Method And Implementation Of Abnormal Event Detection System Based On UAV Platform

In recent years,with the development of civil and industrial UAVs,rotor UAVs are widely used in plant protection,fire protection and security fields.Many researchers began to transplant various computer vision technologies into embedded platform to build intelligent monitoring system based on UAV.However,these monitoring systems based on UAV are far from the ideal state,and they fail to fully tap the potential of UAV's high mobility,and also fail to effectively utilize the ability of UAV video monitoring.At present,most UAV monitoring systems only run the similar monitoring tasks with the fixed camera monitoring system in hover state.How to make the UAV move during the mission?How to further explore the potential of UAV video surveillance?How to improve the accuracy of the detection of abnormal events on UAV platform?How to use the limited hardware resources of UAV embedded platform to complete the above tasks?These are all the problems that need to be solved.This paper focuses on the detection of abnormal events based on UAV platform,and explores the algorithm and system implementation of UAV monitoring system.In view of the above problems,the main work of this paper is as follows:1)A two-stage monitoring framework for UAV abnormal events is designed.In the first stage,the application of airborne computing resources in the flight state is to realize the coarse-grained unsupervised abnormal event monitoring.The purpose is to find out the ground coarse-grained abnormal event,and provide the UAV monitoring system when to perform hovering flight attitude,and then carry out the next fine-grained monitoring task.In the second stage,the detection task is to use the network bandwidth resources to transmit the ground situation in hover state to the remote server,and then monitor the violence anomaly based on C3d neural network.2)Establish the data set of UAV and UAV abnormal events and UAV campus violence events.According to the characteristics of UAV and the needs of monitoring tasks,the data sets of two-stage UAV abnormal event detection tasks are established.3)A frame difference search method for moving object detection is proposed.The algorithm is used to improve the accuracy and efficiency of moving target detection from the perspective of UAV.In this paper,we try to introduce the idea of selective search algorithm while using the frame difference method,and add the results of the frame difference algorithm as an image region feature into the search algorithm based on image segmentation,merge the relevant regions according to the image region similarity,and finally directly get multiple nominations of moving objects.And in the multiple nominations of single frame detection results,the method of fusing and pruning multiple nominations is used to get the single frame detection results in an heuristic way.The experimental results show that the fusion of the two algorithms takes both accuracy and execution efficiency into account.The optimized algorithm can get real-time moving target detection results in the embedded platform of UAV.4)A trajectory clustering anomaly detection method for unsupervised anomaly event detection is proposed.This method is used in the first phase of the unsupervised abnormal event detection task,the purpose of which is to get the preliminary detection results of the ground abnormal situation in the flight state of the UAV.Firstly,the track feature is extracted from the track information of moving object detection.Then we use the clustering based anomaly detection method to get the anomaly detection results on the ground,and achieve 86.7%detection accuracy and 91.1%anomaly detection recall on the data set.5)A method of violence anomaly detection based on C3d neural network is proposed.This method is used in the second phase of the surveillance of violent incidents detection task,the purpose is to get more accurate ground anomaly detection results after hovering UAV.First,the C3d neural network model used in this chapter is introduced,and then three violence data sets selected for monitoring tasks are introduced:hockey fit data set,movies data set and campus violence data set collected by ourselves.Finally,model training and testing are carried out on three data sets,and the detection accuracy is generally better than that of traditional feature extraction.6)Design and implement the anomaly monitoring system based on UAV platform.The system is suitable for the monitoring task of ground anomaly.Through the two-stage monitoring task mentioned above,the execution efficiency and response time of the whole system are taken into account while ensuring the detection accuracy.Finally,the system functions are realized by several software modules on raspberry pie,Android mobile phone and remote server.