CN113657161A - Non-standard small obstacle detection method and device and automatic driving system - Google Patents
Non-standard small obstacle detection method and device and automatic driving system Download PDFInfo
- Publication number
- CN113657161A CN113657161A CN202110799043.0A CN202110799043A CN113657161A CN 113657161 A CN113657161 A CN 113657161A CN 202110799043 A CN202110799043 A CN 202110799043A CN 113657161 A CN113657161 A CN 113657161A
- Authority
- CN
- China
- Prior art keywords
- detection
- segmentation model
- small
- model
- negative
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 181
- 230000011218 segmentation Effects 0.000 claims abstract description 79
- 238000012549 training Methods 0.000 claims abstract description 47
- 230000000694 effects Effects 0.000 claims abstract description 30
- 238000012795 verification Methods 0.000 claims abstract description 23
- 238000012360 testing method Methods 0.000 claims abstract description 20
- 238000000034 method Methods 0.000 claims description 20
- 238000012545 processing Methods 0.000 claims description 14
- 238000013461 design Methods 0.000 claims description 6
- 238000004364 calculation method Methods 0.000 claims description 5
- 238000007781 pre-processing Methods 0.000 claims description 5
- 238000013135 deep learning Methods 0.000 abstract description 12
- 230000001788 irregular Effects 0.000 abstract description 2
- 238000010586 diagram Methods 0.000 description 7
- 230000006870 function Effects 0.000 description 5
- 238000005516 engineering process Methods 0.000 description 4
- 238000012544 monitoring process Methods 0.000 description 3
- 230000004913 activation Effects 0.000 description 2
- 238000013527 convolutional neural network Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 230000018109 developmental process Effects 0.000 description 2
- 230000004927 fusion Effects 0.000 description 2
- 238000010606 normalization Methods 0.000 description 2
- 238000010200 validation analysis Methods 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 1
- 230000004888 barrier function Effects 0.000 description 1
- 238000013145 classification model Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000003064 k means clustering Methods 0.000 description 1
- 238000002372 labelling Methods 0.000 description 1
- 238000011176 pooling Methods 0.000 description 1
- 238000012805 post-processing Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
-
- 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
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a non-standard small obstacle detection method, a non-standard small obstacle detection device and an automatic driving system, which are used for solving the problems that the existing target detection based on deep learning has small size difference and good detection effect of regular shape, but has poor effect on large size difference, irregular shape and very complex background of the target. The non-standard small obstacle detection method comprises the following steps: dividing image data into a training set, a verification set and a test set; training a detection segmentation model based on a training set, verifying the trained detection segmentation model through a verification set, carrying out parameter adjustment on the detection segmentation model according to verification performance, evaluating the performance of the detection segmentation model on a test set collected from a real road scene, and selecting the detection segmentation model with the optimal effect for detecting the nonstandard small obstacles.
Description
Technical Field
The invention relates to the technical field of intelligent driving, in particular to a non-standard small obstacle detection method and device and an automatic driving system.
Background
With the development of society, intelligent driving technology is in a rapid development period. The ability to perceive and understand the environment is the basis and premise of automotive intelligence systems. After the intelligent vehicle senses the surrounding environment and makes analysis, the information can be provided for the control system, and the vehicle is guided to run by avoiding the obstacle. However, for some non-standard obstacles with a height much smaller than the height of the vehicle or with a small size, such as stones and garbage on the road, the obstacle is often not large in size, but may cause unpredictable damage to the vehicle. In the field of image target detection, attempts have been made to detect small obstacles by using a fast area convolutional neural network and a faster area convolutional neural network, and although the detection accuracy is high, the detection speed of these algorithms is slow, and these algorithms are not suitable for detecting small obstacles in real time by vehicles. In order to solve the problem, the detection network based on the single-stage method is adopted to detect the small obstacles, so that the detection time can be greatly shortened under the condition of ensuring the accuracy of category judgment.
The invention provides a method for detecting the target size based on deep learning, which is based on the facts that the existing target detection based on deep learning has small size difference and regular shape detection effect, but the existing target detection based on deep learning has poor effect on large size difference, irregular shape and very complex background.
