Research On Key Techniques Of The Crowd Abnormal Behavior Detection In Surveillance Videos

Crowd abnormal behavior detection based on surveillance videos is a key issue in the smart surveillance system.Also,it is a hot issue in the area of computer vision.For the uncrowded scene,anomaly detection is easy because the image foreground is sample.In the crowded scene,because many moving targets and events will happen simultaneously and the image background is disorder,anomaly detection becomes very difficult.In this dissertation,aiming at the global and local anomaly detection problems under the crowded scenes,the feature representation of crowd behavior and detection model for abnormal behavior are studied to improve the robustness and accuracy of anomaly detection.The main contributions are as follows:(1)Histogram of maximal optical flow projection(HMOFP)is proposed.As a low-level visual feature,optical flow has obvious advantages in describing crowd basic behavior.To eliminate the interference of background noise for the optical flow and show the difference between the features of normal and abnormal movements,the maximal optical flow projection vector on the angular bisector in a bin is set as the motion vector.This will remove the effects of the small optical flow vectors caused by the small movements or noise.Combined with the one-class support vector machine,the advantages and effectiveness of our proposed HMOFP are verified.(2)A dictionary construction method based on the training sample set optimization is proposed.In order to improve the robustness of sparse reconstruction dictionary and reduce the computational cost of dictionary training,the training set optimization and dictionary learning are combined.Before dictionary learning,the training samples that never used for the representation of the other samples in the feature pool are deleted,such that the feature pool can be more compact and typical for the normal samples.Then considering both training speed and detection accuracy,the improved online dictionary learning method is utilized to obtain the dictionary for sparse reconstruction,which can improve the sparse representation ability of the dictionary for crowd normal behavior.(3)An anomaly detection model based on low-rank structure constraint is proposed.In the process of dictionary construction,based on the low-rank property of the feature representation of crowd behavior in the training set,joint minimization of the-norm and nuclear norm is imposed to force the reconstruction coefficient vectors of the training samples to compactly surround a certain center.In the testing stage,according to the constrain of the reconstruction coefficient vectors,the reconstruction coefficient vectors of normal testing samples will also distribute around this center compactly,which can reduce the reconstruction error of normal samples and enlarge the reconstruction error of abnormal samples,such that the detection accuracy of crowd abnormal behavior can be improved.(4)A crowd anomaly detection framework based on deep learning is proposed.A prediction based deep neural network is proposed by combining the Convolutional Neural Network(CNN)and a variant of the Recurrent Neural Network (RNN)-Long Short-Term Memory(LSTM).Instead of the traditional handcraft feature,the output image feature of CNN and the predicted feature of the next frame obtained by LSTM are used to calculate the prediction error.Then the testing frame can be classified as normal or abnormal according to the value of prediction error.