CN108171217A - A kind of three-dimension object detection method based on converged network - Google Patents
A kind of three-dimension object detection method based on converged network Download PDFInfo
- Publication number
- CN108171217A CN108171217A CN201810081797.0A CN201810081797A CN108171217A CN 108171217 A CN108171217 A CN 108171217A CN 201810081797 A CN201810081797 A CN 201810081797A CN 108171217 A CN108171217 A CN 108171217A
- Authority
- CN
- China
- Prior art keywords
- point
- network
- converged network
- frame
- input
- 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.)
- Withdrawn
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/60—Type of objects
- G06V20/64—Three-dimensional objects
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
A kind of three-dimension object detection method based on converged network proposed in the present invention, main contents include:Point cloud network, converged network, intensive fusion forecasting score function, its process is, point cloud network model intake original point cloud, learn the space encoding each put and polymerization global point cloud feature, by these features for classification and semantic segmentation, converged network is using the corresponding points cloud feature generated using the sub-network of the characteristics of image of convolutional neural networks extraction and point converged network as input, these combination of function are got up and are that target object exports a three-dimensional boundaries frame by it, with the supervision direct training network of score function, whether future position is in object boundary frame, and unsupervised score function can help network to select optimum prediction point.The present invention can directly learn most optimally to combine image and depth information, avoid quantization or projection etc. and damage input pretreatment, have general applicability, and its accuracy also greatly improves.
Description
Technical field
The present invention relates to object detecting areas, more particularly, to a kind of three-dimension object detection side based on converged network
Method.
Background technology
The detection and identification of three-dimension object are the important research directions of computer vision field, and it is a kind of general to put fusion
Three-dimension object detection method, image and three-dimensional point cloud information can be utilized simultaneously.Three dimensional point cloud industrially, especially
Application in reverse-engineering is more and more universal.The main application scenarios of three-dimension object detection and identification exist on point cloud data
In identifying existing three-dimensional object model from the point cloud data obtained.Three-dimension object detection technique can be applied various each
Urban planning, construction and the management project of sample, for detecting the pedestrian in City scenarios, vehicle, shop etc.;It can also be by certainly
The auxiliary drivers such as dynamic detection pedestrian, vehicle, traffic sign and road conditions drive vehicle, it might even be possible to applied to following nothing
The fields such as people's autonomous driving technology and mobile robot.In addition to this, in electronic communication monitoring, industrial detection automation, military affairs
The every field such as investigation and Medical Instruments also have extensive application demand.And existing three-dimension object detection technique is often only applicable to
The object of such as automobile etc, but it is not suitable for the detection of other critical objects such as pedestrian or cycling, therefore without general
Time applicability, it is difficult to be applied in practice.
The present invention proposes a kind of three-dimension object detection method based on converged network, and the model intake of point cloud network is original
Point cloud learns the space encoding each put and polymerization global point cloud feature, by these features for classifying and semantic segmentation, fusion
Network makees the corresponding points cloud feature generated using the sub-network of the characteristics of image of convolutional neural networks extraction and point converged network
For input, these combination of function are got up and are that target object exports a three-dimensional boundaries frame by it, direct with supervision score function
Training network, whether future position is in object boundary frame, and unsupervised score function can help network to select optimum prediction point.This
Invention can directly learn most optimally to combine image and depth information, avoid quantization or projection etc. and damage input in advance
Processing has general applicability, and its accuracy also greatly improves.
Invention content
Do not have the problem of universality for existing method, the purpose of the present invention is to provide one kind based on point converged network
Three-dimension object detection method, point cloud network model intake original point cloud learns the space encoding each put and polymerization global point
Cloud feature, by these features for classification and semantic segmentation, converged network will be using the characteristics of image of convolutional neural networks extraction
The corresponding points cloud feature generated with the sub-network of converged network as input, got up and be target pair by it by these combination of function
As exporting a three-dimensional boundaries frame, with the supervision direct training network of score function, whether future position is in object boundary frame, and nothing
Supervision score function can help network to select optimum prediction point.
To solve the above problems, the present invention provides a kind of three-dimension object detection method based on converged network, it is main
Content includes:
(1) cloud network is put;
(2) converged network;
(3) intensive fusion forecasting score function.
Wherein, point fusion, there are three chief components:The point converged network variant of extraction point cloud feature carries
Take convolutional neural networks (CNN), two features of combination and the converged network for exporting three-dimensional boundaries frame of picture appearance feature.
Wherein, the point cloud network realizes the constant of unordered 3D point cloud collection using symmetric function (maximum pond) first
Property processing;The model absorbs original point cloud, and learns the global point cloud feature for the space encoding and polymerization each put, by these
Feature is for classification and semantic segmentation;
Point converged network can directly handle original point, without the operation that damages of voxelization or projection, and with input
The point converged network formula that the quantity of point is linear but original cannot be used for 3D recurrence, it is therefore desirable to carry out batch and return
One changes and inputs normalization.
