv LSTM and reconstruction error.The network uses2 DCNN for spatial semantic information learning and uses Conv LSTM for temporal feature information learning.The algorithm uses reconstruction error and attention map to make the reconstructed video as similar to the input video as possible.It judges abnormal behavior through the relationship between reconstruction error and threshold.Experiments show that the proposed algorithm exhibits good performance in both Avenue and UCSD Ped datasets and gets 87.70%accuracy in the datasets.(3)The thesis proposes classification algorithms for abandoned objects and abnormal behavior.For abandoned objects,a few-shot classification algorithm based on metric learning is used.The algorithm uses a feature extraction network to extract features from samples to be classified and samples of known types.It inputs the extracted high-dimensional features into a shallow neural network for feature encoding,and then performs two similarity measures to determine the specific types of samples to be classified.For abnormal behavior classification,a skeleton based algorithm for abnormal behavior classification is used.The algorithm obtains human skeleton information at different times through Alphapose.The algorithm uses LSTM to fuse the time series features of the processed information and then classifies behavior and alarms abnormal behavior.Experiments show that the abandoned objects classification algorithm based on metric learning can classify the abandoned objects in surveillance videos even with zero samples.The abnormal behavior classification algorithm based on human skeleton reaches 90.61% accuracy in the dataset.