Research Of Road Foreign Object Detection Method Based On 3D Point Cloud Data

Airport runway foreign object debris pose a huge threat to the aircraft’s take-off and landing process and the safety of all passengers and crew.It is a key monitoring area for airport safety.Therefore,it has very important research significance for foreign body detection methods on special roads.In response to the demand for foreign debris detection on the road surface,combined with the three-dimensional data processing technology and three-dimensional neural network that have been fast-growing in recent years,this paper proposes a road foreign debris detection fusion model based on the improved PointCNN and VGG16.Because of the limitation of the airport’s special environment,this paper chooses to do foreign debris detection research on ordinary roads.First of all,this article uses traditional image feature recognition network as a comparison to illustrate the difficulties of deep learning to process 3D data,draws out and analyzes the current methods of dealing with 3D data irregularities.Then introduce the equipment used in the experiment,the principle of data collection,and denoise the collected data based on the improved filter.For the vibration problem in the 3D data collection process,based on the analysis of the vibration form,a transverse convolution method is proposed to compensate the vibration data,which better eliminates the impact of equipment vibration.Then nonlinearly enhance the three-dimensional data in the spatial domain to highlight the features of foreign objects.Secondly,the relationship from depth image to point cloud is analyzed,and various point cloud data downsampling methods are tested and analyzed.Due to the poor sampling effect of the existing methods on pavement point cloud data,this paper proposes a gradient downsampling method by analyzing the distribution features of the road surface foreign debris data.This sampling method not only effectively retains the target structure characteristics but also have higher computational efficiency than the average downsampling method.Then the common pavement foreign debris and the foreign debris used in the test are described,and the road foreign debris data set of this article is constructed.Finally,the point cloud detection method of foreign debris on the road surface based on deep learning is studied.Through experiments,a variety of point cloud feature detection models are analyzed,and the PointCNN network with the best initial training results is selected as the basic model.Through redesign the classification structure and improve the parameters,the classification accuracy rate of the model for the point cloud of foreign matter on the road surface reaches 90%.At the same time,combined with the detection advantages of the traditional CNN network on the image,the three-dimensional data is reduced to the two-dimensional image for detection through the VGG16 network.The result matrix of the point cloud model and the image model is weighted calculation by the Log Softmax.Data test show that through the optimal combination the accuracy of the fusion model reaches 96.8%,which is close to 6.8% improvement compared to the accuracy of a single three-dimensional model.In summary,this article proposes a lateral convolution algorithm to solve the common vibration problems in 3D data acquisition.Then,facing the problem of excessive data volume,a gradient down-sampling method was designed to down-sample the point cloud.Finally,the PointCNN and VGG16 models are fused to detect foreign objects on the road.The systematic research is carried out from the collection and processing of road data to foreign debris detection,which has application value and reference significance for foreign body detection on key roads.