CN117907970A - Method and device for generating target detection model of laser radar and method and device for detecting target - Google Patents
Method and device for generating target detection model of laser radar and method and device for detecting target Download PDFInfo
- Publication number
- CN117907970A CN117907970A CN202410311817.4A CN202410311817A CN117907970A CN 117907970 A CN117907970 A CN 117907970A CN 202410311817 A CN202410311817 A CN 202410311817A CN 117907970 A CN117907970 A CN 117907970A
- Authority
- CN
- China
- Prior art keywords
- target detection
- data set
- detection model
- data
- reasoning
- 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.)
- Granted
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 260
- 238000000034 method Methods 0.000 title claims abstract description 49
- 238000012549 training Methods 0.000 claims abstract description 33
- 238000012545 processing Methods 0.000 claims description 34
- 101100517651 Caenorhabditis elegans num-1 gene Proteins 0.000 claims description 12
- 230000008569 process Effects 0.000 claims description 8
- 238000012216 screening Methods 0.000 claims description 8
- 230000008685 targeting Effects 0.000 claims description 6
- 230000006870 function Effects 0.000 description 9
- 230000006872 improvement Effects 0.000 description 7
- 230000007246 mechanism Effects 0.000 description 3
- 238000013459 approach Methods 0.000 description 2
- 238000004590 computer program Methods 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 230000011218 segmentation Effects 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 238000013136 deep learning model Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- 238000002372 labelling Methods 0.000 description 1
- 238000013507 mapping Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 230000000750 progressive effect Effects 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 239000004065 semiconductor Substances 0.000 description 1
- 230000003068 static effect Effects 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Landscapes
- Optical Radar Systems And Details Thereof (AREA)
Abstract
The invention provides a method and a device for generating a target detection model of a laser radar, wherein the method for generating the target detection model comprises the following steps: acquiring a marked first data set, and training a target detection model based on the first data set; Obtaining an unlabeled second dataset using a target detection modelReasoning the second data set to obtain labels of the second data set; respectively reasoning the second data set by using N trained target detection models to generate a target detection model based on the point cloudTarget detection modelTarget detection for lidar; training target detection model based on first data setThe final target detection model is the target detection model. The generating method can generate the target detection model of the laser radar.
Description
Technical Field
The invention relates to the technical field of laser radars, in particular to a method and a device for generating and detecting a target detection model of a laser radar.
Background
The current development of automobiles is focused on automatic driving technology, which refers to automobile control technology with fully automatic and highly centralized control of the work performed by the automobile driver. In an automatic driving system, the laser radar is more and more widely applied, and the basic principle is as follows: the laser pulse is used for measuring and sensing the environment around the vehicle, so that high-precision three-dimensional scene reconstruction and target detection are realized, and then an automobile control instruction is made.
It will be appreciated that the target detection algorithm is a very important technique, and at present, the target detection algorithm generally includes: 1. a point cloud based target detection algorithm, which relies mainly on the processing and feature extraction of point cloud data, typically identifies and classifies different objects by clustering, segmentation and feature extraction of the point cloud. However, due to sparsity and noise interference of the point cloud data, these algorithms have limited accuracy and robustness in complex scenarios; 2. deep learning-based target detection algorithms, generally including deep learning models (e.g., convolutional neural networks, recurrent neural networks, etc.), which automatically learn the ability to extract features from point cloud data and classify targets; 3. the target detection algorithm based on multi-sensor fusion can comprehensively utilize the advantages of different sensors and provide a more comprehensive and accurate target detection result. By fusing the data of multiple sensors, the target can be better detected and tracked in different environments and situations.
In summary, providing a target detection method for laser radar is a problem to be solved.
Disclosure of Invention
The invention aims to provide a method and a device for generating and detecting a target detection model of a laser radar.
