CN116402874A - Spacecraft depth complementing method based on time sequence optical image and laser radar data - Google Patents

Spacecraft depth complementing method based on time sequence optical image and laser radar data Download PDF

Info

Publication number
CN116402874A
CN116402874A CN202310393175.2A CN202310393175A CN116402874A CN 116402874 A CN116402874 A CN 116402874A CN 202310393175 A CN202310393175 A CN 202310393175A CN 116402874 A CN116402874 A CN 116402874A
Authority
CN
China
Prior art keywords
depth
image
level
feature
target
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202310393175.2A
Other languages
Chinese (zh)
Inventor
汪洪源
刘祥
李子奇
王梓健
陈昱
王秉文
颜志强
金后
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Harbin Institute of Technology
Original Assignee
Harbin Institute of Technology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Harbin Institute of Technology filed Critical Harbin Institute of Technology
Priority to CN202310393175.2A priority Critical patent/CN116402874A/en
Publication of CN116402874A publication Critical patent/CN116402874A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/50—Depth or shape recovery
    • G06T7/521—Depth or shape recovery from laser ranging, e.g. using interferometry; from the projection of structured light
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T17/00—Three dimensional [3D] modelling, e.g. data description of 3D objects
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/11—Region-based segmentation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/194—Segmentation; Edge detection involving foreground-background segmentation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/50—Depth or shape recovery
    • G06T7/55—Depth or shape recovery from multiple images
    • G06T7/593—Depth or shape recovery from multiple images from stereo images
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
    • Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
    • Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Software Systems (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • Mathematical Physics (AREA)
  • Health & Medical Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Computing Systems (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Geometry (AREA)
  • Optics & Photonics (AREA)
  • Computer Graphics (AREA)
  • Image Processing (AREA)

Abstract

A spacecraft depth complement method based on time sequence optical images and laser radar data belongs to the technical field of space target three-dimensional structure perception. The method aims at solving the problem that the existing single-frame target depth completion method ignores time related information of continuous frames, so that target time sequence depth completion results are inconsistent. The method comprises the steps of obtaining a target gray level image time sequence of a monocular visible light camera and a sparse depth image time sequence of a laser radar based on a space target three-dimensional model simulation, and generating a target dense depth ground label of an image; and training the time sequence spacecraft depth complement neural network model by using the target gray image time sequence and the sparse depth image time sequence as samples, obtaining the trained time sequence spacecraft depth complement neural network model, embedding the trained time sequence spacecraft depth complement neural network model into a satellite-borne platform, and realizing the prediction of the dense depth of the space target based on real-time sensing data acquired by a monocular visible light camera and a laser radar. The method is used for depth completion of spacecraft detection.