Further, it is described that batch is gone to normalize, in original point converged network is realized, all full articulamentums all with
A batch normalization layer;But batch normalization hampers the estimation performance of three-dimensional boundaries frame;Batch normalization is intended to eliminate
Scale and deviation in input data, but task is returned for 3D, the absolute figure for putting position is helpful;Therefore, point melts
It closes network variations and deletes all batch normalization layers.
Further, input normalization obtains image by searching for all the points in projectable to frame in scene
The corresponding 3D point cloud of bounding box;However, the spatial position of 3D points is related to 2D frame position heights, this can introduce deviation;Point fusion
Network application space transformer network (STN) carrys out the specification input space;But STN cannot correct these deviations completely, therefore use instead
Known geometry camera calculates specification spin matrix Rc;RcThe z-axis of camera frame will be rotated to by the light at 2D frames center.
Wherein, the converged network, by the characteristics of image extracted using standard CNN and the sub-network of point converged network
The corresponding points cloud feature of generation is as input;These combination of function are got up and are that target object exports a 3D bounding box by it;
Converged network includes global converged network and novel intensive converged network.
Further, the global converged network handles, and directly to object boundary image and point cloud feature
The three-dimensional position in eight corners of frame is returned;The loss function of global converged network is:
Wherein,It is the corner location of true frame demarcated, xiBe prediction corner location, LstnIt is the space introduced
Regularization loss is converted, for the orthogonality of mandatory learning space conversion matrices;A but major defect of global converged network
It is regressive objectVariance directly depend on concrete condition.
Further, the intensive converged network, the main thought of intensive converged network model are three using input
Dimension point returns the absolute position of the corner location of 3D bounding boxes as intensive space anchor point rather than directly, for each defeated
The three-dimensional point entered is predicted from this to the spatial deviation of the corner location of neighbouring bounding box;Use a converged network variant
To export point-by-point feature;For each point, point converged network variant is connect with global point converged network feature and characteristics of image,
Generate the input tensor of n × 3136;Intensive converged network handles the input using multiple layers, and it is pre- to export 3D bounding boxes
The score surveyed and each put;In the testing time, the prediction with top score is selected as finally predicting;Intensive converged network
Loss function be:
Wherein, N is the number of input point,Be the true frame demarcated corner location and i-th input point it
Between offset,Be prediction offset, LscoreIt is the loss of score function.
Wherein, the intensive fusion forecasting score function, LscoreThe target of function is by close-target by Web Cams
Learn spatial deviation on the point of frame;Specifically, score function includes:
(1) score function is supervised:Direct training network, to predict a point whether in object boundary frame;
(2) unsupervised score function:Network selection is allowed to lead to the point of optimum prediction.
Further, the supervision score function and unsupervised score function, supervision scoring loss training network prediction
Whether one point is in target frame;The offset of point i is returned loss to be expressed asI-th point of binary classification is lost into table
It is shown asThen:
Wherein, mi∈ { 0,1 } indicates at i-th point whether in object boundary frame, LscoreIt is to intersect entropy loss, punishment is closed
In set point whether the incorrect prediction in frame;As defined, the supervision score function by Web Cams in study, with pre-
Survey the spatial deviation of the point in object boundary frame;However, it may not provide optimum, because the point in frame may not be
Point with optimum prediction;
The target of unsupervised scoring is that network is allowed to be directly acquainted with which point may provide best hypothesis;Need to network into
Row training, to determine that there may be the high confidence levels well predicted;The formula includes two loss conditions vied each other:Selection
The high confidence level c of all the pointsi, however, the prediction error of corner location is directly proportional to this confidence level;DefinitionCollection is combined into
The inflection point offset of point i returns loss;Then loss becomes:
Wherein, w is the weight factor between two;Best w is found by rule of thumb, and enables w=in all experiments
0.1。
Description of the drawings
Fig. 1 is a kind of system framework figure of the three-dimension object detection method based on converged network of the present invention.
Fig. 2 is a kind of point converged network system figure of the three-dimension object detection method based on converged network of the present invention.
Fig. 3 is a kind of input normalization of three-dimension object detection method based on converged network of the present invention.
Specific embodiment
It should be noted that in the absence of conflict, the feature in embodiment and embodiment in the application can phase
It mutually combines, the present invention is described in further detail in the following with reference to the drawings and specific embodiments.
Fig. 1 is a kind of system framework figure of the three-dimension object detection method based on converged network of the present invention.Mainly include
Point cloud network, converged network, intensive fusion forecasting score function.
There are three chief components for point fusion:The point converged network variant of extraction point cloud feature, extraction picture appearance are special
Convolutional neural networks (CNN), two features of combination and the converged network for exporting three-dimensional boundaries frame of sign.
Point converged network realizes that the invariance of unordered 3D point cloud collection is handled using symmetric function (maximum pond) first;The mould
Type absorbs original point cloud, and learns the global point cloud feature for the space encoding and polymerization each put, these features are used to divide
Class and semantic segmentation;
Point converged network can directly handle original point, without the operation that damages of voxelization or projection, and with input
The point converged network formula that the quantity of point is linear but original cannot be used for 3D recurrence, it is therefore desirable to carry out batch and return
One changes and inputs normalization.