In order to achieve one of the above objects, an embodiment of the present invention provides a method for generating an object detection model of a lidar, including the steps of: acquiring a marked first data set and generating a target detection model based on point cloudThe first data set is acquired by a laser radar, and the target detection model/>Target detection for lidar; training a target detection model/>, based on the first data set; Acquiring an unlabeled second data set, wherein the second data set is acquired by a laser radar; using target detection model/>Reasoning the second data set to obtain labels of the second data set; respectively reasoning the second data set by using N trained target detection models, and carrying out the following processing on each data in the second data set: n first voting results obtained by reasoning data of N trained target detection models are obtained, and a target detection model/> isobtainedThe second voting result is processed by the data, when the mode number of the N first voting results is equal to the second voting result, the data are deleted from the second data set, the data are added into the first data set, and the pseudo tag corresponding to the data is the second voting result; wherein N is a natural number; generating a point cloud based target detection model/>Target detection model/>Target detection for lidar; training a target detection model/>, based on the first data setThe final target detection model is the target detection model/>。
As a further improvement of an embodiment of the present invention, the method further comprises the steps of: the initial value of i is 2, and the following operations are continuously executed: using object detection modelsReasoning the second data set to obtain labels of the second data set, respectively reasoning the second data set by using N trained target detection models, initializing Num < 1 > =0, obtaining the number Num < 2 > of data in the second data set, and performing the following processing on each data in the second data set: n first voting results obtained by reasoning data of N trained target detection models are obtained, and a target detection model/> isobtainedThe second voting result of the data processing, delete the data from the second data set when the mode of N first voting results is equal to the second voting result, increase 1 in Num1 value, and add the data to the first data set, and the pseudo tag corresponding to the data is the second voting result; stopping performing the latter operation when Num1/Num2> is a preset threshold; otherwise, the value of i is increased by 1, and a target detection model/>, based on the point cloud, is generatedTarget detection model/>Target detection for lidar; training a target detection model/>, based on the first data set; Wherein 0< preset threshold <1; the final object detection model is replaced by the object detection model/>。
As a further improvement of an embodiment of the present invention, the acquiring of the unlabeled second data set, the second data set being acquired by the lidar, specifically includes using a target detection modelReasoning the second data set to obtain a label of the second data set, wherein in the reasoning process, the initial label is based on entropy/>To measure the reliability of the result, the screening rule is thatWherein/>，/>For the number of categories in the second dataset,/>For the entropy of the detection result,/>Is the probability of targeting the c-th category, the sum of the confidence levels of all categories in the second dataset is 1,/>For a preset entropy threshold,/>The pseudo tag label representing the sample for category c, the function argmax () is output such that the function value/>Category at maximum.
As a further improvement of one embodiment of the present invention, n=3, 3 trained object detection models are specifically: pointpillars target detection model, pointNet target detection model, and VoxelNet target detection model.
The embodiment of the invention also provides a device for generating the target detection model of the laser radar, which comprises the following modules: an initialization module for acquiring the marked first data set and generating a target detection model based on the point cloudThe first data set is acquired by a laser radar, and the target detection model/>Target detection for lidar; training a target detection model/>, based on the first data set; The first processing module is used for acquiring an unlabeled second data set, and the second data set is acquired by the laser radar; using target detection model/>Reasoning the second data set to obtain labels of the second data set; respectively reasoning the second data set by using N trained target detection models, and carrying out the following processing on each data in the second data set: n first voting results obtained by reasoning data of N trained target detection models are obtained, and a target detection model/> isobtainedThe second voting result is processed by the data, when the mode number of the N first voting results is equal to the second voting result, the data are deleted from the second data set, the data are added into the first data set, and the pseudo tag corresponding to the data is the second voting result; wherein N is a natural number; model training module for generating target detection model/>, based on point cloudTarget detection model/>Target detection for lidar; training target detection model based on first data setThe final target detection model is the target detection model/>。
As a further improvement of an embodiment of the invention, the device further comprises the following modules: a second processing module for: the initial value of i is 2, and the following operations are continuously executed: using object detection modelsReasoning the second data set to obtain labels of the second data set, respectively reasoning the second data set by using N trained target detection models, initializing Num < 1 > =0, obtaining the number Num < 2 > of data in the second data set, and performing the following processing on each data in the second data set: n first voting results obtained by reasoning data of N trained target detection models are obtained, and a target detection model/> isobtainedThe second voting result of the data processing, delete the data from the second data set when the mode of N first voting results is equal to the second voting result, increase 1 in Num1 value, and add the data to the first data set, and the pseudo tag corresponding to the data is the second voting result; stopping performing the latter operation when Num1/Num2> is a preset threshold; otherwise, the value of i is increased by 1, and a target detection model/>, based on the point cloud, is generatedTarget detection model/>Target detection for lidar; training a target detection model/>, based on the first data set; Wherein 0< preset threshold <1; the final object detection model is replaced by the object detection model/>。
As a further improvement of an embodiment of the present invention, the acquiring of the unlabeled second data set, the second data set being acquired by the lidar, specifically includes using a target detection modelReasoning the second data set to obtain a label of the second data set, wherein in the reasoning process, the initial label is based on entropy/>To measure the reliability of the result, the screening rule is thatWherein/>，/>For the number of categories in the second dataset,/>For the entropy of the detection result,/>Is the probability of targeting the c-th category, the sum of the confidence levels of all categories in the second dataset is 1,/>For a preset entropy threshold,/>The pseudo tag label representing the sample for category c, the function argmax () is output such that the function value/>Category at maximum.
