| Airplanes play an important role in modern society as a safe and fast means of transportation.With the continuous development of aviation technology,the stability of aircraft components has also been continuously improved.However,aircraft safety inspection is still an important task.Among them,the inspection around the aircraft before takeoff is of great significance as the last safeguard measure for flight.This inspection involves the inspection of the appearance of several important parts of the aircraft,which not only requires a lot of manpower and material resources,but also cause misjudgments and missed judgments due to the mistakes of the ground staff.With the rapid development of artificial intelligence and computer vision technology,this problem can be solved in an intelligent way.This thesis proposes aircraft appearance intelligent detection algorithm mounted on an airport inspection robot.This algorithm can detect key parts of the aircraft based on the images collected by the camera during the inspection robot’s winding process,find out safety problem and locate the problem.The algorithm mainly includes two parts:the object detection algorithm and the binocular ranging algorithm.In order to be mounted on the inspection robot,this thesis finally deploys the algorithm on the NVIDIA embedded platform Jetson AGX Xavier,and completes the accurate detection by improving and optimizing the algorithm with real-time running targets.In the part of object detection algorithm,in view of the low detection accuracy of small target on the traditional one-stage object detection model,this thesis uses the YOLOX network as the framework and uses the Swin-Transformer structure to improve the YOLOX backbone network and improve the feature extraction capability of the backbone network;To improve the detection ability of difficult-to-detect targets,this thesis uses Focal Loss to improve the loss function of the network,so that the model can focus more on difficult samples.The experimental results show that the improved network in this thesis achieves an accuracy of 86.24%on the aircraft appearance inspection dataset,which is 1.27%higher than that of the YOLOX network.After the acceleration of the NVIDIA platform TensorRT,the single-frame operation speed reaches 7.03ms,which ensures the real-time performance of the operation.In the part of binocular ranging algorithm,in view of the problem of poor real-time performance of traditional binocular ranging algorithms,this thesis proposes a fast ranging algorithm based on the extracted obj ects of interest.The most commonly used SGBM algorithm in binocular vision has high accuracy.but it cannot meet the requirements of this thesis in terms of real-time performance.The improved algorithm in this thesis is based on the improvement of the SGBM algorithm,which can be combined with the results of the object detection algorithm to greatly reduce the amount of computation in the matching process without losing accuracy.At the same time,this thesis proposes the MAD-Census transform to improve the SGBM algorithm based on mutual information matching cost calculation method.This method not only directly reduces the computational complexity,but also has good adaptability to multithreading,which greatly improves the application process.Compared with the SGBM algorithm,the error matching rate is only 5.2%.The final distance test experiment shows that the ranging error of the algorithm is basically controlled below 10%.Finally,this thesis designs an aircraft appearance intelligent detection system based on this algorithm,and tests the function of each module of the system and the real-time performance of the system.The system includes object detection and ranging module,robot motion control module,and network module.The test results show that each module can output the results correctly.After the improvement of the algorithm of each module and the optimization of multithreading,the operation efficiency is significantly improved.The time-consuming of a single frame operation of the system is about 60ms,which can reach 16 to 17 frames per second,and the processing time can meet the real-time requirements. |
