Spectrum three-dimensional reconstruction method and system based on relocation technology, and storage medium

Abstract

The invention relates to a spectrum three-dimensional reconstruction method and system based on a relocation technology, and a storage medium. The method comprises the following steps that a dense point cloud of a field RGB image is constructed through an SLAM technology; a field three-dimensional model is constructed based on the dense point cloud of the RGB image; a ray casting algorithm is utilized to complete texture mapping of the on-site three-dimensional model; a deep convolutional neural network is utilized to realize feature extraction of a spectral image, and pose estimation of a spectrogram is completed based on an extracted feature descriptor, so that relocation of spectral features is completed. On the basis of three-dimensional reconstruction of the RGB image, three-dimensional spectral imaging is achieved in combination with the matched multispectral image, and the method can be widely applied to three-dimensional spectral imaging material evidence inspection.

Description

technical field [0001] The invention relates to a spectral three-dimensional reconstruction method, system and storage medium based on relocation technology, and relates to the technical fields of computer science, pattern recognition and visual algorithm. Background technique [0002] The traditional multispectral camera can obtain the composition information of the substance, but because the information it collects is two-dimensional, there are omissions of the detailed information of the physical evidence, and the spatial position relationship of the substance cannot be obtained. However, the existing three-dimensional reconstruction technology based on color RGB images does not take into account the image pixel information of the multi-spectral image, making it impossible to clearly display important physical evidence such as fingerprints, blood stains, and semen spots in three-dimensional space. When collecting on-site physical evidence information, there will be omissi...

AI Technical Summary

Problems solved by technology

[0003] 1. Spectral images cannot be directly used for 3D reconstruction

In order to be able to clearly capture information such as fingerprints and blood stains, spectral images are all cameras with large apertures and close depth of field. Although local details are relatively clear when shooting, the imaging of the overall object at the far end of the focal plane is relatively virtual, and the entire physical evidence object or scene cannot be captured. All clear shots

Therefore, it is necessary to combine the images of high-definition cameras to complete 3D reconstruction, but how to accurately map spectral information to high-definition 3D models has not been effectively realized so far;

[0004] 2. Traditional feature descriptors are not suitable for spectral information mapping

For example, scale-invariant feature transform descriptor (SIFT) and binary-based feature extraction descriptor (BRISK), etc., although feature extraction and description calculations can be performed on spectral images, the accuracy of extraction is still different from that of ordinary RGB images. This gap is mainly reflected in the difference in pixel values ​​and pixel distribution between the multispectral image and the color RGB image

Moreover, the calculated descriptors are also difficult to match with the extracted descriptors on ordinary RGB images, which brings difficulties to the relocation and registration of spectral images on 3D models;

[0005] 3. The data volume of spectral photos is huge, and the matching dimension is too large

A pose will have dozens or even hundreds of spectral image sets with different wavelengths. As the scale increases, the amount of data that needs to be matched is huge. The appearance of perceptual aliasing makes the matching of traditional methods blurred and weakens the positioning. Robustness of

Images