As a further improvement of one embodiment of the present invention, n=3, 3 trained object detection models are specifically: pointpillars target detection model, pointNet target detection model, and VoxelNet target detection model.
The embodiment of the invention also provides a target detection method for the laser radar, which comprises the following steps: acquiring a marked first data set and an unmarked second data set, wherein the first data set and the second data set are acquired by a laser radar; executing the generating method to generate a final target detection model; and acquiring the data to be processed acquired by the laser radar, and processing the data to be processed by using a final target detection model.
The embodiment of the invention also provides a target detection device for the laser radar, which comprises the following modules: the model generation module is used for acquiring a marked first data set and an unmarked second data set, wherein the first data set and the second data set are acquired by the laser radar; executing the generating method to generate a final target detection model; and the third processing module is used for acquiring the data to be processed acquired by the laser radar and processing the data to be processed by using a final target detection model.
Compared with the prior art, the invention has the technical effects that: the embodiment of the invention provides a method and a device for generating and detecting a target detection model of a laser radar, wherein the method for generating the target detection model comprises the following steps: acquiring a marked first data set, and training a target detection model based on the first data set; Obtaining an unlabeled second dataset using a target detection model/>Reasoning the second data set to obtain labels of the second data set; reasoning the second data set by using the N trained target detection models respectively to generate a target detection model/>, based on the point cloudTarget detection model/>Target detection for lidar; training a target detection model/>, based on the first data setThe final target detection model is the target detection model/>. The generating method can generate the target detection model of the laser radar.
Drawings
Fig. 1 is a flowchart of a method for generating a target detection model of a lidar according to an embodiment of the present invention.
Detailed Description
The following description and the drawings sufficiently illustrate specific embodiments herein to enable those skilled in the art to practice them. Portions and features of some embodiments may be included in, or substituted for, those of others. The scope of the embodiments herein includes the full scope of the claims, as well as all available equivalents of the claims. The terms "first," "second," and the like herein are used merely to distinguish one element from another element and do not require or imply any actual relationship or order between the elements. Indeed the first element could also be termed a second element and vice versa. Moreover, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a structure, apparatus, or device that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such structure, apparatus, or device. Without further limitation, an element defined by the phrase "comprising one … …" does not exclude the presence of other like elements in a structure, apparatus or device that comprises the element. Various embodiments are described herein in a progressive manner, each embodiment focusing on differences from other embodiments, and identical and similar parts between the various embodiments are sufficient to be seen with each other.
The terms "longitudinal," "transverse," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," and the like herein refer to an orientation or positional relationship based on that shown in the drawings, merely for convenience of description herein and to simplify the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operate in a particular orientation, and thus are not to be construed as limiting the invention. In the description herein, unless otherwise specified and limited, the terms "mounted," "connected," and "coupled" are to be construed broadly, and may be, for example, mechanically or electrically coupled, may be in communication with each other within two elements, may be directly coupled, or may be indirectly coupled through an intermediary, as would be apparent to one of ordinary skill in the art.
An embodiment of the present invention provides a method for generating a target detection model of a laser radar, as shown in fig. 1, including the following steps:
step 101: acquiring a marked first data set and generating a target detection model based on point cloud The first data set is acquired by a laser radar, and the target detection model/>Target detection for lidar; training a target detection model/>, based on the first data set; In the inventors' experiments, the first dataset was KITTI dataset, KITTI dataset contained lidar sensor data and labeling information of the target. When the first target detection model is trained, the KITTI data set is divided into a training set and a verification set, and proper training parameters such as a learning rate, a batch size, training iteration times and the like are set, so that the KITTI data set is obtained through iterative training.
Step 102: acquiring an unlabeled second data set, wherein the second data set is acquired by a laser radar; using object detection modelsReasoning the second data set to obtain labels of the second data set; respectively reasoning the second data set by using N trained target detection models, and carrying out the following processing on each data in the second data set: n first voting results obtained by reasoning data of N trained target detection models are obtained, and a target detection model/> isobtainedThe second voting result is processed by the data, when the mode number of the N first voting results is equal to the second voting result, the data are deleted from the second data set, the data are added into the first data set, and the pseudo tag corresponding to the data is the second voting result; wherein N is a natural number; here, the second data set may be a local data set acquired using a lidar, and the local data set should be as close to the actual application scenario as possible with respect to the first data set. Here, the pseudo tag method takes the class having the largest prediction probability as a pseudo tag as a supervised paradigm learned from unlabeled data and labeled data at the same time. After formalization, the entropy minimization is equivalent, since the decision boundary should be as sparse as possible, i.e. low density, regions, so that dense sample data points are avoided to be divided to both sides of the decision boundary, i.e. the model needs to make low entropy predictions on unlabeled data, i.e. entropy minimization. N first voting results obtained by reasoning the data by using N trained target detection models are used, N voting results are generated by the N trained target detection models in each data, wherein some voting results are the same (the number of the voting results is the voting number), some voting results are different, and then the voting result with the largest voting number is the mode of the N first voting results.
