| Vehicle load is one of the most important load of bridges.It is the main cause of fatigue deterioration of bridges,and is also an important basis for bridge time-varying reliability analysis,remaining life prediction and ultimate bearing capacity evaluation.Combining bridge monitoring video and the data from the weight in motion system to carry out bridge vehicle load monitoring research has good application prospects and has attracted more and more attention in recent years.How to establish the corresponding relationship of vehicles in different monitoring images,so as to realize the direct migration of data from the weight in motion system is a key issue in this research.As a computer vision method to establish correspondence between images,the image feature matching method is an effective solution.In view of the challenging problems that affect image matching effect such as image scale,angle,and illumination changes in the bridge monitoring scene,this paper proposes an image matching method based on the data-driven Hard Net deep learning descriptor.Compared with the traditional image feature descriptors whose matching results are not robust enough due to the loss of feature information through manual design of feature extraction rules,data-driven deep learning descriptors can fully extract and retain feature information of feature regions to achieve ideal matching effect.The main research contents of this paper are as follows:(1)This paper has established an image matching method based on Hard Net deep learning descriptor.The method is divided into three stages: feature detection,feature description and feature matching.In the feature detection part,the AGAST feature detection algorithm is used to quickly detect stable point features in the image scale pyramid.In the feature description part,the detected point feature and its surrounding area are input as the feature area into the Hard Net deep neural network pre-trained with large-scale image data to obtain the Hard Net deep learning descriptor.In the feature matching part,the brute force matching method and SIFT matching criteria are employed to establish the descriptor matching relationship between images,and the RANSAC method is combined to assist in eliminating the wrong matching relationship.(2)The performance of the proposed image matching method is verified by using real bridge static monitoring images,and a comparative study is carried out with three traditional feature matching methods.The experimental results show that the method proposed in this paper has strong robustness to image changes such as scale,angle,and illumination changes in the actual bridge monitoring scene,and can obtain more and more stable point feature matching relationships than traditional methods.In the large-scale vehicle image test data set,the m AP and rank-k indicators were introduced into the performance quantitative evaluation test.The image matching method based on the Hard Net deep learning descriptor obtained93.79% of m AP and 98.89% of the rank-1 indicator values,respectively.Greater than the result of the comparison method,it is verified that the method in this paper can achieve the matching of the target vehicle image with higher accuracy and more stability.(3)A strategy of matching target vehicles based on video search is established,and the real bridge dynamic monitoring video data is used to further verify the practicability of the proposed Hard Net deep learning descriptor image matching method.The test results show that the method can still match the target vehicle image with high accuracy in the video,and can effectively establish the corresponding matching relationship of vehicles in different surveillance videos.Finally,based on the image matching method,the bridge vehicle load identification analysis process is established,and the continuous analysis of the bridge vehicle load distribution can be realized under the premise of the data from weight in motion system. |
PDF Full Text Request