Research On Crowd Abnormal Behavior Detection Method Based On Video

With the population growth and economic development of the society,the types of crowd activities are becoming more and more extensive,and the probability of large crowd in public places is increasing.Due to the complexity of real scenes and uncertainty of crowd behavior,the public safety problem of crowd is becoming more and more prominent.Therefore,in order to reduce the loss of people and property when unexpected situations occur,real-time monitoring of crowd activities in scenes becomes a hot topic nowadays.Current research on crowd activity is divided into two main directions: crowd density estimation and crowd behavior analysis.This paper conducts research on crowd density estimation and abnormal behavior detection,with the main research contents as follows.(1)To address the phenomenon that the accuracy of existing classification algorithms varies widely in different density scenes,an algorithm that fuses pixel statistical features and texture features is proposed to classify crowd density.The algorithm first extracts the foreground images containing moving targets in the video by Vibe algorithm,then extracts and fuses the area features and GLCM features from foreground images,and finally achieves crowd density classification by using the fused density features and SVM classification model.Experiments are carried out on public datasets,and the results verify that the algorithm can achieve better classification results in different density scenes.The comparison experiments with current mainstream algorithms prove the effectiveness in classification and real-time performance of the algorithm.(2)Since the velocity features are prone to misjudgment and delay when analyzing crowd motion status,this paper improves velocity features and adopts acceleration features to detect crowd anormal behavior.According to the idea of extracting velocity features by optical flow method,we use LK optical flow method to calculate three consecutive frames of images to obtain crowd acceleration features,and then input them into SVM model for training to obtain crowd abnormal behavior detection model.The experimental results show that the accuracy of crowd anomalous behavior detection based on acceleration feature algorithm has been improved substantially compared with that of velocity feature.(3)In order to learn spatial features and temporal features in videos,a 3D CNNbased crowd abnormal behavior detection algorithm is designed.On the one hand,the3 D convolutional kernel is used to automatically extract the temporal and spatial features in videos;on the other hand,the large-margin Soft Max loss function is used to improve the classification ability of the model for an unbalanced sample size.Finally,the algorithm model is trained and tested on public datasets,and the experimental results show that the improved C3 D network model has high accuracy,robustness in crowd anormal behavior detection.Although the two anomaly detection algorithms studied in this paper can both achieve the detection of crowd anormal behavior,they are applicable to different scenarios.The acceleration feature-based detection algorithm is usually used in scenarios with fixed types of anormal behavior;the 3D CNN-based detection algorithm can detect more types of crowd anormal behavior.