Batch is gone to normalize, in original point converged network is realized, all full articulamentums all follow a batch normalizing
Change layer;But batch normalization hampers the estimation performance of three-dimensional boundaries frame;Batch normalization is intended to eliminate the ruler in input data
Degree and deviation, but task is returned for 3D, the absolute figure for putting position is helpful;Therefore, point converged network variant is deleted
All batch normalization layers.
The corresponding points that converged network generates the sub-network of the characteristics of image extracted using standard CNN and point converged network
Cloud feature is as input;These combination of function are got up and are that target object exports a 3D bounding box by it;Converged network includes
Global converged network and novel intensive converged network.
Intensive fusion forecasting score function, LscoreThe target of function is to learn Web Cams on the point by close-target frame
Spatial deviation;Specifically, score function includes:
(1) score function is supervised:Direct training network, to predict a point whether in object boundary frame;
(2) unsupervised score function:Network selection is allowed to lead to the point of optimum prediction.
Whether supervision scoring loss training network predicts a point in target frame;The offset of point i is returned loss to represent
ForI-th point of binary classification loss is expressed asThen:
Wherein, mi∈ { 0,1 } indicates at i-th point whether in object boundary frame, LscoreIt is to intersect entropy loss, punishment is closed
In set point whether the incorrect prediction in frame;As defined, the supervision score function by Web Cams in study, with pre-
Survey the spatial deviation of the point in object boundary frame;However, it may not provide optimum, because the point in frame may not be
Point with optimum prediction;
The target of unsupervised scoring is that network is allowed to be directly acquainted with which point may provide best hypothesis;Need to network into
Row training, to determine that there may be the high confidence levels well predicted;The formula includes two loss conditions vied each other:Selection
The high confidence level c of all the pointsi, however, the prediction error of corner location is directly proportional to this confidence level;DefinitionCollection is combined into
The inflection point offset of point i returns loss;Then loss becomes:
Wherein, w is the weight factor between two;Best w is found by rule of thumb, and enables w=in all experiments
0.1。
Fig. 2 is a kind of point converged network system figure of the three-dimension object detection method based on converged network of the present invention.Its
In, Fig. 2 (A) represents the point converged network variant of extraction point cloud feature, for handling original point cloud data;Fig. 2 (B) expressions are used for
Extract the convolutional neural networks of picture appearance feature;Fig. 2 (C) represents intensive converged network;Fig. 2 (D) represents global converged network;
Fig. 2 (E) represents final prediction result.
Wherein, the main thought of intensive converged network model is to use the three-dimensional point of input as intensive space anchor point,
Rather than directly return the absolute position of the corner location of 3D bounding boxes, for the three-dimensional point of each input, predict from the point to
The spatial deviation of the corner location of neighbouring bounding box;Point-by-point feature is exported using a converged network variant;For each
Point, point converged network variant are connect with global point converged network feature and characteristics of image, generate the input of n × 3136
Amount;Intensive converged network handles the input using multiple layers, and exports the prediction of 3D bounding boxes and the score each put;It is surveying
The time is tried, the prediction with top score is selected as finally predicting;The loss function of intensive converged network is:
Wherein, N is the number of input point,Be the true frame demarcated corner location and i-th input point it
Between offset,Be prediction offset, LscoreIt is the loss of score function.
Wherein, global converged network handles, and directly to eight angles of object boundary frame image and point cloud feature
The three-dimensional position fallen is returned;The loss function of global converged network is:
Wherein,It is the corner location of true frame demarcated, xiBe prediction corner location, LstnIt is the space introduced
Regularization loss is converted, for the orthogonality of mandatory learning space conversion matrices;A but major defect of global converged network
It is regressive objectVariance directly depend on concrete condition.
Fig. 3 is a kind of input normalization of three-dimension object detection method based on converged network of the present invention.By searching for
All the points in scene in projectable to frame obtain the corresponding 3D point cloud of image boundary frame;However, spatial position and the 2D of 3D points
Frame position height is related, this can introduce deviation;Point converged network application space converter network (STN) carrys out the specification input space;
But STN cannot correct these deviations completely, therefore use known geometry camera instead to calculate specification spin matrix Rc;RcIt will pass through
The light at 2D frames center rotates to the z-axis of camera frame.
For those skilled in the art, the present invention is not limited to the details of above-described embodiment, in the essence without departing substantially from the present invention
In the case of refreshing and range, the present invention can be realized in other specific forms.In addition, those skilled in the art can be to this hair
Bright to carry out various modification and variations without departing from the spirit and scope of the present invention, these improvements and modifications also should be regarded as the present invention's
Protection domain.Therefore, appended claims are intended to be construed to include preferred embodiment and fall into all changes of the scope of the invention
More and change.
