CN109255052B - Three-stage vehicle retrieval method based on multiple features - Google Patents
Three-stage vehicle retrieval method based on multiple features Download PDFInfo
- Publication number
- CN109255052B CN109255052B CN201810996825.1A CN201810996825A CN109255052B CN 109255052 B CN109255052 B CN 109255052B CN 201810996825 A CN201810996825 A CN 201810996825A CN 109255052 B CN109255052 B CN 109255052B
- Authority
- CN
- China
- Prior art keywords
- vehicle
- image
- area
- straight line
- window
- 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.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/25—Determination of region of interest [ROI] or a volume of interest [VOI]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V2201/00—Indexing scheme relating to image or video recognition or understanding
- G06V2201/08—Detecting or categorising vehicles
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
- Image Processing (AREA)
Abstract
The invention provides a vehicle retrieval method, and particularly relates to a three-stage vehicle retrieval method based on multiple features. The method searches by sequentially utilizing the characteristics from coarse to fine, so as to find the vehicle set most similar to the target vehicle. Firstly, vehicle screening is carried out according to vehicle colors in a first stage, and a vehicle set with the same color as that of a target retrieval vehicle is found through screening; then, in the second stage, screening is carried out according to the vehicle type characteristics, and a vehicle set with high similarity to the target vehicle type is found; and finally, searching according to the vehicle characteristics of the vehicle different from other same vehicle types, and finding out the vehicle data set most similar to the target vehicle. The invention has the advantages that: the retrieval problem is subdivided into a plurality of sub-problems by the multi-stage retrieval mode, the difficulty in solving the overall problem can be reduced, and meanwhile, for mass data retrieval, the retrieval is performed in stages by using the characteristics from coarse to fine, so that the retrieval efficiency can be effectively improved.
Description
Technical Field
The invention relates to the technical field of digital image processing, in particular to a three-stage vehicle retrieval method based on multiple features.
Background
With the important role played by video surveillance in the field of public safety, how to quickly find a target vehicle locked by a public safety department in a video image of a complex environment becomes an urgent need. At present, the combination of the automatic license plate recognition technology and market demands is mature, and the social benefit and the economic benefit of the license plate recognition applied to an actual traffic management system are obvious. However, the license plate recognition system cannot be used for vehicles with the conditions of fouling, shielding, fake plate and the like.
While the conventional recognition technology encounters a bottleneck, an object re-recognition technology based on image analysis is becoming a hot point of research in recent years. However, at present, research on object re-recognition is mainly focused on the pedestrian re-recognition field, and few scholars have started to try the military vehicle re-recognition field in recent years. For example, the trypan et al propose (a vehicle weight recognition method [ J ] based on feature fusion and an L-M algorithm, electronic technology, 2018,4(31):12-15) to fuse HSV features and LBP features of a vehicle image, perform singular value decomposition on a fusion feature matrix, extract feature values, and then optimize a BP neural network by adopting an L-M adaptive adjustment algorithm to realize vehicle matching. The method only uses the traditional simple characteristics and is difficult to meet the vehicle retrieval requirement under the complex scene. Lululuol (MVBI hash algorithm-based vehicle re-identification method research [ D ]. Anhui university, 2018) puts the research center of gravity on the efficiency problem of retrieval, firstly, the vehicle bottom layer characteristics and the depth characteristics are subjected to characteristic fusion by using a correlation analysis algorithm, then, the principal component analysis is used for characteristic dimension reduction, and finally, the characteristic descriptors are mapped to a Hamming space by an MVBI hash algorithm for comparison.
Disclosure of Invention
Aiming at the problems in the prior art, the invention provides a multi-feature-based three-stage vehicle retrieval method, which is used for searching by sequentially utilizing features from thick to thin so as to find a vehicle set most similar to a target vehicle. Firstly, vehicle screening is carried out according to vehicle colors in a first stage, and a vehicle set with the same color as that of a target retrieval vehicle is found through screening; then, in the second stage, screening is carried out according to the vehicle type characteristics, and a vehicle set with high similarity to the target vehicle type is found; and finally, searching according to the vehicle characteristics of the vehicle different from other same vehicle types, and finding out the vehicle data set most similar to the target vehicle. Meanwhile, the invention also introduces a vehicle image region segmentation and reconstruction method based on the vehicle window in the image preprocessing stage.
