| With the improvement of public safety awareness,crowd abnormal behavior detection attracts more and more attention,which makes the research of crowd abnormal behavior an academic hotspot in computer vision field.At present,most monitoring systems applied for public areas in China are based on visible light.However,crowd abnormal behavior detection based on visible light is greatly influenced by environment,meanwhile temporary large-scale venues presents higher challenges to fixed video monitoring systems.Considering the strong mobility of the quadrotor UAV and the differences between infrared images of the crowd and the environment,this paper studies the algorithm of infrared crowd abnormal behavior monitoring applied to the quadrotor plat form and its system implementation.Firstly,the dynamics model of the quadrotor UAV is established by Newton-Euler mechanics.According to the characteristics of infrared imaging technology and the push-to-weight ratio of the UAV,we select an infrared vision system consisting of infrared camera FLIR TAU2-336 and image processor Jetson TX1.The reliability of the system is verified through the actual flight test.Aiming at the problems of poor real-time,low classification accuracy and less features of crowd abnormal behavior detection,we design an abnormal behavior detection method based on crowd density estimation and crowd motion estimation.Since there are no published infrared-based abnormal behavior datasets,we self-build a new infrared dataset,which includes sample pictures and videos in different scenes of public areas.According to the relationship among crowd abnormal behavior,crowd density and the average speed of the crowd,the crowd abnormal behavior is detected by combining density and average speed of the crowd.Multi-task cascading convolutional neural network(CNN)is designed to divide crowd density estimation into two related sub-tasks: crowd count classification(which we call as high-level prior)and density map estimation in a cascaded fashion.By embedding high-level prior knowledge into the density map estimation task,a high-quality density map with low-count error is obtained.In order to estimate the average speed of crowd,the pyramid LK optical flow method is used to track th e motion corners,and then the motion vectors of the motion corner points in two consecutive frames is obtained.Finally,aiming at two crowd abnormal behaviors of aggregating and escaping,the experimental results show that the monitoring UAV system can achieve the detection of crowd abnormal behavior in public areas effectively. |
