Research On Crowd Abnormal Behavior Detection Algorithm Based On Surveillance Video

With the rapid development of society and the increasing number in population,the safety problem caused by crowd abnormal behavior has attracted much attention.How to supervise the crowd in public places has become an urgent social problem that needs to be solved.In the intelligent monitoring systems,crowd abnormal behavior detection has always been an essential part,and it is also a popular research direction in the field of computer vision and image processing.In recent years,although crowd abnormal behavior detection methods have achieved good development,due to the complex diversity of crowd scenes,crowd abnormal behavior detection under monitoring is still faced with many problems and challenges,such as crowd occlusion,complex background,and other interference factors,resulting in the accuracy and real-time performance of the crowd abnormal behavior detection algorithm cannot meet the practical needs.Therefore,on the basis of the theory of computer vision and deep learning,a crowd abnormal behavior detection algorithm based on improved SSD and a crowd abnormal behavior detection algorithm based on knowledge distillation are studied in this thesis.The main work of this thesis can be summarized as follows:Ⅰ.An algorithm of crowd abnormal behavior detection based on improved SSD is proposedAiming at the problems of high algorithm complexity and low detection accuracy caused by overlapping occlusion in crowd abnormal behavior detection tasks,an algorithm based on improved Single Shot Multi-box Detector(SSD)for crowd abnormal behavior detection is proposed in this thesis.Firstly,a lightweight network MobileNet V2 is used to replace the original feature extraction network VGG-16 to reduce the size of model parameters.Then,the deforming convolution module is used to construct the convolutional layer to enhance the receptive field,and the location information is integrated into the channel attention for feature enhancement.The remote dependence between spatial positions can be captured,and the occlusion part can be predicted by learning the context relation,to better deal with the overlapping occlusion problem.After experimental verification,the proposed algorithm detected 26.59 fps and.25.41 fps on the two scenes of the UCSD dataset with AUC results of 74.50%and 88.93%,respectively,which have improved the detection speed and accuracy in different degrees compared with other methods.II.An algorithm of crowd abnormal behavior detection based on knowledge distillation is proposedAiming at the problem that the convolutional neural network is difficult to be applied in real life due to a large number of parameters and calculation amount,a crowd abnormal behavior detection algorithm based on knowledge distillation is proposed in this thesis.By constructing a lightweight small model and using the supervision information of the larger model with better performance to training the small model,the small model has the same performance as the large model,but the number of parameters and calculations amount is greatly reduced.Thus,the model can be compressed and accelerated.Firstly,the channel attention is embedded into the improved MobileNet V2.The detection accuracy of the student network can be improved by embedding the channel attention,and the lightweight model is constructed as the student network.Then,the crowd abnormal behavior detection model based on the improved SSD proposed in(1)is taken as the teacher network.By distilling the last layer of the network,the mapping of the teacher network can be learned by the student network,so that the performance of the student network is close to the teacher network while reducing the number of parameters and calculation of the network model.Finally,it can be concluded from the experimental results that although the mAP of the distilled student network is slightly lower than that of the teacher network,the number of model parameters is reduced by 73.30%and the amount of calculation is reduced by 89.15%.It shows that the method proposed in this thesis can further reduce the number of parameters and the amount of calculation without the loss of recognition accuracy,so as to obtain a more compact and powerful high-performance network,which proves the effectiveness of the method proposed in this thesis.III.Completed the GUI design of crowd abnormal behavior detection systemAiming at the algorithm of abnormal crowd behavior detection proposed in this thesis,a convenient graphical user interface for abnormal crowd behavior is designed.The detection interface includes the login interface and two main functional interfaces.The Qt Designer tool in PyQt5 is used to design the graphical user interface of the system,and its functions are tested.The test results show that this interface has good feasibility and practicability.In summary,the problems of low detection accuracy due to overlapping occlusion in crowd abnormal behavior detection,the large number of parameters and calculations in convolutional neural networks are studied in this thesis,which make it difficult to apply in real life,and the crowd abnormal behavior detection algorithm based on improved SSD and the crowd abnormal behavior detection algorithm based on knowledge distillation are proposed.Then interface for abnormal behavior detection is designed,which basically meets the detection requirements.