Step 103: generating a point cloud based target detection modelTarget detection model/>Target detection for lidar; training a target detection model/>, based on the first data setThe final target detection model is the target detection model。
Here, a target detection model is obtained by retraining using the first data setThe object detection model/>Can be used as a final target detection model.
In the pseudo tag method, a pseudo tag of an unlabeled sample may have noise or errors. The pseudo tag method generates a pseudo tag depending on a prediction result of a model, and if the prediction of the model is wrong, noise is introduced into the pseudo tag. This may lead to negative effects of noise data on the model, especially in the initial stages. In the generating method of the embodiment of the invention, a plurality of trained target detection models are used for reducing noise in the pseudo tag.
The generation method of the embodiment of the invention can generate more accurate pseudo tags. Generally, the method of generating pseudo tags by an initial model which is self-trained often has noise because the inference results of the initial model may be erroneous. To solve this problem, the method introduces a multi-model voting mechanism, and multi-model voting can perform consistency detection by adopting detection results generated by a plurality of models. For the same sample, if the predictions of the different models agree, then the pseudo tag may be considered reliable. Conversely, if the predictions of the different models are inconsistent, the pseudo tag may be unreliable and need to be corrected or discarded.
In this embodiment, the method further includes the following steps: the initial value of i is 2, and the following operations are continuously executed: using object detection modelsReasoning the second data set to obtain labels of the second data set, respectively reasoning the second data set by using N trained target detection models, initializing Num < 1 > =0, obtaining the number Num < 2 > of data in the second data set, and performing the following processing on each data in the second data set: n first voting results obtained by reasoning data of N trained target detection models are obtained, and a target detection model/> isobtainedThe second voting result of the data processing, delete the data from the second data set when the mode of N first voting results is equal to the second voting result, increase 1 in Num1 value, and add the data to the first data set, and the pseudo tag corresponding to the data is the second voting result; stopping performing the latter operation when Num1/Num2> is a preset threshold; otherwise, the value of i is increased by 1, and a target detection model/>, based on the point cloud, is generatedTarget detection model/>Target detection for lidar; training a target detection model/>, based on the first data set; Wherein 0< preset threshold <1; the final object detection model is replaced by the object detection model/>。
Here, the first data set is continuously utilized to train the target detection modelUntil NuN1/NuN2> is preset. In the generation method, a multi-model prediction method is added in the pseudo tag algorithm screening process, so that a new first data set is obtained, and a new target detection model/> isobtainedAnd then, carrying out repeated iterative training by combining the first data set, and obtaining an optimal laser radar target detection model after algorithm convergence.
Here, the accuracy of the model is further improved by iterative training. Combining a multi-model voting mechanism, the iterative training method utilizes a new model obtained by training the public data set and the pseudo tag data set together to reasoner local data and generate new labels, and simultaneously, the multi-model voting mechanism is continuously used to obtain new pseudo tags. Until the accuracy converges to get a more accurate model.
In this embodiment, the acquiring the second unlabeled dataset, where the second dataset is acquired by the lidar specifically includes using a target detection modelReasoning the second data set to obtain a label of the second data set, wherein in the reasoning process, the initial label is based on entropy/>To measure the reliability of the result, the screening rule is thatWherein/>，/>For the number of categories in the second dataset,/>For the entropy of the detection result,/>Is the probability of targeting the c-th category, the sum of the confidence levels of all categories in the second dataset is 1,/>For a preset entropy threshold,/>The pseudo tag label representing the sample for category c, the function argmax () is output such that the function value/>Category at maximum.
Here the number of the elements is the number,Can also be understood as/>The confidence of the category, and furthermore the sum of the confidence of all categories in the input example should be 1. That is, entropy will be minimized when the predicted value of a certain class approaches 1, while the predicted values of all other classes approach 0. In the reasoning process, the part with higher reliability is screened out, when the screening rule is thatIn this case, the sample with the highest confidence can be selected. However, these labels are not all correct. These false labels may produce noise which in turn negatively affects the accuracy of the subsequent models.
In this embodiment, n=3, and the 3 trained target detection models are specifically: pointpillars target detection model, pointNet target detection model, and VoxelNet target detection model.
