Research On Mutli-target Tracking Technology Of Air-to-ground Observation

The multi-target tracking system equipped on the airborne platform for earth observation has the advantages of strong mobility,good tracking and monitoring effect,and high adaptability.And it has a wide range of application prospects in both military and civilian fields.The distance between airborne platform and air-to-ground observation targets is far,so then targets have limited imaging pixels on the image sensor and the characteristics of small targets.Furthermore,targets may be partially occluded,which is increasing the difficulty of target detection and tracking.Based on the Aviation Science Fund project,this paper focuses on detecting and tracking technologies of small targets for air-to-ground observation on airborne platforms in order to achieve accurately identifying and real-time tracking of small targets.And the technologies have a good robustness for tracking small targets in complex terrain backgrounds.The research work in the paper mainly includes:1)A database of air-to-ground observation of dim and small target detecting was built.Due to the weakness of space-to-earth observation targets and the complexity and particularity of the background,the use of ordinary public datasets cannot effectively verify the performance of the algorithm in tracking targets under the space-to-earth observation environment,and above 12,000 targets data of space-to-earth observation were constructed by means of target and background modeling for detection model training.The multi-target tracking dataset was made to provide effective support for the verification of multi-target tracking algorithm.2)A dim and small multi-target detection network was studied.For the target detection algorithm,the YOLOv5 m network model was selected as a research object to improve the detection ability of dim and small targets by adding the detection layer of small targets.In the combination of the original feature fusion network Feature Pyramid Network and Pyramid Attention Network structure,four down-sampled feature maps were used as the input of the network to reduce the receptive field of the down-sampled feature map.At the post-processing end,DIo U-NMS was selected to take the place of the traditional non-maximum suppression NMS,and the GIo U＿Loss of YOLOV5 m was selected to replace YOLOV5 m in DIo U＿Loss to accelerate the bounding box regression rate and improve positioning accuracy,so as to improve the detection ability of dim and small targets.The accuracy and adaptability of the improved algorithm were analyzed through experimental comparison.3)A dim and small multi-target tracking algorithm was studied.For the multi-target tracking algorithm,the mainstream multi-target tracking algorithm Deepsort framework was used to increase the long-term occlusion matching method of the target occlusion problem,so that the target can be tracked for a long time,the prediction retention time of the tracking algorithm was extended,the correlation accuracy was improved,and the target ID conversion rate was greatly reduced.Finally,the accuracy and stability of the tracking algorithm were analyzed in the public dataset and the geodetic background dataset.