The technical scheme of the invention is as follows:
a three-stage vehicle retrieval method based on multiple features is characterized by comprising the following steps:
step 1: firstly, training to obtain a detection model and a classification model which are needed in a retrieval process, wherein the training comprises the following steps: the method comprises the following steps of (1) vehicle window detection model winDetect, vehicle color recognition model colorClassify based on a vehicle face region, depth feature extraction network model carTypefeature based on a vehicle type reconstructed image, and depth feature extraction network model carFefeature based on a vehicle reconstructed image; the car face region is obtained according to a car window-based car region segmentation method;
step 2: recording the searched target vehicle as I, and recording the vehicle set in the vehicle search base as X ═ X0,x1,...,xn-1}；
And step 3: identifying the colors of the target vehicle I and all vehicles in the vehicle set X by using a color classification model colorClassify, then finding the vehicles with the same color as the target vehicle I from the X to form a vehicle set S1；
And 4, step 4: obtaining a target vehicle I and a vehicle set S1Corresponding reconstructed image of vehicle type, marked as ITAnd S1 TExtracting I by using a deep network model cartepFeaturTAnd S1 TFinding out a vehicle set S by using the corresponding vehicle type feature vector1The Euclidean distance between all the vehicle type characteristic vectors meeting the requirements of the corresponding vehicle type characteristic vector and the vehicle type characteristic vector I is less than lambda1Constitute a new vehicle set, denoted S2(ii) a Wherein λ1The maximum Euclidean distance threshold value of the preset same-model vehicle in the feature space is obtained;
and 5: obtaining a target vehicle I and a vehicle set S2Corresponding reconstructed image of vehicle, noted as ICAnd S2 CExtracting I by using a deep network model carFeatureCAnd S2 CFinding out a vehicle set S according to the corresponding vehicle feature vector2The Euclidean distance between all the vehicle characteristic vectors meeting the requirement of corresponding vehicle characteristic vector and I is less than lambda2Constitute a new vehicle set, denoted S3，S3Finally finding a vehicle set which is most similar to the target vehicle from the search library; wherein λ2A maximum Euclidean distance threshold value in a feature space for a preset same vehicle;
the three-stage vehicle retrieval method based on multiple features is characterized in that the vehicle region segmentation method based on the vehicle window in the step 1 specifically comprises the following steps:
step 1.1: for any vehicle image A, firstly, a detection model winDetect is utilized to position the approximate position of a vehicle window, and the vehicle window image A is obtained through interceptionWThen determining the window image AWThe detection areas of the upper, lower, left and right sidelines of the upper vehicle window are respectively as follows: dL(0,0,w/3,h)、DR(2*w/3-1,0,w/3,h)、DT(0,0,w,h/5)、DB(0,4 x h/5-1, w, h/5), wherein w and h respectively represent the car window image AWWidth and height of (1);
step 1.2: respectively finding four side lines of the upper side, the lower side, the left side and the right side of the car window in the corresponding detection areas by using any straight line detection method, and respectively recording the corresponding straight lines of the four side lines on the image A as LT、LR、LB、LL；
Step 1.3: determining the areas of the face, windows, roof, left body and right body of the vehicle according to the following rules:
a) the car face area: straight line L on image AR、LB、LLAny polygonal area surrounded by the bottom edge of the image A;
b) the window area is as follows: straight line L on image AR、LB、LL、LTA quadrilateral area is formed by surrounding;
c) roof area: straight line L on image AR、LB、LLAny polygonal area surrounded by the top edge of the image A;
d) left body area: straight line L on image ALAn arbitrary polygon region on the left side;
e) right body area: straight line L on image ARAny polygonal area on the right side;
the three-stage vehicle retrieval method based on multiple features is characterized in that the specific steps of acquiring the vehicle type reconstruction image in the step 4 are as follows:
step 4.1: for any vehicle image AFirstly, a vehicle image area segmentation method based on a vehicle window is utilized to segment a vehicle image area, and then a corresponding vehicle face area image and a corresponding vehicle body area image obtained on an image A are respectively marked as A1，A2(ii) a The vehicle body area image is an area image corresponding to a larger area in the left and right vehicle body areas;
step 4.2: if the body area image A2If the right vehicle body area corresponds to the left vehicle body area, horizontally turning the image;
step 4.3: image A1，A2And uniformly zooming to 200 px-200 px, and horizontally splicing according to the sequence from left to right to obtain a vehicle type reconstructed image corresponding to the vehicle A.
