Airport Surface Object Detection Based On Deep Learning

Modern airports is a complex system,the controller’s monitoring on airport surface activities is becoming increasingly difficult,leading to the more and more obvious potential safety hazard,and so it is of great importance to develop an airport surface video surveillance system.Object detection,as the first step of video surveillance,is directly related to the whole system,thus playing a crucial role.However,due to the existence of object occlusion,changeful scene,camouflage effects,small target and motion blur,object detection has been a hotspot and difficulty in research of computer vision all the time,a large number of scholars are trying to study the various problems in it.In recent years,deep learning,a special kind of neural network,is becoming more and more popular.Compared to the traditional object detection method,deep learning method can learn richer and more robust feature from samples thus yielding better results,therefore,it is a meaningful and challenging work to study the object detection based on deep learning method and apply it to the airport surface monitoring.The main content of this article is as follows.1.We have discussed the design idea of object detection algorithm based on deep learning.According to the survey of existing object detection algorithms,the idea can be divided into two parts: on the one hand,region proposal algorithm such as selective search,region proposal network is applied first,and then followed by a well-designed detector;on the other hand,a single neural network directly outputs detection results.The latter is more advantageous in speed and suitable for our main topic.2.Based on the Single Shot Multi Box Detector(SSD),which takes a good trade-off between speed and accuracy,we summarize the idea of improving the algorithm by improving its base network.The three classification networks: Inception V3,Res Net50 and Mobile Net,are integrated into SSD in the performance and calculation consumption.The experiments show a better performance than the original algorithm.3.In order to solve the problem of imbalance in the one-stage object detection algorithm,the focal loss is introduced to improve the training process.Then,the network structure of Retina Net is introduced and optimized.Finally,the experiment shows that the improved detection algorithm has higher detection accuracy and the detection speed can meet the actual demand.