Researches On Abnormal Events Detection Of Video Surveillance

Abnormal event detection of video surveillance is a cross research which requires knowledge and skills on both video processing and data mining.The significance of this research lies in the potential and vast demand on video analysis.Modeling normal data and abnormal data are two possible solutions for anomaly detection.However,it is difficult to collect adequate abnormal data.Therefore,modeling normal data is the more commonly used method.The thesis summarizes the framework for patch based visual anomaly detection methods.The major work consists of the following four parts:(1)We summarizes the principles,features and models the state-of-art methods adopt.The commonly used features for visual anomaly detection includes HOF,3D-gradient,foreground counter and a descriptor of combined features.HOF is able to catch the speed and direction of motion;3D-gradient is able to combine texture information with motion;foreground counter is a map of motion density;and the descriptor contains information from multiple features.Anomaly detection is a process digging out objects with unexpected behaviors.In data mining,the methods can be classified into cluster-based,stochastic,classification based,etc.Anomaly detection heavily relies on the application.We extract features and models from the recent published works seeking for the best combination.We select the maximum model,one-class SVM and sparse model as candidate models for evaluation.(2)We implement an abnormal event detection system of video surveillance.The system uses HOF as the major part of the descriptor and implement the maximum model,one-class SVM and sparse model for anomaly detection.It is able to evaluate methods if the dataset contains ground-truth videos.All the training,detection and evaluation work can be done with one-click.(3)We compare our methods with six start-of-art methods on a benchmark dataset.The results suggests that the combination of HOF feature with the maximum model is effective and rank third among all the methods on the dataset.The running speed of our method reaches real-time(30.67 FPS)and exceeds all other methods except for the Sparse Combination method,which claims to reach a detection speed of 150 FPS.(4)Detection results on real surveillance videos suggests our method is effective in both day and night.Our method is able to detect abnormal behaviors of vehicles,motorcycles and pedestrians in complicated and crowded scenes.By analyzing the results,we also find the main causes for false alarm and missed target lies in occlusion,bad light conditions and small size of targets,bad light conditions,respectively.