The three-stage vehicle retrieval method based on multiple features is characterized in that the specific steps of acquiring the vehicle reconstruction image in the steps 1 and 5 are as follows:
step 5.1: for any vehicle image A, firstly, a vehicle image area segmentation method based on windows is utilized to carry out vehicle image area segmentation, and then corresponding face area image, window area image, vehicle body area image and roof area image obtained on the image A are respectively marked as A1，A2，A3，A4(ii) a The vehicle body area image is an area image corresponding to a larger area in the left and right vehicle body areas;
step 5.2: if the body area image A3If the right vehicle body area corresponds to the left vehicle body area, horizontally turning the image;
step 5.3: image A1，A2，A3，A4And uniformly zooming to 200 px-200 px, and horizontally splicing according to the sequence from left to right to obtain a vehicle reconstructed image corresponding to the vehicle A.
The three-stage vehicle retrieval method based on multiple features is characterized in that in the step 1,
vehicle window detection model winDetect: training based on a yolo detection network;
vehicle color recognition model colorClassify based on the vehicle face area: the method comprises the steps that the statistic HSV color histogram feature of a car face region is obtained based on training of an SVM (support vector machine) classifier, the used feature is the statistic HSV color histogram feature of the car face region and consists of a 36-dimensional H component, a 20-dimensional S component and a 20-dimensional V component, and the SVM color classifier can recognize black, white, red, yellow, blue and green and other six vehicle colors;
extracting a network model carTypeFeature based on the depth characteristics of a vehicle type reconstructed image: the network model structure for feature extraction consists of a classification network and a dual-input image matching network, the basic network structure for forming the network model structure is cafeNet, and the dimensionality of the network extraction features is 2048; the training data set of the car type feature extraction network model based on the depth features of the vehicle type reconstructed image is a fine category vehicle type data set;
extracting a network model carFeature based on the depth feature of the vehicle type reconstructed image: the network model is the same as a depth feature extraction network model carTypefeature based on a vehicle type reconstructed image, and a training data set of the depth feature extraction network model carFefeature based on the vehicle type reconstructed image is a vehicle re-identification data set.
The three-stage vehicle retrieval method based on multiple features is characterized in that the straight line detection method specifically comprises the following steps:
1) for a detection image X, firstly, giving an angle range of a target straight line, and marking the angle range as [ angleMin, anglemx ];
2) performing gaussian filtering, Canny edge detection and Hough line detection on the image X to obtain a line set L ═ { li | i ═ 0,1, …, NL-1}, wherein NL is the number of all detected lines;
3) finding the angle in the line set L to satisfy [ angleMin, angleMax]All the straight lines in the range, the straight line with the maximum length in the straight lines is the key straight line of the target in the image; wherein, the straight line liLength is a straight line liAll edge points that pass by are the distance of the two points that are farthest apart.