Claims (10)
1. a kind of three-dimension object detection method based on converged network, which is characterized in that main to include point cloud network (one);Melt
Close network (two);Intensive fusion forecasting score function (three).
2. based on the point fusion described in claims 1, which is characterized in that there are three chief components for point fusion:Extraction point
Two the point converged network variant of cloud feature, the convolutional neural networks (CNN) for extracting picture appearance feature, combination features simultaneously export
The converged network of three-dimensional boundaries frame.
3. based on the point cloud network (one) described in claims 1, which is characterized in that point converged network uses symmetric function first
(maximum pond) come realize the invariance of unordered 3D point cloud collection handle;The model absorbs original point cloud, and learns the space each put
Coding and the global point cloud feature of polymerization, by these features for classification and semantic segmentation;
Point converged network can directly handle original point, without the operation that damages of voxelization or projection, and with input point
Quantity is linear, but original point converged network formula cannot be used for 3D recurrence, it is therefore desirable to carry out batch and normalize
It is normalized with input.
4. based on batch being gone to normalize described in claims 3, which is characterized in that in original point converged network is realized,
All full articulamentums all follow a batch normalization layer;But batch normalization hampers the estimation performance of three-dimensional boundaries frame;Batch
Amount normalization is intended to eliminate scale and deviation in input data, but returns task for 3D, and the absolute figure for putting position is that have
It helps;Therefore, point converged network variant deletes all batch normalization layers.
5. based on the input normalization described in claims 3, which is characterized in that by searching in projectable to frame in scene
All the points obtain the corresponding 3D point cloud of image boundary frame;However, the spatial position of 3D points is related to 2D frame position heights, this can draw
Enter deviation;Point converged network application space converter network (STN) carrys out the specification input space;But STN cannot correct these completely
Deviation, therefore known geometry camera is used instead to calculate specification spin matrix Rc;RcTo phase be rotated to by the light at 2D frames center
The z-axis of machine frame.
6. based on the converged network (two) described in claims 1, which is characterized in that converged network will be carried using standard CNN
The corresponding points cloud feature that the characteristics of image and the sub-network of point converged network taken generates is as input;It closes these group of functions
Come and be target object export a 3D bounding box;Converged network includes global converged network and novel intensive converged network.
7. the global converged network described in based on claims 6, which is characterized in that global converged network is to image and Dian Yunte
Sign is handled, and directly the three-dimensional position in eight corners of object boundary frame is returned;The loss of global converged network
Function is:
Wherein,It is the corner location of true frame demarcated, xiBe prediction corner location, LstnBe introduce spatial alternation just
Then change loss, for the orthogonality of mandatory learning space conversion matrices;But a major defect of global converged network is to return
TargetVariance directly depend on concrete condition.
8. the intensive converged network described in based on claims 6, which is characterized in that the main thought of intensive converged network model
It is to use the three-dimensional point of input as intensive space anchor point rather than the absolute position of the direct corner location for returning 3D bounding boxes
It puts, for the three-dimensional point of each input, predicts from this to the spatial deviation of the corner location of neighbouring bounding box;Use a point
Converged network variant exports point-by-point feature;For each point, point converged network variant and global point converged network feature and
Characteristics of image connects, and generates the input tensor of n × 3136;Intensive converged network handles the input using multiple layers, and
The prediction of output 3D bounding boxes and the score each put;In the testing time, the prediction with top score is selected as finally in advance
It surveys;The loss function of intensive converged network is:
Wherein, N is the number of input point,It is inclined between the corner location for the true frame demarcated and i-th of input point
Shifting amount,Be prediction offset, LscoreIt is the loss of score function.
9. the intensive fusion forecasting score function (three) described in based on claims 1, which is characterized in that LscoreThe target of function
It is that Web Cams are learnt into spatial deviation on the point by close-target frame;Specifically, score function includes:
(1) score function is supervised:Direct training network, to predict a point whether in object boundary frame;
(2) unsupervised score function:Network selection is allowed to lead to the point of optimum prediction.