Disclosure of Invention
The invention mainly aims to disclose a non-standard small obstacle detection method, a non-standard small obstacle detection device and an automatic driving system, which are used for solving the problem that the detection effect of the non-standard small obstacle in the prior art is not ideal.
In order to achieve the above object, according to one aspect of the present invention, a method for detecting a non-standard small obstacle is disclosed, and the following technical solutions are adopted:
a non-standard small obstacle detection method includes: dividing image data into a training set, a verification set and a test set; training a detection segmentation model based on a training set, verifying the trained detection segmentation model through a verification set, carrying out parameter adjustment on the detection segmentation model according to verification performance, evaluating the performance of the detection segmentation model on a test set collected from a real road scene, and selecting the detection segmentation model with the optimal effect for detecting the nonstandard small obstacles.
Further, training the detection segmentation model based on the training set includes: generating actual distribution of the actual pavement nonstandard small obstacles by a hard clustering method according to the sizes of the pavement nonstandard small obstacles; and designing a detection default box of the detection segmentation model through actual distribution, and approximately representing a rectangular box of the whole training set by using N detection default boxes through clustering to obtain N types of actual detection boxes with the length, width and height close to the length, width and height of the whole data set.
Further, the network default box for detecting the segmentation model through the actual distribution design comprises: n network default frames are set on 3 feature layers on the EfficientDet, and N/3 frames are preset in each feature layer, wherein the feature layer with a large scale predicts small target features, and the feature layer with a small scale predicts large target features.
Further, training the detection segmentation model based on the training set further comprises: on the basis of the EfficientDet of the basic network model, adding an FPN module to carry out two-classification segmentation of the foreground and the background, wherein the specific calculation mode is as follows:
wherein Accuracy represents model Accuracy, and Recall represents model Recall; TP represents true positive, i.e., determined to be a positive sample, which is in fact also a positive sample; TN represents true negative and is judged as negative, and is actually also negative; FP represents false positive, judged as positive, but in fact negative; FN represents false negative, judged as negative, but in fact positive; the Intersection and Union are the Intersection and Union of the "predicted bounding box" and the "real bounding box", respectively.
Further, after the determining the detection segmentation model with the optimal effect is used for detecting the non-small standard obstacle, the method for detecting the non-small standard obstacle further includes: preprocessing a target detection picture and inputting the preprocessed target detection picture into a detection segmentation model; the detection segmentation model outputs the position and the confidence coefficient of a detection frame on a target detection picture; and processing the detection frame to obtain the detection effect with the highest confidence coefficient and no redundant frame.
According to another aspect of the present invention, a non-standard small obstacle detection device is provided, and the following technical solutions are adopted:
a non-standard small obstacle detection device includes: the dividing module is used for dividing the image data into a training set, a verification set and a test set; and the detection module is used for training the detection segmentation model based on the training set, verifying the trained detection segmentation model through the verification set, adjusting parameters of the detection segmentation model according to verification performance, evaluating the performance of the detection segmentation model on a test set acquired from a real road scene, and selecting the detection segmentation model with the optimal effect for detecting the nonstandard small obstacles.
Further, the detection module includes: the generating module is used for generating actual distribution of the actual pavement non-standard small obstacles by a hard clustering method according to the sizes of the pavement non-standard small obstacles; and the design module is used for designing a detection default box of the detection segmentation model through actual distribution, approximately representing the rectangular box of the whole training set by N detection default boxes through clustering, and obtaining N types of actual detection boxes with the length, width and height close to the length, width and height of the whole data set.
Further, the detection module further comprises: and the computing module is used for adding an FPN module to carry out two-classification segmentation on the foreground and the background on the basis of the efficientDet of the basic network model, and the specific computing mode is as follows:
wherein Accuracy represents model Accuracy, and Recall represents model Recall; TP represents true positive, i.e., determined to be a positive sample, which is in fact also a positive sample; TN represents true negative and is judged as negative, and is actually also negative; FP represents false positive, judged as positive, but in fact negative; FN represents false negative, judged as negative, but in fact positive; the Intersection and Union are the Intersection and Union of the "predicted bounding box" and the "real bounding box", respectively.