The invention has the beneficial effects that:
1) the retrieval problem is subdivided into a plurality of sub-problems by a multi-stage retrieval mode, the difficulty in solving the overall problem can be reduced, and meanwhile, for mass data retrieval, the retrieval is performed in stages by using the characteristics from coarse to fine, so that the retrieval efficiency can be effectively improved;
2) the vehicle window-based vehicle region dividing and image reconstructing method has the advantages that: extracting different images of a target reconstruction aiming at different characteristics, removing redundant information and improving the accuracy of characteristic expression; the unified layout of the image areas of the vehicles is realized, and the retrieval accuracy of the vehicles with different deflection angles is greatly improved;
drawings
FIG. 1: a deep network model structure diagram for training;
FIG. 2: a deep network model structure chart for feature extraction;
FIG. 3: an example vehicle image;
FIG. 4: a window positioning results map for an example vehicle;
FIG. 5: a model reconstruction image of the example vehicle;
FIG. 6: a vehicle reconstructed image of an example vehicle.
Detailed Description
The following describes in detail a process implemented by the present invention (a multi-feature based three-stage vehicle search method) with reference to specific examples.
Step 1: firstly, training to obtain a detection model and a classification model which are needed in a retrieval process, wherein the training comprises the following steps: the method comprises the following steps of (1) vehicle window detection model winDetect, vehicle color recognition model colorClassify based on a vehicle face region, depth feature extraction network model carTypefeature based on a vehicle type reconstructed image, and depth feature extraction network model carFefeature based on a vehicle reconstructed image; the car face region is obtained according to a car window-based car region segmentation method;
in this example, the detection model and the classification model are specifically as follows:
1) vehicle window detection model winDetect: training based on a yolo detection network;
2) vehicle color recognition model colorclassic: the method comprises the steps that the method is obtained based on training of an SVM (support vector machine) classifier, the used features are statistic HSV color histogram features of a car face region, the statistic HSV color histogram features comprise 36-dimensional H components, 20-dimensional S components and 20-dimensional V components, and the SVM color classifier can recognize black, white, red, yellow, blue-green and other six vehicle colors;
3) deep feature extraction network model cartype feature: the network model structure for training is shown in fig. 1, the network model structure for feature extraction is shown in fig. 2, the network model structure is composed of a classification network and a dual-input image matching network, the basic network structure of the network model structure is cafenet, and the dimensionality of the network extraction features is 2048; the training data set of the model is a data set of vehicle types of a fine category;
4) deep feature extraction network model carFeature: the network model structure for training and detecting is the same as 3), and the training data set of the model is a vehicle re-identification data set;
step 2: recording the searched target vehicle as I, and recording the vehicle set in the vehicle search base as X ═ X0,x1,...,xn-1}；
And step 3: identifying the colors of the target vehicle I and all vehicles in the vehicle set X by using a color classification model colorClassify, then finding the vehicles with the same color as the target vehicle I from the X to form a vehicle set S1；
And 4, step 4: obtaining a target vehicle I and a vehicle set S1Corresponding reconstructed image of vehicle type, marked as ITAnd S1 TExtracting I by using a deep network model cartepFeaturTAnd S1 TFinding out a vehicle set S by using the corresponding vehicle type feature vector1The Euclidean distance between all the vehicle type characteristic vectors meeting the requirements of the corresponding vehicle type characteristic vector and the vehicle type characteristic vector I is less than lambda1Constitute a new vehicle set, denoted S2(ii) a Wherein λ1The maximum Euclidean distance threshold value of the preset same-model vehicle in the feature space is obtained;
