Video Surveillance Based Intelligent Abnormal Behavior Warning For Small And Medium Crowds

Crowd abnormal behavior detection as a key technical issue for intelligent surveillance is an important topic in the field of computer vision.Although many scholars have paid great attention to pedestrian anomaly detection and also achieved certain results,some problems still keep open with the aspect of main target extraction and pedestrian behavior analysis,due to the difficulty of extracting information that characterizes pedestrian behavior in complex environments.For this reason,this thesis aims at the crowd's main target extraction and pedestrian behavior abnormality recognition,by which some methodological achievements can be acquired based on feature extraction and convolutional neural network(CNN).The main works and achievements are summarized as follows:A.Aiming at the problem that the threshold radius of the traditional Vi Be algorithm has a significant influence on target extraction effect,an adaptive threshold radius scheme is designed,and later an improved Vi Be algorithm is proposed.For the difficulty of main target extraction in small or medium populations,a main target extraction algorithm is developed based on the feature vectors of gradient histogram and optical flow direction histogram,relying upon the well-known support vector machine model.Numerically Comparative experiments show that the algorithm can effectively extract the main targets in small and medium populations,with low computational complexity and high accuracy.B.Based on the improved Vi Be algorithm,a pedestrian fall recognition algorithm is developed to detect the falling state of crowd movement,depending on the characteristic measurements of aspect ratio,centroid and velocity for the bounding rectangle of the main target.Further,as related to the improved Vi Be algorithm,a pedestrian wandering recognition algorithm is designed to detect the state of pedestrian wandering,after two indices of displacement difference and motion direction are designed to characterize the behavior of the main target.With the help of public data sets,comparative experiments show that such two abnormal behavior recognition algorithms can significantly improve the accuracy of pedestrian abnormal behavior detection.C.After an in-depth analysis of the basic principle,basic structure and functional design of CNN,an improved convolutional neural network is proposed to solve crowd abnormal behavior detection,by means of the strategies of crowd anomaly detection,Dropout mechanism and dataenhancement.With the help of open source data sets,numerical experiments show that the improved convolutional neural network has certain advantages in the identification of crowd abnormal behavior.