| Underground pipelines and voids are two very important targets in road detection,where underground pipelines can hinder road construction and voids defects can affect the use and life of roads.Traditional detection methods such as drilling and gravity are not efficient,and the detection range of the pipeline detector is limited.Ground-penetrating radar technology can detect road underground targets non-destructively and quickly,but the amount of data collected by ground-penetrating radar is very large,and data analysis mainly relies on manual.Therefore,this thesis proposes a deep learning-based ground-penetrating radar road underground multi-target detection method,first using a ground-penetrating radar detection vehicle equipped with a deep learning model to conduct a large-scale general detection of the road and mark the suspected target area;then using a combination of handpush ground-penetrating radar and deep learning model to conduct detailed detection of the suspected area and obtain the precise distribution of underground targets.Firstly,the radar data of the void and pipeline targets are collected using the hand-push ground-penetrating radar,and the acquired raw data are subjected to FIR digital filtering,signal gain adjustment,direct wave removal and Hilbert transform in a series of operations to enhance the features of the targets,and the processed radar data are output into image format;random cropping,mirror flip and contrast enhancement methods are used to enhance the quality of the image data set;Labelme tool is used to annotate the images and construct the image dataset of ground-penetrating radar road underground targets needed for deep learning models.Secondly,this thesis proposes an improved method for rapid detection of underground multiple targets based on RetinaNet,which aims to meet the requirements of detection speed and accuracy in the process of ground-penetrating radar detection vehicle general detection.The feature expression capability is improved by adding the attention mechanism CBAM to the RetinaNet feature extraction network;the neural architecture search method(NAS)is used to find a multi-scale feature fusion method that fits the dataset of this thesis to enhance the fusion capability of the model for different scale features;the GIo U loss function is introduced to calculate the bounding box error more accurately to improve the detection accuracy of the model.After the improvement of the three methods,the m AP of RetinaNet network at Io U=0.5 is 88.7%,while the detection speed of 20.3 frames/s is maintained,and the improved RetinaNet network can meet the demand for speed and accuracy of ground-penetrating radar detection vehicles.Finally,the hand-pushed ground-penetrating radar is used for detailed detection of the suspected areas marked during the general detection,when the model is required to have high detection accuracy.Therefore,this thesis optimizes the high-precision detection model Cascade R-CNN to further improve its detection accuracy on the dataset of this thesis.ResNe Xt-101 with fused group convolution is used as the feature extraction network to enhance the feature extraction capability of the model;the bidirectional strategy PA-FPN is used to fuse multi-scale features to improve the semantic expression capability of the network;the non-maximum suppression method Soft-NMS is used to reduce target miss detection and improve the accuracy of detection results.The m AP of the improved Cascade R-CNN model reaches 92.6% at Io U=0.5,and the effectiveness of the improved Cascade RCNN model is verified by ablation experiments and comparison experiments.In summary,this thesis improves the RetinaNet model with high detection speed and the Cascade R-CNN model with high detection accuracy for the general and detailed detection stages of the road detection process,respectively,thus realizing the rapid and accurate detection of road underground targets and providing important reference values for road construction and maintenance units. |
PDF Full Text Request