Further, the non-standard small obstacle detection device further includes: the input module is used for preprocessing a target detection picture and inputting the preprocessed target detection picture into the detection segmentation model; the output module is used for detecting the position and the confidence coefficient of the detection frame output by the segmentation model on the target detection picture; and the processing module is used for processing the detection frame to obtain the detection effect with the highest confidence coefficient and no redundant frame.
According to another aspect of the invention, an automatic driving system is provided, and the following technical scheme is adopted:
the automatic driving system comprises the non-standard small obstacle detection device.
The invention provides a multi-task non-standard small obstacle detection method based on deep learning, aiming at overcoming the defects existing in the problems of the existing deep learning method, and improving the small obstacle detection method. The target detection uses EfficientNet as a backbone network, EfficientDet as a detector and FPN as a divider. The detection accuracy of the lifting model which can be greatly combined with the segmentation is detected, the condition that the target is segmented and adhered can be reduced to a great extent, real-time monitoring of small non-standard obstacles on the road surface is realized through a deep learning technology, and through improving a network, the detection of the small obstacles with different shapes and sizes and classification detection of the three common road obstacles on the road surface can be adapted to more, so that the detection and segmentation accuracy is improved.
Drawings
In order to more clearly illustrate the embodiments of the present application or technical solutions in the prior art, the drawings needed to be used in the embodiments will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments described in the present invention, and other drawings can be obtained by those skilled in the art according to the drawings.
Fig. 1 is a flowchart of a non-standard small obstacle detection method according to an embodiment of the present invention;
FIG. 2 is a schematic diagram of a training procedure of a non-standard small obstacle detection model according to an embodiment of the present invention;
FIG. 3 is a first network structure according to an embodiment of the present invention;
fig. 4 is a second network structure diagram according to the embodiment of the present invention;
FIG. 5 is a schematic diagram illustrating a segmentation effect according to an embodiment of the present invention;
fig. 6 is a schematic diagram of a non-standard small obstacle detection model inference step according to an embodiment of the present invention;
FIG. 7 is a diagram illustrating the detection effect according to an embodiment of the present invention; and
fig. 8 is a schematic view of a non-standard small obstacle detection device according to an embodiment of the present invention.
Detailed Description
The embodiments of the invention will be described in detail below with reference to the drawings, but the invention can be implemented in many different ways as defined and covered by the claims.
Fig. 1 is a flowchart of a non-standard small obstacle detection method according to an embodiment of the present invention.
Referring to fig. 1, the non-standard small obstacle detection method includes:
s101: dividing image data into a training set, a verification set and a test set;
s103: training a detection segmentation model based on a training set, verifying the trained detection segmentation model through a verification set, carrying out parameter adjustment on the detection segmentation model according to verification performance, evaluating the performance of the detection segmentation model on a test set collected from a real road scene, and selecting the detection segmentation model with the optimal effect for detecting the nonstandard small obstacles.
First, in step S101, an annotated image data set is divided into a training set, a validation set, and a test set. The training set and the test set are specifically as follows 8: 2, wherein the training set comprises real pictures and pictures generated after data enhancement, and the test set is pictures acquired in real road scenes.
Then, in step S103, the detection segmentation model is trained based on the training set.
Fig. 2 is a schematic diagram of a training procedure of a non-standard small obstacle detection model according to an embodiment of the present invention. The non-standard small obstacle detection model training step comprises the following steps:
step 1031: dividing non-standard small obstacle data into a training set, a verification set and a test set;
step 1033: clustering the size proportion of a model default frame according to the size of the non-standard small obstacle;
step 1035: designing a detection model network structure, and adding an FPN (field programmable gate array) segmentation module;
step 1037: and carrying out model hyper-parameter adjustment and detection performance evaluation.
Specifically, according to the size of the small road obstacles, the actual distribution of the actual small road obstacles is generated through a K-Means clustering method, and the default frame of the network model is designed through the actual distribution, so that the model can be more easily regressed, for example, each feature adopts N default frames, and the default value of N is 9. The Means principle is a method for automatically dividing data without labels into several classes, and clustering regression is performed on the width and height of a labeled box. The rectangular frame of the whole data set is approximately represented by N frames through clustering, and the obtained N types of length and width are close to the distribution condition of the width and height frames of the whole data set, so that the model can better regress the default frame. Because the sizes of the small obstacles are different, the sizes of the labeling boxes are different, and the default boxes set by the model can better solve the distribution of real data through the clustered boxes.