Description

Spacecraft depth complementing method based on time sequence optical image and laser radar data
Technical Field
The invention relates to a spacecraft depth complement method based on time sequence optical images and laser radar data, and belongs to the technical field of space target three-dimensional structure perception.
Background
With the rapid development of space technology, the type of space on-orbit task for human gradually shows a diversified development trend. The three-dimensional structure of the spacecraft is perceived, and the point cloud data are acquired, so that a plurality of space on-orbit tasks are smoothly executed, such as fragment cleaning, on-orbit maintenance, intersection butt joint and the like.
Currently, the mainstream three-dimensional structure perception schemes of spatial targets can be mainly divided into a stereoscopic vision system, a time-of-flight (TOF) camera, and a combination of a monocular camera and a lidar. The stereoscopic vision system recovers the depth of the extracted characteristic points by using a triangulation principle, and has poor effect on a smooth surface or an object with repeated textures; in addition, the base line of the binocular camera greatly limits the working distance of the system, and the space on-orbit task requirement is difficult to meet. TOF cameras calculate the exact depth of a target by measuring the time delay between transmitting and receiving a laser pulse. Although accurate depth can be obtained at high density, the working distance of TOF cameras is typically less than 10m, which is limited by the in-orbit power limit, preventing their use in practical applications. The working distance between the monocular camera and the laser radar is far, and the dense depth of the spacecraft is recovered by utilizing the optical image and the sparse ranging information. Compared with a binocular system and a TOF camera, the combination of the monocular camera and the laser radar can effectively increase the working distance of the system, reduce the sensitivity to illumination conditions and materials, and is more suitable for being used in space practical application.
Since achieving dense depth restoration of targets based on monocular cameras and lidar has important applications in many scenarios, a large number of deep learning-based target depth complement algorithms have been proposed in recent years to meet various depth-based application requirements. Although the problem of target depth completion based on single-frame sensor data has made an important progress, the data to be processed in the actual on-orbit working process is sensor sequence data, and the existing single-frame target depth completion method ignores time related information of continuous frames, so that the target time sequence depth completion results are inconsistent. The present invention is therefore aimed at studying a method of spatial spacecraft depth complement based on sequence data.
Disclosure of Invention
Aiming at the problem that the existing single-frame target depth completion method ignores time-related information of continuous frames, so that target time sequence depth completion results are inconsistent, the invention provides a spacecraft depth completion method based on time sequence optical images and laser radar data.
The invention relates to a spacecraft depth complementing method based on time sequence optical images and laser radar data, which comprises the following steps of,
collecting a plurality of space target three-dimensional models, and setting simulation working conditions of the space target three-dimensional models and sensor parameters of a monocular visible light camera and a laser radar; obtaining a target gray level image time sequence of a monocular visible light camera and a sparse depth image time sequence of a laser radar based on simulation of a plurality of space target three-dimensional models, and generating a target dense depth ground label of the image;
constructing a time sequence spacecraft depth complement neural network model: comprising a plurality of target depth prediction branches cascaded in chronological order;
each target depth prediction branch includes an encoding stage and a decoding stage:
the coding stage comprises a foreground segmentation module, a gray image feature extraction module, a morphological preprocessing module and a depth image feature extraction module; the decoding stage comprises an LSTM module and a deconvolution layer;
the prediction process for the spatial target depth at the time t comprises the following steps:
sparse depth image I at time t s,t Preprocessing by a morphological preprocessing module to obtain a preprocessed depth image, and inputting the preprocessed depth image into a depth image feature extraction module;
target gray scale image I at time t g,t Extracting multi-scale gray scale image feature images with different semantic levels and different resolutions through a gray scale image feature extraction module;
sparse depth image I for time t s,t And target gray scale image I g,t After cascade operation, inputting the image to a foreground segmentation module for target foreground segmentation to obtain a foreground segmentation image;
the depth image feature extraction module firstly processes the preprocessed depth image through a convolution layer, and then performs feature fusion with a gray image feature image step by step according to the sequence from large scale to small scale to obtain a multi-mode data fusion feature F at the moment t t ；
Multi-mode data fusion feature F at t moment t Modeling a time sequence relation with the corresponding level characteristic state at the time t-1 step by step through an LSTM module, and performing characteristic decoding through a deconvolution layer; finally, combining the layer characteristic decoding result with the foreground segmentation image to obtain a target depth prediction result at the moment t;
training a sequential spacecraft depth complement neural network model by using a target gray image sequential sequence of a monocular visible light camera and a sparse depth image sequential sequence of a laser radar as sample images to obtain a trained sequential spacecraft depth complement neural network model;
and embedding the trained time sequence spacecraft depth complement neural network model into a satellite-borne platform, and realizing prediction of dense depth of the space target based on real-time sensing data acquired by the monocular visible light camera and the laser radar.
According to the spacecraft depth complement method based on time sequence optical image and laser radar data, the target gray level image I g,t Extracting gray image feature images with five sizes respectively as target gray images I by a gray image feature extraction module g,t 1/2, 1/4, 1/8, 1/16 and 1/32.
According to the spacecraft depth complement method based on time sequence optical images and laser radar data, a depth image feature extraction module obtains t-moment multi-mode data fusion features F t The process of (1) comprises:
the depth image feature extraction module firstly processes the preprocessed depth image through a convolution layer, and then the preprocessed depth image is processed through the primary residual error module and fused with the 1/2 gray image feature image through the primary feature fusion module to obtain primary data fusion features; the primary data fusion features are processed by a secondary residual error module and then fused with the 1/4 gray image feature map by a secondary feature fusion module to obtain secondary data fusion features; the secondary data fusion features are processed by the tertiary residual error module and then fused with the 1/8 gray level image feature map by the tertiary feature fusion module to obtain tertiary data fusion features; the three-level data fusion features are processed by a four-level residual error module and then fused with the 1/16 gray image feature map by a four-level feature fusion module to obtain four-level data fusion features; the four-level data fusion feature is processed by a five-level residual error module and then is subjected to element addition with the 1/32 gray level image feature map to obtain the five-level data fusion feature which is used as a t-moment multi-mode data fusion feature F t 。