and 5: obtaining a target vehicle I and a vehicle set S2Corresponding reconstructed image of vehicle, noted as ICAnd S2 CExtracting I by using a deep network model carFeatureCAnd S2 CCorresponding vehicleFeature vector, find vehicle set S2The Euclidean distance between all the vehicle characteristic vectors meeting the requirement of corresponding vehicle characteristic vector and I is less than lambda2Constitute a new vehicle set, denoted S3，S3Finally finding a vehicle set which is most similar to the target vehicle from the search library; wherein λ2A maximum Euclidean distance threshold value in a feature space for a preset same vehicle;
the three-stage vehicle retrieval method based on multiple features is characterized in that the vehicle region segmentation method based on the vehicle window in the step 1 specifically comprises the following steps:
step 1.1: for any vehicle image A, firstly, a detection model winDetect is utilized to position the approximate position of a vehicle window, and the vehicle window image A is obtained through interceptionWThen determining the window image AWThe detection areas of the upper, lower, left and right sidelines of the upper vehicle window are respectively as follows: dL(0,0,w/3,h)、DR(2*w/3-1,0,w/3,h)、DT(0,0,w,h/5)、DB(0,4 x h/5-1, w, h/5), wherein w and h respectively represent the car window image AWWidth and height of (1);
step 1.2: respectively finding four side lines of the upper side, the lower side, the left side and the right side of the car window in the corresponding detection areas by using any straight line detection method, and respectively recording the corresponding straight lines of the four side lines on the image A as LT、LR、LB、LL；
In the present example, the straight line detection method used is described as follows:
for the detected image X, the angular range of the target straight line is first given and marked as [ angleMin, angleMax](ii) a And performing Gaussian filtering, Canny edge detection and Hough line detection on the image X to obtain a line set L ═ Li|i＝0,1,…,NL-1}，NLThe number of all detected straight lines; finding the angle in the line set L to satisfy [ angleMin, angleMax]All the straight lines in the range, the straight line with the maximum length in the straight lines is the key straight line of the target in the image; wherein, the straight line liLength is a straight line liThe distance between the two farthest points of all the passing edge points;
step 1.3: determining the areas of the face, windows, roof, left body and right body of the vehicle according to the following rules:
f) the car face area: straight line L on image AR、LB、LLAny polygonal area surrounded by the bottom edge of the image A;
g) the window area is as follows: straight line L on image AR、LB、LL、LTA quadrilateral area is formed by surrounding;
h) roof area: straight line L on image AR、LB、LLAny polygonal area surrounded by the top edge of the image A;
i) left body area: straight line L on image ALAn arbitrary polygon region on the left side;
j) right body area: straight line L on image ARAny polygonal area on the right side;
the specific steps of acquiring the vehicle type reconstruction image in the step 4 are as follows:
step 4.1: for any vehicle image A, firstly, a vehicle image area segmentation method based on windows is utilized to carry out vehicle image area segmentation, and then corresponding vehicle face area images and vehicle body area images obtained on the image A are respectively marked as A1，A2(ii) a The vehicle body area image is an area image corresponding to a larger area in the left and right vehicle body areas;
step 4.2: if the body area image A2If the right vehicle body area corresponds to the left vehicle body area, horizontally turning the image;
step 4.3: image A1，A2Uniformly zooming to 200 px-200 px, and horizontally splicing according to the sequence from left to right to obtain a vehicle type reconstructed image corresponding to the vehicle A;
the specific steps of acquiring the vehicle reconstructed image in the steps 1 and 5 are as follows:
step 5.1: for any vehicle image A, firstly, a vehicle image area segmentation method based on windows is utilized to carry out vehicle image area segmentation, and then corresponding face area image, window area image, vehicle body area image and roof area image obtained on the image A are respectively marked as A1，A2，A3，A4(ii) a The vehicle body area image is an area image corresponding to a larger area in the left and right vehicle body areas;
step 5.2: if the body area image A3If the right vehicle body area corresponds to the left vehicle body area, horizontally turning the image;
step 5.3: image A1，A2，A3，A4And uniformly zooming to 200 px-200 px, and horizontally splicing according to the sequence from left to right to obtain a vehicle reconstructed image corresponding to the vehicle A.
In this example, for the vehicle image shown in fig. 3, the corresponding window positioning result, vehicle type reconstructed image, and vehicle reconstructed image are shown in fig. 4, 5, and 6, respectively.