Pointpillars the object detection model is described in paper PointPillars: fast Encoders for Object Detection from Point Clouds, which is an article published on CVPR 2019 on laser point cloud 3D object detection, and which proposes a new point cloud coding method for extracting point cloud features from PointNet, and mapping the extracted features into 2D pseudo images for object detection in a 2D object detection manner.
PointNet the target detection model is disclosed in the paper PointNet: DEEP LEARNING on Point Sets for 3D Classification and Segmentation, the download address of which is: https:// arxiv.
The VoxelNet object detection model can treat the 3D point cloud data as individual voxels (i.e., stereo blocks). In general, voxelNet has a network structure divided into three parts, namely: 1. feature learning network, 2, middle convolution layer, 3, RPN layer.
The second embodiment of the invention provides a device for generating a target detection model of a laser radar, which comprises the following modules: an initialization module for acquiring the marked first data set and generating a target detection model based on the point cloudThe first data set is acquired by a laser radar, and the target detection model/>Target detection for lidar; training a target detection model/>, based on the first data set; The first processing module is used for acquiring an unlabeled second data set, and the second data set is acquired by the laser radar; using target detection model/>Reasoning the second data set to obtain labels of the second data set; respectively reasoning the second data set by using N trained target detection models, and carrying out the following processing on each data in the second data set: n first voting results obtained by reasoning data of N trained target detection models are obtained, and a target detection model/> isobtainedThe second voting result is processed by the data, when the mode number of the N first voting results is equal to the second voting result, the data are deleted from the second data set, the data are added into the first data set, and the pseudo tag corresponding to the data is the second voting result; wherein N is a natural number; model training module for generating target detection model/>, based on point cloudTarget detection model/>Target detection for lidar; training target detection model based on first data setThe final target detection model is the target detection model/>。
In this embodiment, the method further includes the following modules: a second processing module for: the initial value of i is 2, and the following operations are continuously executed: using object detection modelsReasoning the second data set to obtain labels of the second data set, respectively reasoning the second data set by using N trained target detection models, initializing Num < 1 > =0, obtaining the number Num < 2 > of data in the second data set, and performing the following processing on each data in the second data set: n first voting results obtained by reasoning data of N trained target detection models are obtained, and a target detection model/> isobtainedThe second voting result of the data processing, delete the data from the second data set when the mode of N first voting results is equal to the second voting result, increase 1 in Num1 value, and add the data to the first data set, and the pseudo tag corresponding to the data is the second voting result; stopping performing the latter operation when Num1/Num2> is a preset threshold; otherwise, the value of i is increased by 1, and a target detection model/>, based on the point cloud, is generatedTarget detection model/>Target detection for lidar; training a target detection model/>, based on the first data set; Wherein 0< preset threshold <1; the final object detection model is replaced by the object detection model/>。
In this embodiment, the acquiring the second unlabeled data set, where the second data set is acquired by the lidar specifically includes: using object detection modelsReasoning the second data set to obtain a label of the second data set, wherein in the reasoning process, the initial label is based on entropy/>To measure the reliability of the result, the screening rule is thatWherein/>，/>For the number of categories in the second dataset,/>For the entropy of the detection result,/>Is the probability of targeting the c-th category, the sum of the confidence levels of all categories in the second dataset is 1,/>For a preset entropy threshold,/>The pseudo tag label representing the sample for category c, the function argmax () is output such that the function value/>Category at maximum.
In this embodiment, n=3, and the 3 trained target detection models are specifically: pointpillars target detection model, pointNet target detection model, and VoxelNet target detection model.
The third embodiment of the invention provides a target detection method for a laser radar, which comprises the following steps: acquiring a marked first data set and an unmarked second data set, wherein the first data set and the second data set are acquired by a laser radar; executing the generating method in the first embodiment to generate a final target detection model; and acquiring the data to be processed acquired by the laser radar, and processing the data to be processed by using a final target detection model.
The fourth embodiment of the invention provides a target detection device for a laser radar, which comprises the following modules: the model generation module is used for acquiring a marked first data set and an unmarked second data set, wherein the first data set and the second data set are acquired by the laser radar; executing the generating method in the first embodiment to generate a final target detection model; and the third processing module is used for acquiring the data to be processed acquired by the laser radar and processing the data to be processed by using a final target detection model.
It should be noted that, although the steps are described above in a specific order, it is not meant to necessarily be performed in the specific order, and in fact, some of the steps may be performed concurrently or even in a changed order, as long as the required functions are achieved.
The present invention may be a system, method, and/or computer program product. The computer program product may include a readable storage medium having computer readable program instructions embodied thereon for causing a processor to implement aspects of the present invention.
