Abnormal Event Detection Method In Railway Scenes Based On Sparse Combination Learning

With the continuous improvement of people’s living standards,the number of tourist is increasing,and the safety of public places is also becoming prominent at the same time.Especially after the high-speed rail has become the preferred mode of travel for people,it is of great significance to carry out all-round and real-time monitoring of railway scenes such as the waiting room in railway station and platforms.At present,video surveillance systems have been widely used in various fields of society,but there are many problems of traditional manual detection,which has been already unable to meet demand.Therefore,the purpose of this thesis is to study a real-time anomaly detection method with high accuracy.The thesis adopts the method based on sparse combination learning to detect abnormal event in railway scenes,which has higher detection accuracy and detection speed than previous detection models,such as models based on classification and clustering and so on.Although the speed of the sparse combination model based on manual features is fast,there is still room for its accuracy to improve.In order to further improve the accuracy of the anomaly detection based on sparse combination,combined with the deep learning which has the rapid development in the field of computer vision,this thesis proposes a method of anomaly detection based on deep spatiotemporal features.The method extracts the spatiotemporal features of the captured video through deep learning,which can extract the temporal and motion information in the video sequence more effectively.Firstly,a deep three-dimensional convolutional network model is designed,and the network is trained through a large behavior database with labels to obtain a network model with optimal parameters.Then,the features of anomaly database are extracted by using the pre-trained network model,and next,the extracted deep spatiotemporal features are used for sparse combination learning and detection.Experiments on the public anomaly databases AVENUE and SUBWAY show that the proposed method has higher detection accuracy.Finally,the method is applied to the database on actual railway scenes,and it is tested under various railway typical scenes such as platform,main track and throat zone,which can obtain better detection results.According to the experiments on the public anomaly databases and the railway database,the feature extraction based on deep learning is combined with the anomaly detection model based on sparse combination learning in this thesis,which gives full play to the advantages of both ways.The feature extraction algorithm takes the original data as input and adopts the process of unsupervised feature learning to solve the problem of difficult acquisition for labeled data.The whole process of feature extraction does not require any artificial feature extraction steps.The stratagem further improves the detection accuracy by using the more essential features of the data under the condition of guaranteeing high detection speed.