10. based on the supervision score function described in claims 9 and unsupervised score function, which is characterized in that supervision scoring
Whether lose training network predicts a point in target frame;The offset of point i is returned loss to be expressed asBy i-th
The binary classification loss of point is expressed asThen:
Wherein, mi∈ { 0,1 } indicates at i-th point whether in object boundary frame, LscoreBe intersect entropy loss, punishment about to
Fixed point whether the incorrect prediction in frame;As defined, the supervision score function by Web Cams in study, to predict mesh
Mark the spatial deviation of the point in bounding box;However, it may not provide optimum, because the point in frame may not be to have
The point of optimum prediction;
The target of unsupervised scoring is that network is allowed to be directly acquainted with which point may provide best hypothesis;It needs to instruct network
Practice, to determine that there may be the high confidence levels well predicted;The formula includes two loss conditions vied each other:Selection is all
The high confidence level c of pointi, however, the prediction error of corner location is directly proportional to this confidence level;DefinitionCollection is combined into point i
Inflection point offset return loss;Then loss becomes:
Wherein, w is the weight factor between two;Best w is found by rule of thumb, and enables w=0.1 in all experiments.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810081797.0A CN108171217A (en) | 2018-01-29 | 2018-01-29 | A kind of three-dimension object detection method based on converged network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810081797.0A CN108171217A (en) | 2018-01-29 | 2018-01-29 | A kind of three-dimension object detection method based on converged network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108171217A true CN108171217A (en) | 2018-06-15 |
Family
ID=62515690
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810081797.0A Withdrawn CN108171217A (en) | 2018-01-29 | 2018-01-29 | A kind of three-dimension object detection method based on converged network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108171217A (en) |
Cited By (35)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109034077A (en) * | 2018-08-01 | 2018-12-18 | 湖南拓视觉信息技术有限公司 | A kind of three-dimensional point cloud labeling method and device based on Analysis On Multi-scale Features study |
CN109086683A (en) * | 2018-07-11 | 2018-12-25 | 清华大学 | A kind of manpower posture homing method and system based on cloud semantically enhancement |
CN109345510A (en) * | 2018-09-07 | 2019-02-15 | 百度在线网络技术(北京)有限公司 | Object detecting method, device, equipment, storage medium and vehicle |
CN109635843A (en) * | 2018-11-14 | 2019-04-16 | 浙江工业大学 | A kind of three-dimensional object model classification method based on multi-view image |
CN109635685A (en) * | 2018-11-29 | 2019-04-16 | 北京市商汤科技开发有限公司 | Target object 3D detection method, device, medium and equipment |
CN109657559A (en) * | 2018-11-23 | 2019-04-19 | 盎锐(上海)信息科技有限公司 | Point cloud depth degree perceptual coding engine |
CN109737974A (en) * | 2018-12-14 | 2019-05-10 | 中国科学院深圳先进技术研究院 | A kind of 3D navigational semantic map updating method, device and equipment |
CN109816714A (en) * | 2019-01-15 | 2019-05-28 | 西北大学 | A kind of point cloud object type recognition methods based on Three dimensional convolution neural network |
CN109829476A (en) * | 2018-12-27 | 2019-05-31 | 青岛中科慧畅信息科技有限公司 | End-to-end three-dimension object detection method based on YOLO |
CN109872288A (en) * | 2019-01-31 | 2019-06-11 | 深圳大学 | For the network training method of image denoising, device, terminal and storage medium |
CN110276317A (en) * | 2019-06-26 | 2019-09-24 | Oppo广东移动通信有限公司 | A kind of dimension of object detection method, dimension of object detection device and mobile terminal |
CN110689008A (en) * | 2019-09-17 | 2020-01-14 | 大连理工大学 | Monocular image-oriented three-dimensional object detection method based on three-dimensional reconstruction |
CN111062423A (en) * | 2019-11-29 | 2020-04-24 | 中国矿业大学 | Point cloud classification method of point cloud graph neural network based on self-adaptive feature fusion |
CN111144304A (en) * | 2019-12-26 | 2020-05-12 | 上海眼控科技股份有限公司 | Vehicle target detection model generation method, vehicle target detection method and device |
CN111222395A (en) * | 2019-10-21 | 2020-06-02 | 杭州飞步科技有限公司 | Target detection method and device and electronic equipment |
CN111401264A (en) * | 2020-03-19 | 2020-07-10 | 上海眼控科技股份有限公司 | Vehicle target detection method and device, computer equipment and storage medium |
WO2020151109A1 (en) * | 2019-01-22 | 2020-07-30 | 中国科学院自动化研究所 | Three-dimensional target detection method and system based on point cloud weighted channel feature |