Claims (5)
1. A three-stage vehicle retrieval method based on multiple features is characterized by comprising the following steps:
step 1: firstly, training to obtain a detection model and a classification model which are needed in a retrieval process, wherein the training comprises the following steps: the method comprises the following steps of (1) vehicle window detection model winDetect, vehicle color recognition model colorClassify based on a vehicle face region, depth feature extraction network model carTypefeature based on a vehicle type reconstructed image and depth feature extraction network model carFefeature based on a vehicle reconstructed image; the car face region is obtained according to a car window-based car region segmentation method;
the vehicle window-based vehicle region segmentation method in the step 1 comprises the following specific processes:
1.1): for any vehicle image A, firstly, a detection model winDetect is utilized to position the approximate position of a vehicle window, and the vehicle window image A is obtained through interceptionWDetermining a window image AWThe detection areas of the upper, lower, left and right sidelines of the upper vehicle window are respectively as follows: dL(0,0,w/3,h)、DR(2*w/3-1,0,w/3,h)、DT(0,0,w,h/5)、DB(0,4 x h/5-1, w, h/5), wherein w and h respectively represent the car window image AWWidth and height of (1);
1.2): respectively finding the upper part, the lower part and the left part of the car window in the corresponding detection areas by utilizing a straight line detection methodAnd the four right edge lines, and respectively recording the corresponding straight lines of the four right edge lines on the image A as LT、LR、LB、LL；
1.3): determining the areas of the face, windows, roof, left body and right body of the vehicle according to the following rules:
a) the car face area: straight line L on image AR、LB、LLAny polygonal area surrounded by the bottom edge of the image A;
b) the window area is as follows: straight line L on image AR、LB、LL、LTA quadrilateral area is formed by surrounding;
c) roof area: straight line L on image AR、LB、LLAny polygonal area surrounded by the top edge of the image A;
d) left body area: straight line L on image ALAn arbitrary polygon region on the left side;
e) right body area: straight line L on image ARAny polygonal area on the right side;
step 2: recording the searched target vehicle as I, and recording the vehicle set in the vehicle search base as X ═ X0,x1,...,xn-1}；
And step 3: identifying the colors of a target vehicle I and all vehicles in a vehicle set X by using a vehicle color identification model colorClassification based on a vehicle face area, then finding the vehicles with the same color as the target vehicle I from the X to form a vehicle set S1；
And 4, step 4: obtaining a target vehicle I and a vehicle set S1Corresponding reconstructed image of vehicle type, marked as ITAnd S1 TExtracting I by using a car type reconstructed image-based depth network model carTypeFeatureTAnd S1 TFinding out a vehicle set S by using the corresponding vehicle type feature vector1The Euclidean distance between all the vehicle type characteristic vectors meeting the requirements of the corresponding vehicle type characteristic vector and the vehicle type characteristic vector I is less than lambda1Constitute a new vehicle set, denoted S2(ii) a Wherein λ1The maximum Euclidean distance threshold value of the preset same-model vehicle in the feature space is obtained;
and 5: obtain the target vehicleVehicle I and vehicle set S2Corresponding reconstructed image of vehicle, noted as ICAnd S2 CExtracting I by using a depth network model carFeature based on a vehicle type reconstructed imageCAnd S2 CFinding out a vehicle set S according to the corresponding vehicle feature vector2The Euclidean distance between all the vehicle characteristic vectors meeting the requirement of corresponding vehicle characteristic vector and I is less than lambda2Constitute a new vehicle set, denoted S3(ii) a Wherein λ2A predetermined maximum euclidean distance threshold in feature space for the same vehicle.
2. The multi-feature-based three-stage vehicle retrieval method according to claim 1, wherein the specific steps of obtaining the vehicle type reconstruction image in the step 4 are as follows:
4.1): for any vehicle image A, firstly, a vehicle image area segmentation method based on windows is utilized to carry out vehicle image area segmentation, and then corresponding vehicle face area images and vehicle body area images obtained on the image A are respectively marked as A1，A2(ii) a The vehicle body area image is an area image corresponding to a larger area in the left and right vehicle body areas;
4.2): if the body area image A2If the right vehicle body area corresponds to the left vehicle body area, horizontally turning the image;
4.3): image A1，A2And uniformly zooming to 200 px-200 px, and horizontally splicing according to the sequence from left to right to obtain a vehicle type reconstructed image corresponding to the vehicle A.
