A Study On Remote Tower Target Detection And Tracking Based On Deep Learning

As an emerging technology with low cost and high efficiency,remote tower can monitor the airport remotely,which is of great significance to reduce the operation cost and ensure the safety of the airport.Target detection and tracking is the core technology of remote tower,and it is also the premise of aircraft target recognition,conflict behavior prediction and other functions.However,because the detection targets of remote tower are mostly remote scenes and small targets,it brings many difficulties to the detection of moving targets.In view of the above problems,this paper studies the remote tower target detection algorithm,and proposes an improved algorithm based on Yolo V5.In terms of reducing the model parameters and the amount of computation,ghost convolution and DW deep separable convolution with less parameters are used in the network structure,and senet attention mechanism is introduced to improve the accuracy;In the aspect of enhancing the detection effect of different scale targets: first,sknet is used to assign different weights to convolution cores of different sizes,so that the network can automatically expand or narrow the receptive field for targets of different scales;second,Bi FPN structure is used in the neck feature fusion part to give different weights to features of different scales.The small-scale box loss is given a regulating factor to improve the detection effect of small targets;The weighted NMS is replaced by Diou NMS to improve the detection effect of mutual occlusion.In this paper,we make and annotate the data set which is suitable for the application scenario,and propose a data enhancement method of small target scale adaptation to improve the detection performance of the model for small targets.Experiments show that the work of this paper can greatly improve the detection effect of the network for small targets.The final training model size is only 3.9m,reasoning speed is only 8.04 ms,using remote tower video data for testing,map is 94%,such a high-precision lightweight model can be well deployed in embedded devices.In this paper,two extended applications of the proposed detection algorithm are made.One is to input the results of target detection to the deep sort algorithm to realize the target tracking.Through the conversion matrix obtained from camera calibration,the pixel coordinates in the video are converted to the world coordinates,and then the detection results are fused with ADS-B data through the world coordinates to display the flight number in the real-time video,The second is to expand the algorithm of aircraft shutdown point detection,which can not only return to the target detection frame of the aircraft,but also return to the nose,tail,left wing and right wing of the aircraft.