CN111583268A (en) * | 2020-05-19 | 2020-08-25 | 北京数字绿土科技有限公司 | Point cloud virtual selection and cutting method, device and equipment |
CN111612059A (en) * | 2020-05-19 | 2020-09-01 | 上海大学 | Construction method of multi-plane coding point cloud feature deep learning model based on pointpilars |
WO2020207166A1 (en) * | 2019-04-11 | 2020-10-15 | 腾讯科技(深圳)有限公司 | Object detection method and apparatus, electronic device, and storage medium |
CN112053374A (en) * | 2020-08-12 | 2020-12-08 | 哈尔滨工程大学 | 3D target bounding box estimation system based on GIoU |
WO2020253121A1 (en) * | 2019-06-17 | 2020-12-24 | 商汤集团有限公司 | Target detection method and apparatus, intelligent driving method and device, and storage medium |
CN112200303A (en) * | 2020-09-28 | 2021-01-08 | 杭州飞步科技有限公司 | Laser radar point cloud 3D target detection method based on context-dependent encoder |
CN112598635A (en) * | 2020-12-18 | 2021-04-02 | 武汉大学 | Point cloud 3D target detection method based on symmetric point generation |
CN112771850A (en) * | 2018-10-02 | 2021-05-07 | 华为技术有限公司 | Motion estimation using 3D assistance data |
CN112904370A (en) * | 2019-11-15 | 2021-06-04 | 辉达公司 | Multi-view deep neural network for lidar sensing |
CN113052109A (en) * | 2021-04-01 | 2021-06-29 | 西安建筑科技大学 | 3D target detection system and 3D target detection method thereof |
CN113128527A (en) * | 2021-06-21 | 2021-07-16 | 中国人民解放军国防科技大学 | Image scene classification method based on converter model and convolutional neural network |
GB2591171A (en) * | 2019-11-14 | 2021-07-21 | Motional Ad Llc | Sequential fusion for 3D object detection |
CN113239726A (en) * | 2021-04-06 | 2021-08-10 | 北京航空航天大学杭州创新研究院 | Target detection method and device based on coloring point cloud and electronic equipment |
CN113269891A (en) * | 2020-02-14 | 2021-08-17 | 初速度(苏州)科技有限公司 | Method and device for determining three-dimensional bounding box of point cloud data |
CN113780257A (en) * | 2021-11-12 | 2021-12-10 | 紫东信息科技(苏州)有限公司 | Multi-mode fusion weak supervision vehicle target detection method and system |
CN114401666A (en) * | 2019-07-15 | 2022-04-26 | 普罗马顿控股有限责任公司 | Object detection and instance segmentation of 3D point clouds based on deep learning |
US11500063B2 (en) | 2018-11-08 | 2022-11-15 | Motional Ad Llc | Deep learning for object detection using pillars |
US11967873B2 (en) | 2019-09-23 | 2024-04-23 | Canoo Technologies Inc. | Fractional slot electric motors with coil elements having rectangular cross-sections |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107092859A (en) * | 2017-03-14 | 2017-08-25 | 佛山科学技术学院 | A kind of depth characteristic extracting method of threedimensional model |
CN107392122A (en) * | 2017-07-07 | 2017-11-24 | 西安电子科技大学 | Polarization SAR silhouette target detection method based on multipolarization feature and FCN CRF UNEs |
-
2018
- 2018-01-29 CN CN201810081797.0A patent/CN108171217A/en not_active Withdrawn
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107092859A (en) * | 2017-03-14 | 2017-08-25 | 佛山科学技术学院 | A kind of depth characteristic extracting method of threedimensional model |
CN107392122A (en) * | 2017-07-07 | 2017-11-24 | 西安电子科技大学 | Polarization SAR silhouette target detection method based on multipolarization feature and FCN CRF UNEs |
Non-Patent Citations (1)
Title |
---|
XU, DANFEI ET AL: ""PointFusion: Deep Sensor Fusion for 3D Bounding Box Estimation"", 《HTTPS://ARXIV.ORG/PDF/1711.10871V1.PDF》 * |
Cited By (58)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109086683A (en) * | 2018-07-11 | 2018-12-25 | 清华大学 | A kind of manpower posture homing method and system based on cloud semantically enhancement |
CN109086683B (en) * | 2018-07-11 | 2020-09-15 | 清华大学 | Human hand posture regression method and system based on point cloud semantic enhancement |
CN109034077A (en) * | 2018-08-01 | 2018-12-18 | 湖南拓视觉信息技术有限公司 | A kind of three-dimensional point cloud labeling method and device based on Analysis On Multi-scale Features study |
CN109034077B (en) * | 2018-08-01 | 2021-06-25 | 湖南拓视觉信息技术有限公司 | Three-dimensional point cloud marking method and device based on multi-scale feature learning |
CN109345510A (en) * | 2018-09-07 | 2019-02-15 | 百度在线网络技术(北京)有限公司 | Object detecting method, device, equipment, storage medium and vehicle |
US11379699B2 (en) | 2018-09-07 | 2022-07-05 | Apollo Intelligent Driving Technology (Beijing) Co., Ltd. | Object detection method and apparatus for object detection |
CN112771850A (en) * | 2018-10-02 | 2021-05-07 | 华为技术有限公司 | Motion estimation using 3D assistance data |
US11688104B2 (en) | 2018-10-02 | 2023-06-27 | Huawei Technologies Co., Ltd. | Motion estimation using 3D auxiliary data |
CN112771850B (en) * | 2018-10-02 | 2022-05-24 | 华为技术有限公司 | Motion compensation method, system and storage medium using 3D auxiliary data |
US11500063B2 (en) | 2018-11-08 | 2022-11-15 | Motional Ad Llc | Deep learning for object detection using pillars |