According to the spacecraft depth complement method based on the time sequence optical image and the laser radar data, the method for obtaining the target depth prediction result at the time t in the decoding stage comprises the following steps:
fusion of t-moment multi-mode data to feature F t Five-stage characteristic state at time t-1 through five-stage LSTM module
Figure BDA0004176570860000031
Modeling the time sequence relationship to obtain five times at the time tStage timing enhancement feature->
Figure BDA0004176570860000032
And five-level characteristic memory state->
Figure BDA0004176570860000033
Figure BDA0004176570860000034
And->
Figure BDA0004176570860000035
Five-level characteristic state of common composition t time>
Figure BDA0004176570860000036
Five-level timing enhancement feature->
Figure BDA0004176570860000037
Performing feature decoding through a five-stage deconvolution layer to obtain a five-stage feature decoding result; the five-level feature decoding result passes through a four-level LSTM module and a four-level feature state at the time t-1
Figure BDA0004176570860000038
Modeling the time sequence relationship to obtain four-level time sequence enhancement characteristic ++at the time t>
Figure BDA0004176570860000039
And four-level characteristic memory state->
Figure BDA00041765708600000310
And->
Figure BDA00041765708600000311
Four-level characteristic states which together form the time t>
Figure BDA00041765708600000312
Four-level timing enhancement feature->
Figure BDA00041765708600000313
Performing feature decoding through a four-level deconvolution layer to obtain a four-level feature decoding result; the four-level characteristic decoding result passes through a three-level LSTM module and the three-level characteristic state at the time t-1>
Figure BDA00041765708600000314
Modeling the time sequence relationship to obtain three-level time sequence enhancement characteristic +.>
Figure BDA00041765708600000315
And three-level characteristic memory state->
Figure BDA00041765708600000316
Figure BDA00041765708600000317
And (3) with
Figure BDA00041765708600000318
Three-level characteristic state of common composition t time>
Figure BDA00041765708600000319
Three-level timing enhancement feature->
Figure BDA00041765708600000320
Performing feature decoding through the three-level deconvolution layer to obtain a three-level feature decoding result; the third-level characteristic decoding result passes through a second-level LSTM module and a second-level characteristic state at the time t-1>
Figure BDA00041765708600000321
Modeling the time sequence relationship to obtain the secondary time sequence enhancement characteristic +.>
Figure BDA00041765708600000322
And a secondary feature memory state
Figure BDA00041765708600000323
And->
Figure BDA00041765708600000324
Second-order characteristic states which together form the time t>
Figure BDA00041765708600000325
Second order timing enhancement feature->
Figure BDA00041765708600000326
Performing feature decoding through a secondary deconvolution layer to obtain a secondary feature decoding result; the secondary characteristic decoding result passes through a primary LSTM module and the primary characteristic state at the time t-1>
Figure BDA00041765708600000327
Modeling the time sequence relationship to obtain the first-level time sequence enhancement characteristic +.>
Figure BDA00041765708600000328
And the first-level characteristic memory state->
Figure BDA0004176570860000041
And->
Figure BDA0004176570860000042
First order characteristic states which together form the time t>
Figure BDA0004176570860000043
One level of timing enhancement feature->
Figure BDA0004176570860000044
Performing feature decoding through the primary deconvolution layer to obtain a primary feature decoding result;
and combining the primary characteristic decoding result with the foreground segmentation image to obtain a target depth prediction result at the moment t.
According to the spacecraft depth complement method based on time sequence optical images and laser radar data, a foreground segmentation module uses sparse depth image I s,t And target gray scale image I g,t As input, predicting each pixel in the cascade image by means of an encoding-decoding structure with a skip connectionAnd setting the pixel depth prediction result lower than the target threshold value to 0 to obtain the image after foreground segmentation.
According to the spacecraft depth complement method based on the time sequence optical image and the laser radar data, the simulation working conditions of the set space target three-dimensional model comprise three-dimensional model material parameter setting, texture map setting, illumination adding, earth background adding, starry sky background adding, setting of relative positions and relative postures of a target and an observation platform and setting of output nodes.
According to the spacecraft depth complementing method based on the time sequence optical image and the laser radar data, in the training process of the time sequence spacecraft depth complementing neural network model, the network weight parameter gradient is calculated by utilizing the error between the target depth prediction result and the target dense depth ground label of the image each time, so that the network parameter updating is realized.
According to the spacecraft depth complementing method based on the time sequence optical image and the laser radar data, after the network parameters of the time sequence spacecraft depth complementing neural network model are updated for set times, performance verification is carried out on the network model by using verification set sample data, and the network parameters corresponding to the optimal performance of the network model are used as the network parameters of the time sequence spacecraft depth complementing neural network model after training.
According to the spacecraft depth complement method based on the time sequence optical image and the laser radar data, a three-dimensional rendering software simulation is adopted on a space target three-dimensional model to obtain a target gray level image time sequence of a monocular visible light camera and a sparse depth image time sequence of the laser radar.
According to the spacecraft depth complementing method based on the time sequence optical image and the laser radar data, in the decoding stage, spacecraft size priori knowledge is introduced, and pixel depths, of which the deviation between a target depth prediction result and a laser radar average ranging result is larger than a preset deviation threshold value, are filtered.
The invention has the beneficial effects that: the method fully utilizes the space-time coherence contained in continuous frame data through the LSTM module in the decoding stage, thereby obtaining the depth complement result with high precision and consistent time of the space target. Firstly, collecting a space target three-dimensional structure model, and realizing simulation of sensor imaging data and automatic generation of a depth label under different working conditions and different camera parameters by utilizing three-dimensional software, thereby constructing a space target time sequence depth complement data set; then model training and parameter updating are carried out on the deep complement neural network model of the sequential spacecraft; and finally, the test data is injected into the trained time sequence spacecraft depth completion neural network model, so that the completion and completion precision evaluation of the target depth data can be realized.
The method fully excavates the target information association relation between different frame image data, improves the accurate recovery of the dense depth of the space target, and obtains the space target depth prediction result with stable time sequence; the method can solve the problem that the inter-frame complement results are inconsistent due to the fact that the existing single-frame depth complement method ignores time sequence data related information, has the advantages of being small in memory occupation, high in accuracy, high in speed, consistent in time sequence prediction result and the like, and achieves accurate recovery of the space target three-dimensional fine structure.
