Abnormal Crowd Behavior Detection Algorithm Research Based On Spatial-Temporal Interesting Points

In the field of computer vision, detecting group abnormal behavior in crowded scenes intelligently is becoming a hot issue. This direction has acdamic and application value in intelligent transportation, security monitoring and human-computer interaction. In this paper, our interesting is detecting abnormal behavior in videos using STIPs (Spatial-Temporal Interesting Points) and analyzing escaping, fighting and other abnormal actions.Traditional methods based on tracking or particles have flaws on illumination changing and occlusion. While human actions in video exist fierce changing points in space and time and STIPs method already reaped success in capture these points, we proposed a new method that describing human behavior in crowded scenes based on STIPs. This method extracts feature points from video and describes them in local. By comparing three different STIPs extraction methods, we chose a scale-invariant feature extraction method which is based on Hessian matrix.When STIPs were extracted, descriptor was constructed for each STIP. To test the algorithm performance under different descriptors, histogram of gradient, histogram of optical flow orientation and spatial-temporal Haar feature were used in this step. Three types of descriptor were also described and tested in detail.Bag-of-words model was used in normal behavior modeling and testing phase. In order to overcome the shortcoming that the traditional K-means method can not accurately describe the characteristics of the data set distribution in generating keywords, Gaussian mixture model(GMM) based on Expectation Maximum(EM) estimation was introduced. It described the data probability distribution effectively. Then each video of normal action is divided into several clips and they were described in vectors using keywords. All vectors constructed normal behavior codebook. In testing phase, input video is also required to be divided into several clips. Then we extract STIPs on each clip and match them with keywords and form a probability vector. Following the test code vector of each input video clip was formed by accumulating all probability vectors. Through calculating the Euclidean distance between test code vector and vectors in normal behavior codebook, the abnormal behavior was detected when the distance exceed an appropriate threshold.At last, we test the algorithm in UMN and UCF dataset. ROC (Receiver Operating Characteristic Curve) and comparison between detecting result and ground truth under different descriptor and model number were given. In UMN dataset, spatial-temporal Haar feature obtained better result. In UCF dataset, histogram of optical flow orientation surpassed other descriptors. Experiments show that the proposed algorithm has outstanding performance in detect group abnormal behavior in crowded scenes.