CN109635843A (en) * | 2018-11-14 | 2019-04-16 | 浙江工业大学 | A kind of three-dimensional object model classification method based on multi-view image |
CN109635843B (en) * | 2018-11-14 | 2021-06-18 | 浙江工业大学 | Three-dimensional object model classification method based on multi-view images |
CN109657559A (en) * | 2018-11-23 | 2019-04-19 | 盎锐(上海)信息科技有限公司 | Point cloud depth degree perceptual coding engine |
CN109657559B (en) * | 2018-11-23 | 2023-02-07 | 盎锐(上海)信息科技有限公司 | Point cloud depth perception coding engine device |
CN109635685B (en) * | 2018-11-29 | 2021-02-12 | 北京市商汤科技开发有限公司 | Target object 3D detection method, device, medium and equipment |
WO2020108311A1 (en) * | 2018-11-29 | 2020-06-04 | 北京市商汤科技开发有限公司 | 3d detection method and apparatus for target object, and medium and device |
CN109635685A (en) * | 2018-11-29 | 2019-04-16 | 北京市商汤科技开发有限公司 | Target object 3D detection method, device, medium and equipment |
CN109737974A (en) * | 2018-12-14 | 2019-05-10 | 中国科学院深圳先进技术研究院 | A kind of 3D navigational semantic map updating method, device and equipment |
CN109829476A (en) * | 2018-12-27 | 2019-05-31 | 青岛中科慧畅信息科技有限公司 | End-to-end three-dimension object detection method based on YOLO |
CN109816714A (en) * | 2019-01-15 | 2019-05-28 | 西北大学 | A kind of point cloud object type recognition methods based on Three dimensional convolution neural network |
WO2020151109A1 (en) * | 2019-01-22 | 2020-07-30 | 中国科学院自动化研究所 | Three-dimensional target detection method and system based on point cloud weighted channel feature |
US11488308B2 (en) | 2019-01-22 | 2022-11-01 | Institute Of Automation, Chinese Academy Of Sciences | Three-dimensional object detection method and system based on weighted channel features of a point cloud |
CN109872288A (en) * | 2019-01-31 | 2019-06-11 | 深圳大学 | For the network training method of image denoising, device, terminal and storage medium |
WO2020207166A1 (en) * | 2019-04-11 | 2020-10-15 | 腾讯科技(深圳)有限公司 | Object detection method and apparatus, electronic device, and storage medium |
US11915501B2 (en) | 2019-04-11 | 2024-02-27 | Tencent Technology (Shenzhen) Company Limited | Object detection method and apparatus, electronic device, and storage medium |
WO2020253121A1 (en) * | 2019-06-17 | 2020-12-24 | 商汤集团有限公司 | Target detection method and apparatus, intelligent driving method and device, and storage medium |
CN110276317A (en) * | 2019-06-26 | 2019-09-24 | Oppo广东移动通信有限公司 | A kind of dimension of object detection method, dimension of object detection device and mobile terminal |
CN110276317B (en) * | 2019-06-26 | 2022-02-22 | Oppo广东移动通信有限公司 | Object size detection method, object size detection device and mobile terminal |
CN114401666A (en) * | 2019-07-15 | 2022-04-26 | 普罗马顿控股有限责任公司 | Object detection and instance segmentation of 3D point clouds based on deep learning |
CN110689008A (en) * | 2019-09-17 | 2020-01-14 | 大连理工大学 | Monocular image-oriented three-dimensional object detection method based on three-dimensional reconstruction |
US11967873B2 (en) | 2019-09-23 | 2024-04-23 | Canoo Technologies Inc. | Fractional slot electric motors with coil elements having rectangular cross-sections |
CN111222395A (en) * | 2019-10-21 | 2020-06-02 | 杭州飞步科技有限公司 | Target detection method and device and electronic equipment |
CN111222395B (en) * | 2019-10-21 | 2023-05-23 | 杭州飞步科技有限公司 | Target detection method and device and electronic equipment |
US11634155B2 (en) | 2019-11-14 | 2023-04-25 | Motional Ad Llc | Sequential fusion for 3D object detection |
GB2591171A (en) * | 2019-11-14 | 2021-07-21 | Motional Ad Llc | Sequential fusion for 3D object detection |
GB2591171B (en) * | 2019-11-14 | 2023-09-13 | Motional Ad Llc | Sequential fusion for 3D object detection |
US11214281B2 (en) | 2019-11-14 | 2022-01-04 | Motional Ad Llc | Sequential fusion for 3D object detection |
CN112904370A (en) * | 2019-11-15 | 2021-06-04 | 辉达公司 | Multi-view deep neural network for lidar sensing |
CN111062423A (en) * | 2019-11-29 | 2020-04-24 | 中国矿业大学 | Point cloud classification method of point cloud graph neural network based on self-adaptive feature fusion |
CN111062423B (en) * | 2019-11-29 | 2022-04-26 | 中国矿业大学 | Point cloud classification method of point cloud graph neural network based on self-adaptive feature fusion |
CN111144304A (en) * | 2019-12-26 | 2020-05-12 | 上海眼控科技股份有限公司 | Vehicle target detection model generation method, vehicle target detection method and device |
CN113269891A (en) * | 2020-02-14 | 2021-08-17 | 初速度(苏州)科技有限公司 | Method and device for determining three-dimensional bounding box of point cloud data |
CN113269891B (en) * | 2020-02-14 | 2022-06-24 | 魔门塔(苏州)科技有限公司 | Method and device for determining three-dimensional bounding box of point cloud data |
CN111401264A (en) * | 2020-03-19 | 2020-07-10 | 上海眼控科技股份有限公司 | Vehicle target detection method and device, computer equipment and storage medium |