According to the method, a cyclic neural network is introduced into a standard decoder structure, so that the network can sense the change of target time sequence characteristics; the cyclic neural network adopts a multi-level embedded form, so that the network can adapt to working condition scenes with different change rates, and the network is more robust. The recurrent neural network can accumulate past target feature knowledge, and network prediction results are more and more accurate along with the time.
Drawings
FIG. 1 is a schematic flow chart of a spacecraft depth completion method based on time-series optical images and laser radar data;
FIG. 2 is an overall framework diagram of a time series spacecraft depth complement neural network model; in the diagram, t-2 is taken as a subscript variable, and the corresponding characteristic state at the moment t-2 is taken as a variable;
FIG. 3 is a schematic flow diagram of a feature fusion module within the encoding stage depth image feature extraction module;
fig. 4 is a block diagram of a long and short term memory network (LSTM module).
Detailed Description
The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
It should be noted that, without conflict, the embodiments of the present invention and features of the embodiments may be combined with each other.
The invention is further described below with reference to the drawings and specific examples, which are not intended to be limiting.
The invention provides a spacecraft depth complementing method based on time sequence optical images and laser radar data, which is shown in the accompanying figures 1 to 3,
collecting a plurality of space target three-dimensional models, and setting on-orbit imaging simulation working conditions of the space target three-dimensional models and sensor parameters of a monocular visible light camera and a laser radar; a target gray level image time sequence of a monocular visible light camera and a sparse depth image time sequence of a laser radar are obtained based on simulation of a plurality of space target three-dimensional models, and a target dense depth ground label of the image is generated, so that a space target time sequence depth complement data set is constructed;
constructing a time sequence spacecraft deep complement neural network model and carrying out network parameter initialization design and super-parameter setting; the network model comprises a plurality of target depth prediction branches which are cascaded in time sequence;
each target depth prediction branch includes an encoding stage and a decoding stage:
the coding stage comprises a foreground segmentation module, a gray image feature extraction module, a morphological preprocessing module and a depth image feature extraction module; the decoding stage comprises an LSTM module and a deconvolution layer; the depth image feature extraction module comprises four feature fusion modules;
the prediction process for the spatial target depth at the time t comprises the following steps:
sparse depth image I at time t s,t Preprocessing by a morphological preprocessing module to obtain a preprocessed depth image, and inputting the preprocessed depth image into a depth image feature extraction module;
target gray scale image I at time t g,t Extracting multi-scale gray scale image feature images with different semantic levels and different resolutions through a gray scale image feature extraction module;
sparse depth image I for time t s,t And target gray scale image I g,t After cascade operation, inputting the image to a foreground segmentation module for target foreground segmentation to obtain a foreground segmentation image;
the depth image feature extraction module firstly processes the preprocessed depth image through a convolution layer, and then performs feature fusion with a gray image feature image step by step according to the sequence from large scale to small scale to obtain a multi-mode data fusion feature F at the moment t t ；
Multi-mode data fusion feature F at t moment t Modeling a time sequence relation with the corresponding level characteristic state at the time t-1 step by step through an LSTM module, and performing characteristic decoding through a deconvolution layer; finally, combining the layer characteristic decoding result with the foreground segmentation image to obtain a target depth prediction result at the moment t;
training a sequential spacecraft depth complement neural network model by using a target gray image sequential sequence of a monocular visible light camera and a sparse depth image sequential sequence of a laser radar as sample images to obtain a trained sequential spacecraft depth complement neural network model;
and embedding the trained time sequence spacecraft depth complement neural network model into a satellite-borne platform, and realizing prediction of dense depth of a space target and recovery of a target three-dimensional structure based on real-time sensing data acquired by a monocular visible light camera and a laser radar.
The spatial target three-dimensional model in this embodiment may be obtained by collection purchase.
The time sequence spacecraft depth complement neural network model adopts an encoder-decoder structure as a network base structure to obtain an end-to-end trainable neural network model.
In target gray level image and sparse depth image feature coding, in order to prevent the damage of too sparse data to convolution operation, pseudo dense depth data is obtained by adopting a morphological preprocessing method and is used as input of a depth image feature extraction module.
The LSTM module in the decoding stage in this embodiment is one of Recurrent Neural Networks (RNNs), and is hierarchically embedded in a standard decoder structure to capture feature deviations between frames of feature maps of different levels, so as to fully utilize time correlation between adjacent frames to generate a dense depth map with stable time sequence.
In this embodiment, in order to filter the interference of the unrelated background on the depth completion of the spacecraft, a foreground segmentation network in a single-frame spacecraft depth completion network (SDCNet) is used to filter the sky background.
In the present embodiment, at the feature decoding stage, the LSTM module is used to fuse the multi-mode data at the time t-1 with the feature F t And the characteristic state s at the current moment t-1 Modeling time sequence relation and outputting time sequence enhancement characteristics
Figure BDA0004176570860000071
Timing enhancement feature->
Figure BDA0004176570860000072
Feature decoding is performed by the deconvolution layer.
Still further, as shown in fig. 2, the target gradation image I in the present embodiment g,t Extracting gray image feature images with five sizes respectively as target gray images I by a gray image feature extraction module g,t 1/2, 1/4, 1/8, 1/16 and 1/32.
Further, referring to fig. 2, the depth image feature extraction module obtains a multi-mode data fusion feature F at time t t The process of (1) comprises:
the depth image feature extraction module firstly processes the preprocessed depth image through a convolution layer, and then fuses the preprocessed depth image with the 1/2 gray image feature image after being processed by the primary residual error module by the primary feature fusion module to obtain a depth imageA stage data fusion feature; the primary data fusion features are processed by a secondary residual error module and then fused with the 1/4 gray image feature map by a secondary feature fusion module to obtain secondary data fusion features; the secondary data fusion features are processed by the tertiary residual error module and then fused with the 1/8 gray level image feature map by the tertiary feature fusion module to obtain tertiary data fusion features; the three-level data fusion features are processed by a four-level residual error module and then fused with the 1/16 gray image feature map by a four-level feature fusion module to obtain four-level data fusion features; the four-level data fusion feature is processed by a five-level residual error module and then is subjected to element addition with the 1/32 gray level image feature map to obtain the five-level data fusion feature which is used as a t-moment multi-mode data fusion feature F t 。
