A Framework For The Crowd Anomaly Detection Using Machine Learning Techniques

Abstract

Currently, the demand of surveillance system is rapidly increasing for the real time security system applications and this type of surveillance systems are majorly used at crowded areas such as commercial complexes, sport stadiums and shopping malls etc. In crowd behavior analysis difficulty occurs during motion analysis that can be obtained with the support of object tracking and trajectories. To provide robust object tracking system under a complicated scenario, the system model should able to track object of interest and apart from environmental conditions, the object itself should have undergone changes. For that, it requires kinematics adaptation technique of tracking model, which will provide tracking of objects. Several methodologies have been proposed for the analysis of crowd behavior, which has limitations for heavily crowded scenarios and when a new object enters a frame etc. In this study, an approach to detect the crowd abnormality behavior in a dense scenario, where the behavior of crowd scenario may be Normal or Abnormal is proposed. Our proposed model is based upon the optical flow approach which has a capabality to examine the unexpected alteration in motion at any type of crowd. Initially, the major motion area is extracted through the motion heat map and the Harris corner detector is used to extract the information of point-of-interest under extracted motion area. The optical flow is analyzed in order to fix the threshold value, where the optical flow provides the energy level of each frame. In addition, SVM classifier is used to classify normal and abnormal frames, which is very useful in video surveillance model. Though in real-time processing, the classifier should be more accurate and the latency should be less for detection, and hence the techniques such as SVM, deep learning techniques such as Artificial Neural Networks are time consuming so it is required to look for other faster and accurate approaches. Further work is carried out, where the video sequences are split into frames.