CN111583268A (en) * | 2020-05-19 | 2020-08-25 | 北京数字绿土科技有限公司 | Point cloud virtual selection and cutting method, device and equipment |
CN111612059B (en) * | 2020-05-19 | 2022-10-21 | 上海大学 | Construction method of multi-plane coding point cloud feature deep learning model based on pointpilars |
CN111612059A (en) * | 2020-05-19 | 2020-09-01 | 上海大学 | Construction method of multi-plane coding point cloud feature deep learning model based on pointpilars |
CN111583268B (en) * | 2020-05-19 | 2021-04-23 | 北京数字绿土科技有限公司 | Point cloud virtual selection and cutting method, device and equipment |
CN112053374A (en) * | 2020-08-12 | 2020-12-08 | 哈尔滨工程大学 | 3D target bounding box estimation system based on GIoU |
CN112200303B (en) * | 2020-09-28 | 2022-10-21 | 杭州飞步科技有限公司 | Laser radar point cloud 3D target detection method based on context-dependent encoder |
CN112200303A (en) * | 2020-09-28 | 2021-01-08 | 杭州飞步科技有限公司 | Laser radar point cloud 3D target detection method based on context-dependent encoder |
CN112598635B (en) * | 2020-12-18 | 2024-03-12 | 武汉大学 | Point cloud 3D target detection method based on symmetric point generation |
CN112598635A (en) * | 2020-12-18 | 2021-04-02 | 武汉大学 | Point cloud 3D target detection method based on symmetric point generation |
CN113052109A (en) * | 2021-04-01 | 2021-06-29 | 西安建筑科技大学 | 3D target detection system and 3D target detection method thereof |
CN113239726B (en) * | 2021-04-06 | 2022-11-08 | 北京航空航天大学杭州创新研究院 | Target detection method and device based on coloring point cloud and electronic equipment |
CN113239726A (en) * | 2021-04-06 | 2021-08-10 | 北京航空航天大学杭州创新研究院 | Target detection method and device based on coloring point cloud and electronic equipment |
CN113128527A (en) * | 2021-06-21 | 2021-07-16 | 中国人民解放军国防科技大学 | Image scene classification method based on converter model and convolutional neural network |
CN113780257A (en) * | 2021-11-12 | 2021-12-10 | 紫东信息科技(苏州)有限公司 | Multi-mode fusion weak supervision vehicle target detection method and system |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108171217A (en) | A kind of three-dimension object detection method based on converged network | |
Pandey et al. | Ford campus vision and lidar data set | |
Benenson et al. | Stixels estimation without depth map computation | |
Paz et al. | Probabilistic semantic mapping for urban autonomous driving applications | |
US9158992B2 (en) | Acceleration of linear classifiers | |
US20230072731A1 (en) | System and method for panoptic segmentation of point clouds | |
WO2022141858A1 (en) | Pedestrian detection method and apparatus, electronic device, and storage medium | |
Gao et al. | Fine-grained off-road semantic segmentation and mapping via contrastive learning | |
Taran et al. | Impact of ground truth annotation quality on performance of semantic image segmentation of traffic conditions | |
CN113807399A (en) | Neural network training method, neural network detection method and neural network detection device | |
Schubert et al. | Visual place recognition: A tutorial | |
CN114088099A (en) | Semantic relocation method and device based on known map, electronic equipment and medium | |
Yu et al. | Accurate and robust visual localization system in large-scale appearance-changing environments | |
CN112070071A (en) | Method and device for labeling objects in video, computer equipment and storage medium | |
Vashisht et al. | Effective implementation of machine learning algorithms using 3D colour texture feature for traffic sign detection for smart cities | |
CN114550091A (en) | Unsupervised pedestrian re-identification method and unsupervised pedestrian re-identification device based on local features | |
Singh et al. | Efficient deep learning-based semantic mapping approach using monocular vision for resource-limited mobile robots | |
Börcs et al. | A model-based approach for fast vehicle detection in continuously streamed urban LIDAR point clouds | |
Nath et al. | Deep learning models for content-based retrieval of construction visual data | |
Yan et al. | Lane information perception network for HD maps | |
Yudin et al. | Hpointloc: Point-based indoor place recognition using synthetic rgb-d images | |
Guo et al. | Optimal path planning in field based on traversability prediction for mobile robot | |
CN107193965B (en) | BoVW algorithm-based rapid indoor positioning method | |
Li et al. | Instance-aware semantic segmentation of road furniture in mobile laser scanning data | |
Changalasetty et al. | Classification of moving vehicles using k-means clustering |
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 | ||
WW01 | Invention patent application withdrawn after publication | ||
WW01 | Invention patent application withdrawn after publication |
Application publication date: 20180615 |