Still further, as shown in fig. 2 and 3, the method for obtaining the target depth prediction result at the time t in the decoding stage includes:
fusion of t-moment multi-mode data to feature F t Five-stage characteristic state at time t-1 through five-stage LSTM module
Figure BDA0004176570860000073
Modeling the time sequence relationship to obtain five-level time sequence enhancement characteristic +.>
Figure BDA0004176570860000074
And five-level characteristic memory state->
Figure BDA0004176570860000075
And->
Figure BDA0004176570860000076
Five-level characteristic state of common composition t time>
Figure BDA0004176570860000077
Five-level timing enhancement feature->
Figure BDA0004176570860000078
Performing feature decoding through a five-stage deconvolution layer to obtain a five-stage feature decoding result; passing the five-level feature decoding result through four levelsLSTM module and four-level characteristic state at t-1 moment
Figure BDA0004176570860000079
Modeling the time sequence relationship to obtain four-level time sequence enhancement characteristic ++at the time t>
Figure BDA00041765708600000710
And four-level characteristic memory state->
Figure BDA00041765708600000711
And->
Figure BDA00041765708600000712
Four-level characteristic states which together form the time t>
Figure BDA00041765708600000713
Four-level timing enhancement feature->
Figure BDA00041765708600000714
Performing feature decoding through a four-level deconvolution layer to obtain a four-level feature decoding result; the four-level characteristic decoding result passes through a three-level LSTM module and the three-level characteristic state at the time t-1>
Figure BDA0004176570860000081
Modeling the time sequence relationship to obtain three-level time sequence enhancement characteristic +.>
Figure BDA0004176570860000082
And three-level characteristic memory state->
Figure BDA0004176570860000083
Figure BDA0004176570860000084
And (3) with
Figure BDA0004176570860000085
Three-level characteristic state of common composition t time>
Figure BDA0004176570860000086
Three-level timing enhancement feature->
Figure BDA0004176570860000087
Performing feature decoding through the three-level deconvolution layer to obtain a three-level feature decoding result; the third-level characteristic decoding result passes through a second-level LSTM module and a second-level characteristic state at the time t-1>
Figure BDA0004176570860000088
Modeling the time sequence relationship to obtain the secondary time sequence enhancement characteristic +.>
Figure BDA0004176570860000089
And a secondary feature memory state
Figure BDA00041765708600000810
And->
Figure BDA00041765708600000811
Second-order characteristic states which together form the time t>
Figure BDA00041765708600000812
Second order timing enhancement feature->
Figure BDA00041765708600000813
Performing feature decoding through a secondary deconvolution layer to obtain a secondary feature decoding result; the secondary characteristic decoding result passes through a primary LSTM module and the primary characteristic state at the time t-1>
Figure BDA00041765708600000814
Modeling the time sequence relationship to obtain the first-level time sequence enhancement characteristic +.>
Figure BDA00041765708600000815
And the first-level characteristic memory state->
Figure BDA00041765708600000816
And->
Figure BDA00041765708600000817
First order characteristic states which together form the time t>
Figure BDA00041765708600000818
One level of timing enhancement feature->
Figure BDA00041765708600000819
Performing feature decoding through the primary deconvolution layer to obtain a primary feature decoding result;
and combining the primary characteristic decoding result with the foreground segmentation image to obtain a target depth prediction result at the moment t.
In the present embodiment, the foreground segmentation module uses sparse depth image I s,t And target gray scale image I g,t The cascade image of (2) is used as input, the probability that each pixel in the cascade image belongs to a target is predicted through the coding-decoding structure with jump connection, and the pixel depth prediction result lower than the target threshold value is set to 0, so that the image after foreground segmentation is obtained.
As an example, the simulation conditions for setting the three-dimensional model of the spatial target include three-dimensional model material parameter setting, texture map setting, illumination adding, earth background adding, starry sky background adding, setting of relative positions and relative postures of the target and the observation platform, and setting of output nodes.
Still further, in the training process of the time sequence spacecraft depth complement neural network model, the network weight parameter gradient is calculated by utilizing the error between the target depth prediction result and the target dense depth ground label of the image each time, so that the network parameter updating is realized.
In this embodiment, after the network parameters of the time-series spacecraft depth-complement neural network model are updated for a set number of times, performance verification is performed on the network model by using verification set sample data, and the network parameters corresponding to the network model with optimal performance are stored as the network parameters of the trained time-series spacecraft depth-complement neural network model.
And taking the saved network parameters as final network weights, loading test data to the network to obtain a deep complement result, and comparing the deep complement result with a ground label to evaluate network accuracy.
As an example, a target gray image time sequence of a monocular visible light camera and a sparse depth image time sequence of a laser radar are obtained by adopting three-dimensional rendering software simulation to a space target three-dimensional model.
In the embodiment, in the decoding stage, spacecraft size priori knowledge is introduced, and pixel depths, where the deviation between a target depth prediction result and a laser radar average ranging result is greater than a preset deviation threshold, are filtered, so that a high-quality depth prediction result is obtained.
The following describes the working process of the deep complement neural network model of the time sequence spacecraft in detail:
referring to fig. 2, a target gray-scale image I is inputted at time t g,t And sparse depth image I s,t Different sensor data features are aggregated by adopting a multi-source deep image feature extraction module, so that a Gao Biaozheng-force multi-mode data fusion feature F is obtained t Can be expressed as:
F t ＝F enc o der ([I g,t ,I s,t ],θ enc o der )，
f in the formula encoder For encoder feature extraction function, θ encoder Representing the network parameters that the encoder structure needs to learn.
In combination with the illustration of fig. 3, the feature fusion module in the encoding stage realizes the fusion of the target gray image features and the depth image features based on the attention mechanism, and provides features with high characterization capability for the subsequent target depth decoding.
The feature fusion module mainly comprises a feature embedding layer, a cross channel fusion layer and a space attention layer; the following is a detailed description through the primary feature fusion module, and the corresponding operation is performed on the primary feature added with the upper right corner mark 1:
the primary characteristic fusion module uses the primary characteristic of the gray image
Figure BDA0004176570860000091
And depth image oneStage feature
Figure BDA0004176570860000092
