Research And Implementation Of Crowd Event Detection Methods In Public Places

As our country has a large population,personal safety in public places cannot be ignored.Although surveillance in public places can monitor specific conditions in public places,the video duration is relatively high and the occurrence frequency of abnormal events is low,which requires a lot of time and manpower to investigate.The existing computer vision technology can be combined with surveillance video to realize the detection of abnormal group events in public places.Rapid treatment in a short time after the occurrence of dangerous events can greatly reduce the impact of abnormal events in the population.This studies the abnormal detection method of surveillance video in public places to detect abnormal events in time.First,an anomaly detection model--TUnet,which combines Transformer and UNet,is proposed.Previous models are built based on convolutional networks and lack global perception ability.By combining Transformer with convolutional neural networks,we can ensure the model of ability to predict details and improve the model of ability to understand global information.The AUC of the proposed method in the three public data sets(UCSD ped2,CUHK Avenue,and Shanghai Tech)is 96.4%,85.9%,and 73.0%,respectively,which is 1%,0.8%,and0.2% higher than that of the prediction-based anomaly detection benchmark model,respectively.Then,based on the previous method,we continue to improve the memory module and put forward the Mem TUnet model.In order to prevent the generation network from being too strong,the quality of the generated abnormal event image is too high.To solve this problem,a memory module is added in front of each decoder layer to help the model record the characteristics of normal events and weaken the fitting ability of abnormal events.And a more concise model structure,and better performance without the use of optical flow and discriminator.Verified on the CUHK Avenue dataset and Shanghai Tech dataset,the AUC of the proposed method is further improved by 0.6% and 0.8% compared with the previous method,and the AUC of the proposed method is the same as that of UCSD Ped2.This study shows how the abnormal detection method of group events in public places improves the ability of model detection.The proposed method can timely reduce the secondary hazards after the occurrence of abnormal group events.