K-N default frames are set on 3 feature layers on the EfficientDet, and N/3 frames are preset in each feature layer. The feature layer with large scale predicts small target features, and the feature layer with small scale predicts large target features. And designing a detection model network structure. The method is improved on the basis of a basic network model EfficientDet, an FPN module is added for carrying out two-classification segmentation on the foreground and the background, the segmentation effect is shown in figure 5, the accuracy, the recall rate and the IoU value of the detection model are improved, and the calculation mode is as follows:
wherein accuracy represents the accuracy of the model, and recall represents the recall rate of the model; TP represents true positive, i.e., determined to be a positive sample, which is in fact also a positive sample; TN represents true negative and is judged as negative, and is actually also negative; FP represents false positive, judged as positive, but in fact negative; FN represents false negative, judged as negative, but in fact positive; the Intersection and Union are the Intersection and Union of the "predicted bounding box" and the "real bounding box", respectively. And finally, carrying out hyper-parameter adjustment according to the performance of the model on the verification set, selecting a training model with the optimal effect, and evaluating the performance of the training model on a test set collected from a real road scene.
The hyper-parameter adjustment is to select the combination corresponding to the result with the optimal effect as the parameter combination finally used by the model through setting the combination of a plurality of groups of hyper-parameters and model training.
In step 1035, designing a detection model network structure and adding an FPN segmentation module specifically includes the following steps, which are shown in fig. 3 to 4.
As shown in fig. 3, a picture is taken as input data to enter a network, after convolution processing, MBConv1 processing is started, then after 6 times of MBConv6 processing, after three times of pooling-batch normalization-Relu function activation, upsampling is performed, feature maps after three times of upsampling are successively fused, wherein the picture after three times of upsampling feature map fusion is transmitted to a segmentation network for segmentation, and is transmitted to a prediction network together with the pictures after two times of upsampling feature map fusion, so that small obstacle classification and prediction of the position and size of a detection frame are performed. More specifically, fig. 4 shows the MB convolution process in the network structure, i.e. the MBConv1 process: after input, batch normalization, activation of a Swish function, deep convolution, a Swish function, global pooling of output, convolution, Swish function processing, convolution again, Sigmoid function processing, and output entering the next stage, namely convolution first, batch normalization and Dropout processing.
More specifically, the hyper-parameters are adjusted according to the performance of the model on the verification set, the training model with the optimal effect is selected, and the performance of the training model is evaluated on the test set collected from the real road scene.
As a preferred embodiment, after the detection segmentation model with the best determination effect is used for detecting the non-small-scale obstacle, the non-small-scale obstacle detection method further includes a non-small-scale obstacle detection model inference step, specifically referring to fig. 6. The non-standard small obstacle detection model reasoning step comprises the following steps:
step 201: cutting according to the detection frame result;
step 203: outputting a category confidence through a classification model;
step 205: and obtaining the detection effect with highest confidence coefficient and no redundant frame and the segmentation effect of the foreground and the background through post-processing.
The method for detecting the non-standard small obstacles of the multi-task network based on the deep learning improves the performance of detecting the non-standard small obstacles, and has good detection effects on the non-standard small obstacles with uneven data distribution and different sizes and shapes, and the specific participation effect schematic diagram is shown in fig. 7.
Fig. 8 is a schematic view of a non-standard small obstacle detection device according to an embodiment of the present invention.
Referring to fig. 8, a non-standard small obstacle detecting apparatus includes: a dividing module 80 for dividing the image data into a training set, a validation set and a test set; and the detection module 82 is used for training the detection segmentation model based on the training set, verifying the trained detection segmentation model through the verification set, adjusting parameters of the detection segmentation model according to verification performance, evaluating the performance of the detection segmentation model on a test set acquired from a real road scene, and selecting the detection segmentation model with the optimal effect for detecting the nonstandard small obstacle.