(C represents the number of characteristic channels, H and W represent the length and width of the characteristic map respectively) as input, and obtaining primary data fusion characteristics>
Figure BDA0004176570860000093
The feature embedding layer encodes the feature graphs of different channels to generate corresponding feature vectors; the characteristic embedding layer respectively characterizes gray scale
Figure BDA0004176570860000094
And depth feature->
Figure BDA00041765708600000911
The method comprises the steps of decomposing the region into M mutually non-overlapping regions (each feature block is of a size of S multiplied by S), and extracting the region features by using a depth separable convolution operation with a convolution kernel of the size of S multiplied by S and a convolution step of S. In addition, pair->
Figure BDA0004176570860000095
And->
Figure BDA00041765708600000912
The regional blocks of the (2) are subjected to maximum pooling operation and average pooling operation to extract regional global features, and finally three features are cascaded to output gray level image primary embedded features +.>
Figure BDA0004176570860000096
First-level embedding of features with depth image->
Figure BDA0004176570860000097
Wherein d is k ＝3×H×W/S 2 。
The cross channel attention layer takes the characteristic encoding result of the characteristic embedding layer as input, firstly adopts linear transformation to respectively calculate gray image primary embedded characteristic query vectors and depth image primary embedded characteristic key vectors of n attention heads, and can be specifically expressed as:
Figure BDA0004176570860000098
in the middle of
Figure BDA0004176570860000099
Linear mapping weight matrix for computing query vector and key vector, respectively, < >>
Figure BDA00041765708600000910
Respectively embedding a feature query vector and a depth image embedded feature key vector for the gray level image of the ith attention head; n is the number of attention headers.
The gray image feature and depth image feature channel association weight matrix of the ith attention head is further calculated by using the scaling dot product attention, and can be specifically expressed as:
Figure BDA0004176570860000101
w in i The weights are associated for the feature channels of the ith attention head, softmax (·) is a normalization function.
Finally, according to the association weight matrix, the gray image primary characteristic is carried out
Figure BDA0004176570860000102
Primary feature of depth image F s 1 Channel fusion is performed. Specifically, the method can be expressed as:
Figure BDA0004176570860000103
h in g 、h s Respectively one-level characteristics of gray-scale images
Figure BDA0004176570860000104
First-level features of depth image->
Figure BDA00041765708600001015
Feature vectors spread in rows, w i For the feature channel associated weight of the ith attention head, reshape (·) is a vector dimension transform operation, +.>
Figure BDA0004176570860000105
The first channel fusion feature for the ith attention head.
Finally, cascading the channel fusion characteristics obtained by calculation of n attention heads, and further realizing multi-head attention characteristic fusion by using convolution operation to obtain primary channel fusion characteristics
Figure BDA0004176570860000106
Specifically, the method can be expressed as:
Figure BDA0004176570860000107
wherein Conv (·) is a convolution operation, [ ·; and is a characteristic cascade operation.
Spatial attention layer with primary channel fusion features
Figure BDA0004176570860000108
First-order features of gray-scale images->
Figure BDA0004176570860000109
For input, features at different spatial positions are characterized by means of channel average pooling and channel maximum pooling operations, and a feature cascade is input to a convolution layer to obtain first-level spatial attention weights. Specifically, the method can be expressed as:
Figure BDA00041765708600001010
in the middle of
Figure BDA00041765708600001011
Is the first-level spatial attention weight; max (max) c (·)、avg c (. Cndot.) is the channel maximum pooling and average pooling operations, respectively; sigma (·) is a Sigmoid normalization function; conv (-) and [; carrying out]A convolution operation and a feature concatenation operation, respectively.
Based on the first-order spatial attention weight, the features at different spatial positions are weighted and summed to obtain a first-order fusion feature
Figure BDA00041765708600001012
Specifically, the method can be expressed as:
Figure BDA00041765708600001013
where +..
Still further, in the feature encoding stage, four levels of fused features
Figure BDA00041765708600001014
The five-level residual error module is used for processing and adding the five-level residual error module with 1/32 gray image characteristic image elements to obtain five-level data fusion characteristics serving as t-moment multi-mode data fusion characteristics F t 。
In the decoding stage, a long and short-term memory network (LSTM module) is introduced before each deconvolution operation, as shown in fig. 4. The LSTM module is mainly composed of an input gate i t Forgetting door f t Output gate o t The flow direction of different frame information is controlled, the internal data processing process is described in detail by taking a five-stage LSTM module as an example, and the corresponding variable is added with an upper right corner mark 5 to correspond to the following:
the same-level characteristic memory state of one frame above five-level LSTM module
Figure BDA0004176570860000111
The same level time sequence enhancement feature of the previous frame>
Figure BDA0004176570860000112
Current frame fusion feature F t For input, output the characteristic memory state of the current frame +.>
Figure BDA0004176570860000113
And current frame timing enhancement features
Figure BDA0004176570860000114
Figure BDA0004176570860000115
And->
Figure BDA0004176570860000116
Characteristic states which together form the current frame +.>
Figure BDA0004176570860000117
Timing enhancement feature->
Figure BDA0004176570860000118
As input to a subsequent deconvolution operation.
Figure BDA0004176570860000119
And->
Figure BDA00041765708600001110
The calculation process of (1) can be expressed as:
Figure BDA00041765708600001111
Figure BDA00041765708600001112
Figure BDA00041765708600001113
Figure BDA00041765708600001114
Figure BDA00041765708600001115
wherein [. Cndot.; carrying out]Characterizing the feature cascade operation, # denotes a convolution operation, +. f 、W i 、W o 、W c 、b f 、b i 、b o 、b c And memorizing the parameters to be learned of the network for a long time. Conv in FIG. 4 represents a convolution operation, and tanh represents a hyperbolic tangent function.
In summary, the method of the invention introduces a cyclic neural network into the decoder structure, so that the network can mine the target time sequence characteristic change; the method comprises the steps of embedding a circulating neural network into the deconvolution operation of different levels in a hierarchical manner by adopting a multi-scale mechanism, so that the network can adapt to target characteristic changes under different movement speeds and different movement modes; and finally, the network prediction result is more stable through the time correlation of the depth mining time sequence data, and the depth complement precision can be continuously improved along with the time.
Although the invention herein has been described with reference to particular embodiments, it is to be understood that these embodiments are merely illustrative of the principles and applications of the present invention. It is therefore to be understood that numerous modifications may be made to the illustrative embodiments and that other arrangements may be devised without departing from the spirit and scope of the present invention as defined by the appended claims. It should be understood that the different dependent claims and the features described herein may be combined in ways other than as described in the original claims. It is also to be understood that features described in connection with separate embodiments may be used in other described embodiments.

