Object Detection Algorithm For Airport Scene

In recent years,China’s civil aviation industry has achieved rapid development,with airport throughput increasing year by year.However,this has also led to the problem of airports becoming increasingly crowded and monitoring scenes becoming more difficult.Currently,most airport scene detection methods rely on manual visual surveillance,but its limitations are too great to ensure absolute airport safety.The airport poses significant challenges for aircraft object detection due to the wide range of scenes,numerous small objects,large differences in object scales,the wing of the aircraft stretched out causing the object to overshoot.This thesis uses deep learning technology to build a detection system.In response to the difficulties mentioned above,research on camera object detection in airport scenes was conducted.An object detection network was designed,and the network was further improved while considering the speed of network inference and improving the accuracy of network detection.The main research work is as follows:(1)This thesis designs an airport object detection network to address the problem of significant differences in object scales and small object proportions in airport scenes.The network includes a backbone and a neck,which consists of gradient flow module and super down-sampling module for better control over the feature transmission path and improved network perception field.Additionally,it also has a tri-directional concatenation module and reparameterization module for more comprehensive utilization of signals at different levels,achieving decoupling between training and inference networks to improve inference speed.YOLO series’ three-head prediction networks are used to enhance the detection effect on objects with different scales.The network achieved an AP score of 79.1%on AGVS airport detection dataset.(2)In order to further improve the accuracy of the algorithm without affecting the speed of network,this thesis decouples the detection head.The classification and regression tasks are calculated using two separate detection heads,thus solving the problem of classification and regression conflicts and improving the upper limit of network performance.Based on this,anchor-free transformation was performed on the network,merging the feature maps generated by the original detection head branch into a single feature vector,reducing candidate box numbers.The SimOTA dynamic label assignment method was used to complete network label assignment work while improving algorithm accuracy and reducing additional computational costs caused by decoupling operations.Compared with previous detection algorithms,this algorithm has improved accuracy by 5.3percentage points with almost no impact on inference speed.(3)In order to further improve the accuracy of algorithm detection and solve the problem of missed small objects,this thesis separately added a coordinate attention module after the backbone.This allows the algorithm to exclude other factors in complex environments and more effectively extract features related to aircraft.The algorithm has improved its accuracy by 1.2 percentage points compared to the decoupled network with almost no additional computational cost.