Optionally, the detection module 82 includes: a generating module (not shown in the figure) for generating the actual distribution of the actual road surface non-standard small obstacles by a hard clustering method according to the sizes of the road surface non-standard small obstacles; and a design module (not shown) designs a detection default box of the detection segmentation model through actual distribution, approximately represents a rectangular box of the whole training set by N detection default boxes through clustering, and obtains N types of actual detection boxes with the length, width and height close to the length, width and height of the whole data set.
Optionally, the detection module 82 further includes: optionally, the calculation module is configured to add an FPN module to perform two-class segmentation of the foreground and the background based on the EfficientDet of the basic network model, and the specific calculation method is as follows:
wherein Accuracy represents model Accuracy, and Recall represents model Recall; TP represents true positive, i.e., determined to be a positive sample, which is in fact also a positive sample; TN represents true negative and is judged as negative, and is actually also negative; FP represents false positive, judged as positive, but in fact negative; FN represents false negative, judged as negative, but in fact positive; the Intersection and Union are the Intersection and Union of the "predicted bounding box" and the "real bounding box", respectively.
As a preferred embodiment, the non-standard small obstacle detecting device further includes: optionally, the input module is configured to input the detection segmentation model after preprocessing the target detection picture; optionally, the output module is configured to detect a position and a confidence of a detection frame output by the segmentation model on the target detection picture; and the processing module is used for processing the detection frame to obtain the detection effect with the highest confidence coefficient and no redundant frame.
The automatic driving system provided by the invention comprises the non-standard small obstacle detection device.
The invention provides a multi-task non-standard small obstacle detection method based on deep learning, aiming at overcoming the defects existing in the problems of the existing deep learning method, and improving the small obstacle detection method. The target detection uses EfficientNet as a backbone network, EfficientDet as a detector and FPN as a divider. The detection accuracy of the lifting model which can be greatly combined with the segmentation is detected, the condition that the target is segmented and adhered can be reduced to a great extent, real-time monitoring of small non-standard obstacles on the road surface is realized through a deep learning technology, and through improving a network, the detection of the small obstacles with different shapes and sizes and classification detection of the three common road obstacles on the road surface can be adapted to more, so that the detection and segmentation accuracy is improved.
The real-time monitoring of the small non-standard obstacles on the road surface is realized through a deep learning technology, the network is improved, the detection and the segmentation of the small obstacles with different shapes and sizes and the classification detection of three common road barriers on the road surface can be adapted more, and the accuracy of the detection and the segmentation is improved.
While certain exemplary embodiments of the present invention have been described above by way of illustration only, it will be apparent to those of ordinary skill in the art that the described embodiments may be modified in various different ways without departing from the spirit and scope of the invention. Accordingly, the drawings and description are illustrative in nature and should not be construed as limiting the scope of the invention.
Claims (10)
1. A method for detecting a non-standard small obstacle, comprising:
dividing image data into a training set, a verification set and a test set;
training a detection segmentation model based on a training set, verifying the trained detection segmentation model through a verification set, carrying out parameter adjustment on the detection segmentation model according to verification performance, evaluating the performance of the detection segmentation model on a test set collected from a real road scene, and selecting the detection segmentation model with the optimal effect for detecting the nonstandard small obstacles.
2. The non-standard small obstacle detection method of claim 1, wherein the training a detection segmentation model based on a training set comprises:
generating actual distribution of the actual pavement nonstandard small obstacles by a hard clustering method according to the sizes of the pavement nonstandard small obstacles;
and designing a detection default box of the detection segmentation model through actual distribution, and approximately representing a rectangular box of the whole training set by using N detection default boxes through clustering to obtain N types of actual detection boxes with the length, width and height close to the length, width and height of the whole data set.
3. The non-nominal small obstacle detection method of claim 2, wherein the network default box for detecting the segmentation model by actual distributed design comprises:
n network default frames are set on 3 feature layers on the EfficientDet, and N/3 frames are preset in each feature layer, wherein the feature layer with a large scale predicts small target features, and the feature layer with a small scale predicts large target features.