Claims (10)

1. A spacecraft depth complementing method based on time sequence optical images and laser radar data is characterized by comprising the following steps of,
collecting a plurality of space target three-dimensional models, and setting simulation working conditions of the space target three-dimensional models and sensor parameters of a monocular visible light camera and a laser radar; obtaining a target gray level image time sequence of a monocular visible light camera and a sparse depth image time sequence of a laser radar based on simulation of a plurality of space target three-dimensional models, and generating a target dense depth ground label of the image;
constructing a time sequence spacecraft depth complement neural network model: comprising a plurality of target depth prediction branches cascaded in chronological order;
each target depth prediction branch includes an encoding stage and a decoding stage:
the coding stage comprises a foreground segmentation module, a gray image feature extraction module, a morphological preprocessing module and a depth image feature extraction module; the decoding stage comprises an LSTM module and a deconvolution layer;
the prediction process for the spatial target depth at the time t comprises the following steps:
sparse depth image I at time t s,t Preprocessing by a morphological preprocessing module to obtain a preprocessed depth image, and inputting the preprocessed depth image into a depth image feature extraction module;
target gray scale image I at time t g,t Extracting multi-scale gray scale image feature images with different semantic levels and different resolutions through a gray scale image feature extraction module;
sparse depth image I for time t s,t And target gray scale image I g,t After cascade operation, inputting the image to a foreground segmentation module for target foreground segmentation to obtain a foreground segmentation image;
the depth image feature extraction module firstly processes the preprocessed depth image through a convolution layer, and then performs feature fusion with a gray image feature image step by step according to the sequence from large scale to small scale to obtain a multi-mode data fusion feature F at the moment t t ；
Multi-mode data fusion feature F at t moment t Modeling a time sequence relation with the corresponding level characteristic state at the time t-1 step by step through an LSTM module, and performing characteristic decoding through a deconvolution layer; finally, combining the layer characteristic decoding result with the foreground segmentation image to obtain a target depth prediction result at the moment t;
training a sequential spacecraft depth complement neural network model by using a target gray image sequential sequence of a monocular visible light camera and a sparse depth image sequential sequence of a laser radar as sample images to obtain a trained sequential spacecraft depth complement neural network model;
and embedding the trained time sequence spacecraft depth complement neural network model into a satellite-borne platform, and realizing prediction of dense depth of the space target based on real-time sensing data acquired by the monocular visible light camera and the laser radar.
2. The spacecraft depth complementing method based on time-series optical images and laser radar data according to claim 1, wherein the target gray scale image I g,t Extracting gray image feature images with five sizes respectively as target gray images I by a gray image feature extraction module g,t 1/2, 1/4, 1/8, 1/16 and 1/32.
3. The spacecraft depth filling method based on time-series optical images and lidar data according to claim 2, wherein,
the depth image feature extraction module obtains a t-moment multi-mode data fusion feature F t The process of (1) comprises:
the depth image feature extraction module firstly processes the preprocessed depth image through a convolution layer, and then the preprocessed depth image is processed through the primary residual error module and fused with the 1/2 gray image feature image through the primary feature fusion module to obtain primary data fusion features; the primary data fusion features are processed by a secondary residual error module and then fused with the 1/4 gray image feature map by a secondary feature fusion module to obtain secondary data fusion features; the secondary data fusion features are processed by the tertiary residual error module and then fused with the 1/8 gray level image feature map by the tertiary feature fusion module to obtain tertiary data fusion features; the three-level data fusion features are processed by a four-level residual error module and then fused with the 1/16 gray image feature map by a four-level feature fusion module to obtain four-level data fusion features; the four-level data fusion characteristic is processed by a five-level residual error module and then is subjected to element addition with the 1/32 gray level image characteristic image to obtain the five-level data fusion characteristic which is used as a t-time multi-mode data fusion characteristicSign F t 。
4. The spacecraft depth filling method based on time-series optical images and lidar data according to claim 3,
the method for obtaining the target depth prediction result at the time t in the decoding stage comprises the following steps:
fusion of t-moment multi-mode data to feature F t Five-stage characteristic state at time t-1 through five-stage LSTM module
Figure FDA0004176570850000021
Modeling the time sequence relationship to obtain five-level time sequence enhancement characteristic +.>
Figure FDA0004176570850000022
And five-level characteristic memory state->
Figure FDA0004176570850000023
Figure FDA0004176570850000024
And->
Figure FDA0004176570850000025
Five-level characteristic state of common composition t time>
Figure FDA0004176570850000026
Five-level timing enhancement feature->
Figure FDA0004176570850000027
Performing feature decoding through a five-stage deconvolution layer to obtain a five-stage feature decoding result; the decoding result of the five-level characteristic is processed by a four-level LSTM module and the four-level characteristic state at the time t-1>
Figure FDA0004176570850000028
Modeling the time sequence relationship to obtain four-level time sequence enhancement characteristic ++at the time t>
Figure FDA0004176570850000029
And four-level characteristic memory state->
Figure FDA00041765708500000210
Figure FDA00041765708500000211
And->
Figure FDA00041765708500000212
Four-level characteristic states which together form the time t>
Figure FDA00041765708500000213
Four-level timing enhancement feature->
Figure FDA00041765708500000214
Performing feature decoding through a four-level deconvolution layer to obtain a four-level feature decoding result; the four-level characteristic decoding result passes through a three-level LSTM module and the three-level characteristic state at the time t-1>
Figure FDA00041765708500000215
Modeling the time sequence relationship to obtain three-level time sequence enhancement characteristic +.>
Figure FDA00041765708500000216
And three-level characteristic memory state->
Figure FDA00041765708500000217
Figure FDA00041765708500000218
And->
Figure FDA00041765708500000219
Three-level characteristic state of common composition t time>
Figure FDA00041765708500000220
Three-level timing enhancement feature->
Figure FDA00041765708500000221
Performing feature decoding through the three-level deconvolution layer to obtain a three-level feature decoding result; the third-level characteristic decoding result passes through a second-level LSTM module and a second-level characteristic state at the time t-1>
Figure FDA0004176570850000031
Modeling the time sequence relationship to obtain the secondary time sequence enhancement characteristic +.>
Figure FDA0004176570850000032
And secondary characteristic memory state->
Figure FDA0004176570850000033
Figure FDA0004176570850000034
And->
Figure FDA0004176570850000035
Second-order characteristic states which together form the time t>
Figure FDA0004176570850000036
Second order timing enhancement feature->
Figure FDA0004176570850000037
Performing feature decoding through a secondary deconvolution layer to obtain a secondary feature decoding result; the secondary characteristic decoding result passes through a primary LSTM module and the primary characteristic state at the time t-1>
Figure FDA0004176570850000038
Modeling the time sequence relationship to obtain the first-level time sequence enhancement characteristic +.>
Figure FDA0004176570850000039
And the first-level characteristic memory state->
Figure FDA00041765708500000310
Figure FDA00041765708500000311
And (3) with
Figure FDA00041765708500000312
First order characteristic states which together form the time t>
Figure FDA00041765708500000313
One level of timing enhancement feature->
Figure FDA00041765708500000314
Performing feature decoding through the primary deconvolution layer to obtain a primary feature decoding result;
and combining the primary characteristic decoding result with the foreground segmentation image to obtain a target depth prediction result at the moment t.
5. The spacecraft depth filling method based on time sequence optical images and laser radar data according to claim 4, wherein the foreground segmentation module uses sparse depth image I s,t And target gray scale image I g,t The cascade image of (2) is used as input, the probability that each pixel in the cascade image belongs to a target is predicted through the coding-decoding structure with jump connection, and the pixel depth prediction result lower than the target threshold value is set to 0, so that the image after foreground segmentation is obtained.
6. The spacecraft depth filling method based on time-series optical images and lidar data according to claim 1, wherein,
the simulation working conditions of the three-dimensional model of the set space target comprise three-dimensional model material parameter setting, texture map setting, illumination adding, earth background adding, starry sky background adding, and setting of relative positions and relative postures of the target and the observation platform and output node setting.
7. The spacecraft depth completing method based on time sequence optical images and laser radar data according to claim 1, wherein in the training process of a time sequence spacecraft depth completing neural network model, a network weight parameter gradient is calculated by utilizing an error between each target depth predicting result and a target dense depth ground label of an image, and network parameter updating is achieved.
8. The spacecraft depth complementing method based on the time-series optical image and the laser radar data, according to claim 7, is characterized in that after the network parameters of the time-series spacecraft depth complementing neural network model are updated for a set number of times, performance verification is carried out on the network model by using verification set sample data, and the network parameters corresponding to the optimal performance of the network model are used as the network parameters of the trained time-series spacecraft depth complementing neural network model.
9. The spacecraft depth filling method based on time-series optical images and lidar data according to claim 1, wherein,
and simulating the space target three-dimensional model by adopting three-dimensional rendering software to obtain a target gray level image time sequence of the monocular visible light camera and a sparse depth image time sequence of the laser radar.
10. The spacecraft depth filling method based on time-series optical images and lidar data according to claim 1, wherein,
and in the decoding stage, introducing priori knowledge of the spacecraft size, and filtering out pixel depths, wherein the deviation between the target depth prediction result and the laser radar average ranging result is larger than a preset deviation threshold value.
CN202310393175.2A 2023-04-13 2023-04-13 Spacecraft depth complementing method based on time sequence optical image and laser radar data Pending CN116402874A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310393175.2A CN116402874A (en) 2023-04-13 2023-04-13 Spacecraft depth complementing method based on time sequence optical image and laser radar data

