Research On Surface Moving Target Detection In Navigation Airports

In recent years,with the introduction of the‘14th Five-Year Plan’aviation plan,the Intelligent civil aviation plan is also constantly advancing and developing,and the number of aircraft is also increasing,the number of aircraft is also increasing,and the runways and taxiways of the airport are becoming larger and more complex.Due to the unreasonable design of the airport surface road and the negligence of the controllers,the probability of unsafe incidents on the ground is increased.In view of this situation,large and medium-sized busy airports have established a scene monitoring system to improve the monitoring efficiency and reduce the risk of airport surface taxi collision.However,due to the shortage of funds,the general aviation airports are often unable to use the scene surveillance radar,multi-point positioning and other surveillance means with good effect but high cost^[1].At the present stage,the scene surveillance means also have the disadvantages of high cost,blind area of surveillance,vulnerability to weather,and non-intuitive display^[2].Therefore,how to realize the all-day and low-cost monitoring of the general aviation airport surface and the detection of the moving targets on the surface has become the key to improving the safety of the general aviation airport^[2].Artificial intelligence is in the development boom,and various algorithms in the field of machine vision emerge in endlessly.The object detection algorithm based on deep learning is a rapidly developing direction in the field of machine vision,and the technology has very high robustness and mobility.Therefore,it is applied in many fields of daily life,but this kind of algorithm has never been applied to the detection of airport surface targets.Therefore,aiming at the defects and problems in the field of general aviation airport surface moving target detection methods,the target detection algorithm in deep learning is integrated with the airport surface operation scene to establish a high-precision,all-day,low-cost general aviation airport surface detection model^[2],in order to improve the operation safety of general aviation airports and assist controllers to make reasonable decisions.This thesis studies and improves the existing target detection algorithm,and constructs a model suitable for moving target detection on the surface of general aviation airport.The main innovations are as follows:（1）Improvement of the backbone network.In response to the issue of redundant information generated by ordinary convolutions in the traditional YOLO（you only look once）v3 algorithm,an improvement has been made to the backbone network.A new backbone network has been constructed using deep separable convolutions instead of the original convolutions,which improves the detection speed of the model without losing detection performance.（2）Improvement of loss function.In view of the problems such as inadequate training and slow convergence speed in the traditional YOOv3 algorithm,the loss function is improved,and the distance intersection over Union（DIoU）function is used as the position loss function^[2]to accelerate the convergence speed and detection accuracy of the model.（3）Improvement of training methods.In the training phase,a combination of transfer training,freezing training,thawing training,breakpoint recovery,and hierarchical training is used to improve the training speed and quality.Freeze training and thawing training accelerate the training speed while improving the training accuracy.Breakpoint recovery and hierarchical training can help fuse the weights of multi scene training to improve the accuracy and speed of target detection.（4）Comparison of detection effects in different weather scenarios.Based on the improved YOLOv3 algorithm,a comparative analysis was conducted on the performance of moving object detection in three different scenarios at a certain airport in Southwest China.The results show that the improved YOLOv3 algorithm performs significantly better in detecting surface moving targets in normal weather scenarios than in foggy and rainy weather scenarios,and in detecting aircraft targets,it outperforms vehicle and pedestrian targets^[2];The detection accuracy,recall rate,and mean average precision（m AP）of the three types of targets have reached 92.96%,80.51%,and 91.96%,respectively.The GPU processing speed is 74f/s,which significantly improves the performance compared to other current popular algorithms.