The readable storage medium may be a tangible device that retains and stores instructions for use by an instruction execution device. The readable storage medium may include, for example, but is not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the readable storage medium would include the following: portable computer disks, hard disks, random Access Memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static Random Access Memory (SRAM), portable compact disk read-only memory (CD-ROM), digital Versatile Disks (DVD), memory sticks, floppy disks, mechanical coding devices, punch cards or in-groove structures such as punch cards or grooves having instructions stored thereon, and any suitable combination of the foregoing.
The foregoing description of embodiments of the invention has been presented for purposes of illustration and description, and is not intended to be exhaustive or limited to the embodiments disclosed. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the various embodiments described. The terminology used herein was chosen in order to best explain the principles of the embodiments, the practical application, or the technical improvements in the marketplace, or to enable others of ordinary skill in the art to understand the embodiments disclosed herein.
Claims (10)
1. The method for generating the target detection model of the laser radar is characterized by comprising the following steps of:
acquiring a marked first data set and generating a target detection model based on point cloud The first data set is acquired by a laser radar, and the target detection model/>Target detection for lidar; training a target detection model/>, based on the first data set；
Acquiring an unlabeled second data set, wherein the second data set is acquired by a laser radar; using object detection modelsReasoning the second data set to obtain labels of the second data set; respectively reasoning the second data set by using N trained target detection models, and carrying out the following processing on each data in the second data set: n first voting results obtained by reasoning data of N trained target detection models are obtained, and a target detection model/> isobtainedThe second voting result is processed by the data, when the mode number of the N first voting results is equal to the second voting result, the data are deleted from the second data set, the data are added into the first data set, and the pseudo tag corresponding to the data is the second voting result; wherein N is a natural number;
Generating a point cloud based target detection model Target detection model/>Target detection for lidar; training a target detection model/>, based on the first data setThe final target detection model is the target detection model/>。
2. The method of generating of claim 1, further comprising the steps of:
the initial value of i is 2, and the following operations are continuously executed:
Using object detection models Reasoning the second data set to obtain labels of the second data set, respectively reasoning the second data set by using N trained target detection models, initializing Num < 1 > =0, obtaining the number Num < 2 > of data in the second data set, and performing the following processing on each data in the second data set: n first voting results obtained by reasoning data of N trained target detection models are obtained, and a target detection model/> isobtainedThe second voting result of the data processing, delete the data from the second data set when the mode of N first voting results is equal to the second voting result, increase 1 in Num1 value, and add the data to the first data set, and the pseudo tag corresponding to the data is the second voting result; stopping performing the latter operation when Num1/Num2> is a preset threshold; otherwise, the value of i is increased by 1, and a target detection model/>, based on the point cloud, is generatedTarget detection model/>Target detection for lidar; training a target detection model/>, based on the first data set; Wherein 0< preset threshold <1; the final object detection model is replaced by the object detection model/>。
3. The method according to claim 1, wherein the acquiring of the unlabeled second data set, the second data set being acquired by the lidar, specifically comprises:
Using object detection models Reasoning the second data set to obtain a label of the second data set, wherein in the reasoning process, the initial label is based on entropy/>To measure the reliability of the result, the screening rule is thatWherein/>，/>For the number of categories in the second dataset,/>For the entropy of the detection result,/>Is the probability of targeting the c-th category, the sum of the confidence levels of all categories in the second dataset is 1,/>For a preset entropy threshold,/>The pseudo tag label representing the sample for category c, the function argmax () is output such that the function value/>Category at maximum.
4. The method of generating according to claim 1, wherein:
N=3, the 3 trained target detection models are specifically: pointpillars target detection model, pointNet target detection model, and VoxelNet target detection model.