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310393175.2A CN116402874A (en) 2023-04-13 2023-04-13 Spacecraft depth complementing method based on time sequence optical image and laser radar data

Publications (1)

Publication Number Publication Date
CN116402874A true CN116402874A (en) 2023-07-07

Family

ID=87019671

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310393175.2A Pending CN116402874A (en) 2023-04-13 2023-04-13 Spacecraft depth complementing method based on time sequence optical image and laser radar data

Country Status (1)

Country Link
CN (1) CN116402874A (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117009750A (en) * 2023-09-28 2023-11-07 北京宝隆泓瑞科技有限公司 Methane concentration data complement method and device for machine learning

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117009750A (en) * 2023-09-28 2023-11-07 北京宝隆泓瑞科技有限公司 Methane concentration data complement method and device for machine learning
CN117009750B (en) * 2023-09-28 2024-01-02 北京宝隆泓瑞科技有限公司 Methane concentration data complement method and device for machine learning

Similar Documents

Publication Publication Date Title
CN110210551B (en) Visual target tracking method based on adaptive subject sensitivity
Gu et al. DenseLiDAR: A real-time pseudo dense depth guided depth completion network
CN110443842B (en) Depth map prediction method based on visual angle fusion
Wang et al. Fadnet: A fast and accurate network for disparity estimation
CN112529015B (en) Three-dimensional point cloud processing method, device and equipment based on geometric unwrapping
Mahjourian et al. Geometry-based next frame prediction from monocular video
Bešić et al. Dynamic object removal and spatio-temporal RGB-D inpainting via geometry-aware adversarial learning
Maslov et al. Online supervised attention-based recurrent depth estimation from monocular video
CN116797787B (en) Remote sensing image semantic segmentation method based on cross-modal fusion and graph neural network
JP2024513596A (en) Image processing method and apparatus and computer readable storage medium
Xiong et al. Contextual Sa-attention convolutional LSTM for precipitation nowcasting: A spatiotemporal sequence forecasting view
CN110852199A (en) Foreground extraction method based on double-frame coding and decoding model
CN116612468A (en) Three-dimensional target detection method based on multi-mode fusion and depth attention mechanism
CN116485867A (en) Structured scene depth estimation method for automatic driving
CN116402874A (en) Spacecraft depth complementing method based on time sequence optical image and laser radar data
CN116205962A (en) Monocular depth estimation method and system based on complete context information
Knöbelreiter et al. Self-supervised learning for stereo reconstruction on aerial images
Du et al. Srh-net: Stacked recurrent hourglass network for stereo matching
CN115511759A (en) Point cloud image depth completion method based on cascade feature interaction
CN117593702B (en) Remote monitoring method, device, equipment and storage medium
CN115131418A (en) Monocular depth estimation algorithm based on Transformer
Neupane et al. Building footprint segmentation using transfer learning: a case study of the city of melbourne
CN117710429A (en) Improved lightweight monocular depth estimation method integrating CNN and transducer
CN112950786A (en) Vehicle three-dimensional reconstruction method based on neural network
CN110532868B (en) Method for predicting free space semantic boundary

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination