| With the continuous advancement of urbanization,the rapidly increasing urban population is bringing severe challenges to urban management.Especially at the time when the people intensively travel,such as holidays,abnormal crowd events can lead to serious consequences.With the widespread application of deep learning,research on how to use artificial intelligence technology to help government workers effectively monitor population density has very practical significance.Accurately estimating the number of crowds from images or videos and judging the state of crowd movement will become an important application of computer vision technology in urban management,and it is more accurate to use drones to observe urban crowds from the air.Based on the needs of actual scenarios,this thesis proposes a crowd counting and anomaly detection method based on the UAV platform.In the field of crowd counting,with the deepening of the network structure and the addition of different functional modules,the structure of the convolutional neural network has become more and more complex,and the calculation speed will inevitably drop.This thesis studies the crowd counting method of real-time calculation on the UAV platform.By cutting branches,merging features in advance and using transposed convolution,this thesis designs a detection accuracy of 80%and can be used on the UAV platform every second.Calculating the network structure of 44 pictures,a good balance between the accuracy and real-time of the algorithm.In the field of crowd anomaly detection,the mainstream crowd anomaly detection method has a large amount of calculation,and the current counting method ignores the spatial distribution information of the density map.This article is based on the population density map for crowd abnormality detection.First,a data set with a total of 7090 pictures was designed,which included five types of crowd movement.Then use the clustering algorithm to obtain the change information of the crowd movement state of the continuous image.Finally,through the "number of clusters" and "sample size of the largest cluster" design indicators "comprehensive confusion value",the population anomaly detection is realized. |