4. The non-standard small obstacle detection method of claim 2, wherein training the detection segmentation model based on the training set further comprises:
on the basis of the EfficientDet of the basic network model, adding an FPN module to carry out two-classification segmentation of the foreground and the background, wherein the specific calculation mode is as follows:
wherein Accuracy represents model Accuracy, and Recall represents model Recall; TP represents true positive, i.e., determined to be a positive sample, which is in fact also a positive sample; TN represents true negative and is judged as negative, and is actually also negative; FP represents false positive, judged as positive, but in fact negative; FN represents false negative, judged as negative, but in fact positive; the Intersection and Union are the Intersection and Union of the "predicted bounding box" and the "real bounding box", respectively.
5. The non-small-scale obstacle detection method according to claim 4, wherein after the determining the detection segmentation model with the best effect is used for non-small-scale obstacle detection, the non-small-scale obstacle detection method further comprises:
preprocessing a target detection picture and inputting the preprocessed target detection picture into a detection segmentation model;
the detection segmentation model outputs the position and the confidence coefficient of a detection frame on a target detection picture;
and processing the detection frame to obtain the detection effect with the highest confidence coefficient and no redundant frame.
6. A non-standard small obstacle detection device, comprising:
the dividing module is used for dividing the image data into a training set, a verification set and a test set;
and the detection module is used for training the detection segmentation model based on the training set, verifying the trained detection segmentation model through the verification set, adjusting parameters of the detection segmentation model according to verification performance, evaluating the performance of the detection segmentation model on a test set acquired from a real road scene, and selecting the detection segmentation model with the optimal effect for detecting the nonstandard small obstacles.
7. The non-standard small obstacle detection apparatus of claim 6, wherein the detection module comprises:
the generating module is used for generating actual distribution of the actual pavement non-standard small obstacles by a hard clustering method according to the sizes of the pavement non-standard small obstacles;
and the design module is used for designing a detection default box of the detection segmentation model through actual distribution, approximately representing the rectangular box of the whole training set by N detection default boxes through clustering, and obtaining N types of actual detection boxes with the length, width and height close to the length, width and height of the whole data set.
8. The non-standard small obstacle detection apparatus of claim 7, wherein the detection module further comprises:
and the computing module is used for adding an FPN module to carry out two-classification segmentation on the foreground and the background on the basis of the efficientDet of the basic network model, and the specific computing mode is as follows:
wherein Accuracy represents model Accuracy, and Recall represents model Recall; TP represents true positive, i.e., determined to be a positive sample, which is in fact also a positive sample; TN represents true negative and is judged as negative, and is actually also negative; FP represents false positive, judged as positive, but in fact negative; FN represents false negative, judged as negative, but in fact positive; the Intersection and Union are the Intersection and Union of the "predicted bounding box" and the "real bounding box", respectively.
9. The non-standard small obstacle detecting device according to claim 8, further comprising:
the input module is used for preprocessing a target detection picture and inputting the preprocessed target detection picture into the detection segmentation model;
the output module is used for detecting the position and the confidence coefficient of the detection frame output by the segmentation model on the target detection picture;
and the processing module is used for processing the detection frame to obtain the detection effect with the highest confidence coefficient and no redundant frame.