5. The device for generating the target detection model of the laser radar is characterized by comprising the following modules:
An initialization module for acquiring the marked first data set and generating a target detection model based on the point cloud The first data set is acquired by a laser radar, and the target detection model/>Target detection for lidar; training a target detection model/>, based on the first data set；
The first processing module is used for acquiring an unlabeled second data set, and the second data set is acquired by the laser radar; using object detection modelsReasoning the second data set to obtain labels of the second data set; respectively reasoning the second data set by using N trained target detection models, and carrying out the following processing on each data in the second data set: n first voting results obtained by reasoning data of N trained target detection models are obtained, and a target detection model/> isobtainedThe second voting result is processed by the data, when the mode number of the N first voting results is equal to the second voting result, the data are deleted from the second data set, the data are added into the first data set, and the pseudo tag corresponding to the data is the second voting result; wherein N is a natural number;
model training module for generating target detection model based on point cloud Target detection model/>Target detection for lidar; training a target detection model/>, based on the first data setThe final target detection model is the target detection model/>。
6. The generating device of claim 5, further comprising the following modules:
A second processing module for: the initial value of i is 2, and the following operations are continuously executed: using object detection models Reasoning the second data set to obtain labels of the second data set, respectively reasoning the second data set by using N trained target detection models, initializing Num < 1 > =0, obtaining the number Num < 2 > of data in the second data set, and performing the following processing on each data in the second data set: n first voting results obtained by reasoning data of N trained target detection models are obtained, and a target detection model/> isobtainedThe second voting result of the data processing, delete the data from the second data set when the mode of N first voting results is equal to the second voting result, increase 1 in Num1 value, and add the data to the first data set, and the pseudo tag corresponding to the data is the second voting result; stopping performing the latter operation when Num1/Num2> is a preset threshold; otherwise, the value of i is increased by 1, and a target detection model/>, based on the point cloud, is generatedTarget detection model/>Target detection for lidar; training a target detection model/>, based on the first data set; Wherein 0< preset threshold <1; the final object detection model is replaced by the object detection model/>。
7. The generating device according to claim 5, wherein the acquiring of the unlabeled second data set, the second data set being acquired by the lidar, specifically comprises:
Using object detection models Reasoning the second data set to obtain a label of the second data set, wherein in the reasoning process, the initial label is based on entropy/>To measure the reliability of the result, the screening rule is thatWherein/>，/>For the number of categories in the second dataset,/>For the entropy of the detection result,/>Is the probability of targeting the c-th category, the sum of the confidence levels of all categories in the second dataset is 1,/>For a preset entropy threshold,/>The pseudo tag label representing the sample for category c, the function argmax () is output such that the function value/>Category at maximum.
8. The generating apparatus according to claim 5, wherein:
N=3, the 3 trained target detection models are specifically: pointpillars target detection model, pointNet target detection model, and VoxelNet target detection model.
9. A target detection method for a lidar, comprising the steps of:
Acquiring a marked first data set and an unmarked second data set, wherein the first data set and the second data set are acquired by a laser radar; performing the generating method of any one of claims 1-4 to generate a final object detection model;
And acquiring the data to be processed acquired by the laser radar, and processing the data to be processed by using a final target detection model.
10. An object detection device for a lidar, comprising the following modules:
The model generation module is used for acquiring a marked first data set and an unmarked second data set, wherein the first data set and the second data set are acquired by the laser radar; performing the generating method of any one of claims 1-4 to generate a final object detection model;
And the third processing module is used for acquiring the data to be processed acquired by the laser radar and processing the data to be processed by using a final target detection model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410311817.4A CN117907970B (en) | 2024-03-19 | 2024-03-19 | Method and device for generating target detection model of laser radar and method and device for detecting target |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202410311817.4A CN117907970B (en) | 2024-03-19 | 2024-03-19 | Method and device for generating target detection model of laser radar and method and device for detecting target |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117907970A true CN117907970A (en) | 2024-04-19 |
| CN117907970B CN117907970B (en) | 2024-05-28 |
Family
ID=90685462
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202410311817.4A Active CN117907970B (en) | 2024-03-19 | 2024-03-19 | Method and device for generating target detection model of laser radar and method and device for detecting target |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117907970B (en) |
Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110363820A (en) * | 2019-06-28 | 2019-10-22 | 东南大学 | It is a kind of based on the object detection method merged before laser radar, image |
| CN111179329A (en) * | 2019-12-31 | 2020-05-19 | 智车优行科技（上海）有限公司 | Three-dimensional target detection method and device and electronic equipment |
| CN112990921A (en) * | 2021-01-29 | 2021-06-18 | 腾讯科技（深圳）有限公司 | Data processing method and device based on block chain, computer and storage medium |
| CN114708474A (en) * | 2022-03-14 | 2022-07-05 | 电子科技大学 | Image semantic understanding algorithm fusing local and global features |
| CN115327572A (en) * | 2022-08-09 | 2022-11-11 | 河南科技大学 | Method for detecting obstacle in front of vehicle |
| CN115761392A (en) * | 2022-10-14 | 2023-03-07 | 际络科技(上海)有限公司 | Laser radar target detection model training method, detection method and device |
| CN115841552A (en) * | 2022-11-30 | 2023-03-24 | 北京百度网讯科技有限公司 | High-precision map generation method and device, electronic equipment and medium |
| CN115877372A (en) * | 2022-11-28 | 2023-03-31 | 重庆长安汽车股份有限公司 | Laser radar detection method, device, vehicle and storage medium |
| CN116343077A (en) * | 2023-01-03 | 2023-06-27 | 西安电子科技大学 | Fire detection early warning method based on attention mechanism and multi-scale characteristics |
| CN116342851A (en) * | 2022-12-29 | 2023-06-27 | 浙江啄云智能科技有限公司 | Target detection model construction method, target detection method and device |
-
2024
- 2024-03-19 CN CN202410311817.4A patent/CN117907970B/en active Active
Patent Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110363820A (en) * | 2019-06-28 | 2019-10-22 | 东南大学 | It is a kind of based on the object detection method merged before laser radar, image |
| CN111179329A (en) * | 2019-12-31 | 2020-05-19 | 智车优行科技（上海）有限公司 | Three-dimensional target detection method and device and electronic equipment |
| CN112990921A (en) * | 2021-01-29 | 2021-06-18 | 腾讯科技（深圳）有限公司 | Data processing method and device based on block chain, computer and storage medium |
| CN114708474A (en) * | 2022-03-14 | 2022-07-05 | 电子科技大学 | Image semantic understanding algorithm fusing local and global features |
| CN115327572A (en) * | 2022-08-09 | 2022-11-11 | 河南科技大学 | Method for detecting obstacle in front of vehicle |
| CN115761392A (en) * | 2022-10-14 | 2023-03-07 | 际络科技(上海)有限公司 | Laser radar target detection model training method, detection method and device |
| CN115877372A (en) * | 2022-11-28 | 2023-03-31 | 重庆长安汽车股份有限公司 | Laser radar detection method, device, vehicle and storage medium |
| CN115841552A (en) * | 2022-11-30 | 2023-03-24 | 北京百度网讯科技有限公司 | High-precision map generation method and device, electronic equipment and medium |
| CN116342851A (en) * | 2022-12-29 | 2023-06-27 | 浙江啄云智能科技有限公司 | Target detection model construction method, target detection method and device |
| CN116343077A (en) * | 2023-01-03 | 2023-06-27 | 西安电子科技大学 | Fire detection early warning method based on attention mechanism and multi-scale characteristics |
Non-Patent Citations (2)
| Title |
|---|
| 王海 等: "基于激光雷达与毫米波雷达融合的车辆目标检测算法", 江苏大学学报, vol. 42, no. 4, 31 July 2021 (2021-07-31), pages 390 - 394 * |
| 田永林: "面向交通场景三维目标检测的平行视觉关键方法研究", 中国科学技术大学博士学位论文, 29 May 2022 (2022-05-29) * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN117907970B (en) | 2024-05-28 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Feng et al. | A review and comparative study on probabilistic object detection in autonomous driving | |
| Nitsch et al. | Out-of-distribution detection for automotive perception | |
| US10318848B2 (en) | Methods for object localization and image classification | |
| US20210089895A1 (en) | Device and method for generating a counterfactual data sample for a neural network | |
| US20210319340A1 (en) | Machine learning model confidence score validation | |
| EP3985552A1 (en) | System for detection and management of uncertainty in perception systems | |
| US11094080B2 (en) | Method and device for determining whether a hand cooperates with a manual steering element of a vehicle | |
| KR20210002018A (en) | Method for estimating a global uncertainty of a neural network | |
| Dong et al. | Automatic aerospace weld inspection using unsupervised local deep feature learning | |
| US11663840B2 (en) | Method and system for removing noise in documents for image processing | |
| CN112633310A (en) | Method and system for classifying sensor data with improved training robustness | |
| Ravindran et al. | Camera, LiDAR, and radar sensor fusion based on Bayesian neural network (CLR-BNN) | |
| Patel et al. | Investigation of uncertainty of deep learning-based object classification on radar spectra | |
| CN117274768A (en) | Training method of target detection network, target detection method and related device | |
| CN117690098A (en) | Multi-label identification method based on dynamic graph convolution under open driving scene | |
| US11551084B2 (en) | System and method of robust active learning method using noisy labels and domain adaptation | |
| CN117907970B (en) | Method and device for generating target detection model of laser radar and method and device for detecting target | |
| Dong et al. | Combination of modified U‐Net and domain adaptation for road detection | |
| CN113095351A (en) | Method for generating marked data by means of an improvement of the initial marking | |
| CN112802034A (en) | Image segmentation and recognition method, model construction method and device and electronic equipment | |
| CN112597997A (en) | Region-of-interest determining method, image content identifying method and device | |
| Alashhab et al. | Precise ship location with cnn filter selection from optical aerial images | |
| US20210201015A1 (en) | System and method for enhancing neural sentence classification | |
| CN111950582A (en) | Determining a perturbation mask for a classification model | |
| Hu et al. | Detection of material on a tray in automatic assembly line based on convolutional neural network |
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 | ||
| GR01 | Patent grant |