Abnormal Detection In Crowd Scenes Based On The Histograms Of Oriented Optical Flow And Sparse Representation

In the field of public safety, abnormal behaviors detection in crowd scenes has caught much attention from the researchers of many fileds, such as physics, sociology, graphics, computer vision, etc.. The purpose of the study is to understand the behavior of group in video surveillance. People from different research fields have different interpretations to the same video scenes and, in the field of computer vision, group behavior analysis mainly refer to classifing group behaviors[1].Traditional behavior analysis was generally modeled for a single person, in which It is often assumed that the movement patterns of moving objects relatively fixed and the background and illumination don’t change. It’s ideal condition,. However, in crowded public areas, there are many persons and occluders among them[2], especially in public places (such as airports, subway stations, etc.). Therefore, the traditional methods often couldn’t meet the our needs when applied in such scenarios. In order to solve the problem above, in this paper we study the methods for the analysis and detection of the various forms of abnormal group behaviors in the crowded scenes, and classify the abnormal events and other emergencies such as groups gathering, scattering, running and fighting suddenly.Most of the movement model used in the past study of behavior analysis did not take into account the motion characteristics of temporal continuity between frames, and the extracted features did not contain such kind of motion information. In this paper we present a new feature descriptor, called the kinematic histograms of oriented optical flow (KHOF), in which the acceleration information is introduced as an important factor in the conventional optical flow histogram. Traditional optical flow histogram only statistics the intensity of optical flow in multiple directions, whlie KHOF proposed in this paper can show the change of optical flow in different directions, instead of only the distribution of the intensity of optical. It makes the motion characteristics richer. In this paper we also introduce a method, called significant determination of temporal visual area, which analyze the characteristics of optical flow of each frame for the tracking of the valid movement information to extract only efficient moving area as a sample. By this way we can reduce the computational expense effectively. Sparse representation theory is also applied to the detection of abnormal behavior of groups in this paper. The sparse reconstruction cost (SRC) is used to distinguish abnonnal behaviors[3]. Compared to the general classification methods used in the past, sparse representation characterized mainly in terms of speed, high recognition rate and stability. The experiments on the standard datasets, UMN and UCSD datasets, and the video made by ourself show the results: our method can effectively identify various types of anomalies, and i is more effective than many existing algorithms.The main work of this paper is divided into the following two parts. (1) The kinematic histograms of oriented optical flow(KHOF)The optical flow method is a very representative motion feature extraction method used widely. Assuming the gray of pixels of two consecutive frames of a video sequence are not changed, the state of motion of the object can be indicated by the calculated instantaneous velocity of the moving object in the video field. Wang et al[4] proposed that the statistical information of direction and speed of movement can be depicted by histogram of oriented optical flow(HOF) and Cong et al.[3] proposed a multi-scale histogram of oriented optical flow(MHOF), which can statistic optical flow on different scales. However, the acceleration, a very important feature in motion information are not included in their characteristics. In our study, abnormal behavior is often manifested in fast moving objects (such as running, the crowd quickly passing vehicles, etc.) and sudden acceleration or deceleration of moving objects (such as panic crowd scattered, assault, etc.). In this paper, we propose a new motion feature descriptor, the kinematic histograms of oriented optical flow (KHOF). In the traditional histogram of oriented optical flow, the optical flow of the image patch, is calculated first then statistic distribution of the optical flow direction and intensity is determined, while in KHOF, not only the distribution of optical flow in different direction, but the change of optical flow in different direction is taken into the histogram.(2) Group anomaly detection based on sparse representationSparse representation theory has been widely used in recent years. Sparse representation is sparse decomposition based on an over complete dictionary. The dictionary is constructed with redundant atoms, instead of traditional orthogonal basis. In some areas, sparse representation has been used very successfully, such as reconstruction of image, face recognition and action recognition etc.. In this paper, the sparse representation theory is applied to abnormal behavior detection of groups. Experimental results show that, compared to the traditional classifying algorithm, the algorithm of sparse representation have higher speed and recognition rate,and it is more stable in performance. Sparse representation is very suitable for small training samples and feature classification of high dimensionality. We introduce sparse representation approach to identify abnormal behavior, reconstruct the over complete dictionary with normal behaviours, and detect the abnormal behaviours with sparse reconstruction costs (SRC).Experiments show that the method proposed in this paper can effectively identify a variety of unusual events, such as running, vehicles passing among the crowd quickly, panic scattered, assault, etc.. In this paper, we also use the temporal region saliency extraction, making the number of redundant samples significantly reduced without affecting the detection result, which greatly improve the detection efficiency. Meanwhile, with KHOF proposed in this paper, some special anomalies can be detected by our algorithm for which other methods can’t detect such as sudden acceleration of deceleration of moved object. As our feature descriptor comprise richer motion information, the method proposed in this paper can achieve higher detection rate and lower error rate in abnormal behaviour detection.