10. An automatic driving system characterized by comprising the non-standard small obstacle detecting device according to any one of claims 6 to 9.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110799043.0A CN113657161A (en) | 2021-07-15 | 2021-07-15 | Non-standard small obstacle detection method and device and automatic driving system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110799043.0A CN113657161A (en) | 2021-07-15 | 2021-07-15 | Non-standard small obstacle detection method and device and automatic driving system |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN113657161A true CN113657161A (en) | 2021-11-16 |
Family
ID=78477410
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110799043.0A Pending CN113657161A (en) | 2021-07-15 | 2021-07-15 | Non-standard small obstacle detection method and device and automatic driving system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113657161A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114782923A (en) * | 2022-05-07 | 2022-07-22 | 厦门瑞为信息技术有限公司 | Vehicle blind area detection system |
| CN116381698A (en) * | 2023-06-05 | 2023-07-04 | 蘑菇车联信息科技有限公司 | Road remains detection method and device and electronic equipment |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180330615A1 (en) * | 2017-05-12 | 2018-11-15 | Toyota Jidosha Kabushiki Kaisha | Road obstacle detection device, method, and program |
| CN110059558A (en) * | 2019-03-15 | 2019-07-26 | 江苏大学 | A kind of orchard barrier real-time detection method based on improvement SSD network |
| CN110852177A (en) * | 2019-10-17 | 2020-02-28 | 北京全路通信信号研究设计院集团有限公司 | Obstacle detection method and system based on monocular camera |
| CN110992325A (en) * | 2019-11-27 | 2020-04-10 | 同济大学 | Target counting method, device and equipment based on deep learning |
-
2021
- 2021-07-15 CN CN202110799043.0A patent/CN113657161A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180330615A1 (en) * | 2017-05-12 | 2018-11-15 | Toyota Jidosha Kabushiki Kaisha | Road obstacle detection device, method, and program |
| CN110059558A (en) * | 2019-03-15 | 2019-07-26 | 江苏大学 | A kind of orchard barrier real-time detection method based on improvement SSD network |
| CN110852177A (en) * | 2019-10-17 | 2020-02-28 | 北京全路通信信号研究设计院集团有限公司 | Obstacle detection method and system based on monocular camera |
| CN110992325A (en) * | 2019-11-27 | 2020-04-10 | 同济大学 | Target counting method, device and equipment based on deep learning |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114782923A (en) * | 2022-05-07 | 2022-07-22 | 厦门瑞为信息技术有限公司 | Vehicle blind area detection system |
| CN114782923B (en) * | 2022-05-07 | 2024-05-03 | 厦门瑞为信息技术有限公司 | Detection system for dead zone of vehicle |
| CN116381698A (en) * | 2023-06-05 | 2023-07-04 | 蘑菇车联信息科技有限公司 | Road remains detection method and device and electronic equipment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108830188B (en) | Vehicle detection method based on deep learning | |
| US8509478B2 (en) | Detection of objects in digital images | |
| Pamula | Road traffic conditions classification based on multilevel filtering of image content using convolutional neural networks | |
| CN105892471A (en) | Automatic automobile driving method and device | |
| CN109993031A (en) | A kind of animal-drawn vehicle target is driven against traffic regulations behavioral value method, apparatus and camera | |
| CN113657161A (en) | Non-standard small obstacle detection method and device and automatic driving system | |
| CN114821665A (en) | Urban pedestrian flow small target detection method based on convolutional neural network | |
| CN116597411A (en) | Method and system for identifying traffic sign by unmanned vehicle in extreme weather | |
| CN112215073A (en) | Traffic marking line rapid identification and tracking method under high-speed motion scene | |
| CN114863122A (en) | Intelligent high-precision pavement disease identification method based on artificial intelligence | |
| Hashemi et al. | A new comparison framework to survey neural networks‐based vehicle detection and classification approaches | |
| CN118196573A (en) | Vehicle detection method and system based on deep learning | |
| CN111339834B (en) | Method for identifying vehicle driving direction, computer device and storage medium | |
| Mouzinho et al. | Hierarchical semantic segmentation based approach for road surface damages and markings detection on paved road | |
| CN116630904A (en) | Small target vehicle detection method integrating non-adjacent jump connection and multi-scale residual error structure | |
| Shen et al. | Car plate detection based on Yolov3 | |
| Wang et al. | Front obstacle detection system based on YOLOv3 | |
| Candradewi et al. | Intelligent traffic monitoring systems: Vehicle type classification using support vector machine | |
| Joshi et al. | Flow analysis of vehicles on a lane using deep learning techniques | |
| Saffarini et al. | Survey of road anomalies detection methods | |
| Triwibowo et al. | Analysis of Classification and Calculation of Vehicle Type at APILL Intersection Using YOLO Method and Kalman Filter | |
| Wang et al. | YOLOX-based dynamic vehicle target detection | |
| CN113111732B (en) | High-speed service area dense pedestrian detection method | |
| Al Khafaji et al. | Traffic Signs Detection and Recognition Using A combination of YOLO and CNN | |
| Senthilnayaki et al. | Traffic Sign Prediction and Classification Using Image Processing Techniques |
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 | ||
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20211116 |