UAV Aerial Image Target Recognition And Tracking

In recent years,with the development of unmanned aerial vehicle technology,unmanned aerial vehicles have been widely used in military,industrial,and civilian fields,and the target recognition and tracking of unmanned aerial vehicles has become a research hotspot.Due to the complex scenes in aerial images,the different target scales,and the problems of illumination changes,occlusion interference and camera shake,target recognition and tracking in aerial images become more difficult.In response to these problems,this paper studies the target recognition and tracking of UAV aerial images,and mainly completes the following tasks.Firstly,a real-time aerial image detection model based on multi-scale feature fusion is proposed by improving the YOLOv4 detection model.By using CSP Darknet as the backbone network,the feature pyramid structure of global and local information fusion is constructed.On the basis of the three detection layers of YOLOv4,a detection layer for small target detection is added,and a four level multi-scale detection layer is constructed.The training and testing are carried out on the visdrone2020 det data set.Compared with the original YOLOv4,the performance of the improved model in the aerial image target detection task is improved,in which the recall rate of pedestrians and cars is increased by 9.6%,the overall recall rate is increased by 5.5%,and the mean Average Precision(m AP)is increased by 5.2%.At the same time,the detection speed almost does not decline,which effectively verifies that the proposed detection algorithm has good detection performance for pedestrians and vehicles in the aerial scene.Then,in order to realize target tracking in aerial images,a multi-target tracking algorithm based on motion prediction and detection data association is proposed.Kalman filter is used to predict the target motion state,and the IOU between the detection result and the prediction result is calculated.As a measure of spatial similarity,the loss matrix is constructed.Then the Hungarian algorithm is used to realize the correlation matching between the detection result and the tracking prediction.At the same time,the management module of the target tracker is introduced to manage the life cycle of the tracking target.The experimental results show that the algorithm can track the target effectively,but it still needs to be improved on the problem of ID switch.Finally,in order to reduce the number of ID switch,a multi-layer Association matching mechanism based on the fusion of spatial information and apparent information is proposed.The experimental results show that the multi-target tracking integrated with the apparent information increases the clues between the detection and tracking tracks,effectively reduces the ID switch,and the target tracking is more stable.