Research On Aerial Target Detection And Tracking Method Based On Convolutional Neural Network

Aerial target detection and tracking technology is the basis for real-time monitoring of the operation of UAV targets on the ground,and one of the key technologies to achieve intelligent application of UAVs.In actual scenes,ground targets have a wide field of view,a small size,and a complicated background.The general algorithm is not ideal for target detection and tracking in special scenarios.Therefore,this paper takes target detection and tracking from the perspective of aerial photography as the research direction.Based on the research of computer vision technology and deep learning theory,we explore more efficient aerial target detection and tracking algorithms to improve the detection ability of aerial small targets in complex environments and enhance the stability of aerial moving target tracking.This paper analyzes the domestic and international research and development status of convolutional neural network,target detection and tracking technology,expounds the difficult problems faced by target detection and tracking in aerial images,and optimizes and improves the existing problems.The contents of the key research are as follows:1.This paper constructs aerial target detection model in a complex environment combined with random depth residual network.First of all,this paper studies and analyzes the impact of different deep networks on model performance.It uses deep residual networks to train detection models to improve the learning ability of the network.Second,through the residual learning mechanism in deep residual networks,it reduces the degradation problems caused by deep networks.Finally,a random depth mechanism is introduced to adjust the network structure and reduce the network complexity.The algorithm realizes the efficient learning performance of deep networks and improves the classification and detection capabilities of aerial targets in complex backgrounds.2.This paper proposes an improved aerial small target detection method based on multi-scale feature pyramid.Since the features of small targets mostly exist in shallow networks,as the network deepens,the features will become fewer and even disappear.In this paper,a pyramid structure of multi-scale feature maps is introduced into the detection network to better integrate the detailed feature information existing in the shallow layer with the rich semantic information contained in the deep layer,extract the features of small targets more fully,and improve the detection ability for small targets in aerial images.3.Using the constructed aerial photography target detection model as a detector,combined with Kalman filtering to construct aerial photography moving target tracking model.The detector is used to detect the target of the initial frame in the aerial video,extract the target position and bounding box information,track the detected target through the Kalman filter,and call the detector again to update the target when the target is lost.The model can detect and track moving targets in aerial video.Tracking experiments were carried out in various scenarios,which verified that the tracking method proposed in this paper has good robustness.