Research On Airport Detection Data Set Based On Deep Learning

Airport scene surveillance is a complex problem.With the development of civil aviation in China,it is more and more difficult for airport management personnel to manage the target in airport surface.There are many safety accidents in recent years.Object detection is the fundamental of airport surface monitoring,and its purpose is to accurately identify and locate the airplane,the detection effect will directly affect the results of the tracking and action recognition,which is very important in the scene surveillance system.In last few years,with the rise of deep learning,the target detection method of deep learning makes the detection accuracy reach a high level.Deep learning relies heavily on the training sample,however,there is no training dataset for airport surface target detection,many special cases in airports cannot be trained on public datasets and it will seriously affect the performance of the detection algorithm.In addition,there is a general imbalance problem in target detection.Due to the large scene area of the airport,the background of the monitoring image is far larger than the target,the imbalance problem is more serious in the airport environment.Therefore,it is a meaningful and challenging task to research and make an airport detection dataset that includes various situations on the airport field and meets the needs of algorithm training,and propose the detection method to solve the serious imbalance problem in the airport based on the dataset.The main content of this article is as follows.1.We have discussed the design scheme of the mainstream object detectiondatasets,and we used YOLO,Faster-RCNN and SSD algorithms to train adetector for each dataset.In the detection results of each detector in the airportscene monitoring data,we found that the existing datasets have great defects inaircraft detection: The samples of the mainstream datasets for the aircraft onlycontain a few angles and attitudes,and these samples,whose the aircraft objectis obvious and the background and the object scale is single,can be very easilydetected.The performance of these detectors is not ideal even in the normalairport environment,so professional airport detection dataset are indispensablefor airport aircraft detection.2.According to the characteristics of each mainstream dataset,this paper designeda data collection and selection scheme for the airport environment,andcompleted data annotation.The image annotation tool of the dataset is Labelimg.The dataset contains a training set of 6,000 images and a test set of 1,500images.In total the dataset which contain objects with different weather,different lighting,different scales and partial occlusion has 20,000 labeledinstances in 7,500 images and these instances contain.The performance of thedetectors which is trained by our dataset is far better than which is trained byothers dataset.3.In order to solve the problem of imbalance in airport detection,we propose amethod for detecting unbalanced learning.Compared with the currentmainstream research on model structure,this method mainly focuses on thetraining process.The innovation of the method is that we use balanced featurepyramid to reduce the imbalance in the process of propagation and introduce thefocal loss to reduce the imbalance in the process of sampling.The methods havebeen verified in our dataset,The result show that the proposed methods caneffectively suppressed the imbalance in the process of training and improve thedetection accuracy..