3. The three-stage vehicle retrieval method based on multiple features of claim 1, wherein the specific process of acquiring the vehicle reconstruction images in the steps 1 and 5 is as follows:
5.1): for any vehicle image A, firstly, a vehicle image area segmentation method based on windows is utilized to carry out vehicle image area segmentation, and then corresponding face area image, window area image, vehicle body area image and roof area image obtained on the image A are respectively marked as A1，A2，A3，A4(ii) a The vehicle body area image is an area image corresponding to a larger area in the left and right vehicle body areas;
5.2): if the body area image A3If the right vehicle body area corresponds to the left vehicle body area, horizontally turning the image;
5.3): image A1，A2，A3，A4And uniformly zooming to 200 px-200 px, and horizontally splicing according to the sequence from left to right to obtain a vehicle reconstructed image corresponding to the vehicle A.
4. The multi-feature based three-stage vehicle retrieval method according to claim 1, wherein in the step 1,
vehicle window detection model winDetect: training based on a yolo detection network;
vehicle color recognition model colorClassify based on the vehicle face area: the method comprises the steps that the method is obtained based on training of an SVM classifier, the used features are statistic HSV color histogram features of a car face region, the statistic HSV color histogram features comprise 36-dimensional H components, 20-dimensional S components and 20-dimensional V components, and the SVM classifier identifies six colors of black, white, red, yellow, blue and green cars;
extracting a network model carTypeFeature based on the depth characteristics of a vehicle type reconstructed image: the network model structure for feature extraction consists of a classification network and a dual-input image matching network, the basic network structure for forming the network model structure is cafeNet, and the dimensionality of the network extraction features is 2048; the training data set of the car type feature extraction network model based on the depth features of the vehicle type reconstructed image is a fine category vehicle type data set;
extracting a network model carFeature based on the depth feature of the vehicle type reconstructed image: the network model is the same as a car type feature extraction network model carTypefeature structure based on a car type reconstructed image, and the training data set of the car type reconstructed image based depth feature extraction network model carFefeature is a vehicle re-identification data set.
5. The multi-feature based three-stage vehicle retrieval method according to claim 1, wherein the straight line detection method is as follows:
1) for a detection image X, firstly, giving an angle range of a target straight line, and marking the angle range as [ angleMin, anglemx ];
2) and performing Gaussian filtering, Canny edge detection and Hough line detection on the image X to obtain a line set L ═ LiI |, 0,1, …, NL-1}, NL being the number of all lines detected;
3) finding the angle in the line set L to satisfy [ angleMin, angleMax]All the straight lines in the range, the straight line with the maximum length in the straight lines is the key straight line of the target in the image; wherein, the straight line liLength is a straight line liAll edge points that pass by are the distance of the two points that are farthest apart.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810996825.1A CN109255052B (en) | 2018-08-29 | 2018-08-29 | Three-stage vehicle retrieval method based on multiple features |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810996825.1A CN109255052B (en) | 2018-08-29 | 2018-08-29 | Three-stage vehicle retrieval method based on multiple features |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109255052A CN109255052A (en) | 2019-01-22 |
| CN109255052B true CN109255052B (en) | 2022-03-25 |
Family
ID=65049406
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810996825.1A Active CN109255052B (en) | 2018-08-29 | 2018-08-29 | Three-stage vehicle retrieval method based on multiple features |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109255052B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110046579B (en) * | 2019-04-18 | 2023-04-07 | 重庆大学 | Deep Hash pedestrian re-identification method |
| CN110503597B (en) * | 2019-07-09 | 2023-05-16 | 浙江浩腾电子科技股份有限公司 | Vehicle viewpoint normalization method based on 2D structure model |
| CN115858846B (en) * | 2023-02-16 | 2023-04-21 | 云南派动科技有限公司 | Skier image retrieval method and system based on deep learning |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105404874A (en) * | 2015-11-27 | 2016-03-16 | 成都神州数码索贝科技有限公司 | Vehicle window recognition system based on projection and hough straight line detection |
| CN106326893A (en) * | 2016-08-25 | 2017-01-11 | 安徽水滴科技有限责任公司 | Vehicle color recognition method based on area discrimination |
| CN107437062A (en) * | 2017-06-27 | 2017-12-05 | 浙江工业大学 | A kind of multi-direction vehicle rough localization method of still image |
| CN107463939A (en) * | 2017-06-27 | 2017-12-12 | 浙江工业大学 | A kind of image key line detection method |
| CN108133170A (en) * | 2017-10-31 | 2018-06-08 | 浙江浩腾电子科技股份有限公司 | A kind of vehicle window localization method of multi-direction vehicle |
-
2018
- 2018-08-29 CN CN201810996825.1A patent/CN109255052B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105404874A (en) * | 2015-11-27 | 2016-03-16 | 成都神州数码索贝科技有限公司 | Vehicle window recognition system based on projection and hough straight line detection |
| CN106326893A (en) * | 2016-08-25 | 2017-01-11 | 安徽水滴科技有限责任公司 | Vehicle color recognition method based on area discrimination |
| CN107437062A (en) * | 2017-06-27 | 2017-12-05 | 浙江工业大学 | A kind of multi-direction vehicle rough localization method of still image |
| CN107463939A (en) * | 2017-06-27 | 2017-12-12 | 浙江工业大学 | A kind of image key line detection method |
| CN108133170A (en) * | 2017-10-31 | 2018-06-08 | 浙江浩腾电子科技股份有限公司 | A kind of vehicle window localization method of multi-direction vehicle |
Non-Patent Citations (1)
| Title |
|---|
| 融合ViBe与帧差法的交叉路口多车辆检测方法;高飞 等;《计算机测量与控制》;20171025;第35-38页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109255052A (en) | 2019-01-22 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Bai et al. | Group-sensitive triplet embedding for vehicle reidentification | |
| CN105046196B (en) | Front truck information of vehicles structuring output method based on concatenated convolutional neutral net | |
| Wang et al. | An effective method for plate number recognition | |
| CN105740886B (en) | A kind of automobile logo identification method based on machine learning | |
| Gonçalves et al. | License plate recognition based on temporal redundancy | |
| Malik et al. | Detection and recognition of traffic signs from road scene images | |
| CN109255052B (en) | Three-stage vehicle retrieval method based on multiple features | |
| Prates et al. | Brazilian license plate detection using histogram of oriented gradients and sliding windows | |
| Shah et al. | OCR-based chassis-number recognition using artificial neural networks | |
| Chantakamo et al. | The multi vehicle recognition using hybrid blob analysis and feature-based | |
| Abedin et al. | Traffic sign recognition using surf: Speeded up robust feature descriptor and artificial neural network classifier | |
| Ellahyani et al. | Traffic sign detection for intelligent transportation systems: a survey | |
| Yao et al. | Coupled multivehicle detection and classification with prior objectness measure | |
| Asgarian Dehkordi et al. | Vehicle type recognition based on dimension estimation and bag of word classification | |
| Aarthi et al. | Vehicle detection in static images using color and corner map | |
| Thumthong et al. | Automatic detection and recognition of Thai vehicle license plate from CCTV images | |
| Symeonidis et al. | Neural attention-driven non-maximum suppression for person detection | |
| Mirza et al. | Urdu caption text detection using textural features | |
| Juntanasub et al. | Car license plate recognition through Hausdorff distance technique | |
| Vinoharan et al. | A wheel-based side-view car detection using snake algorithm | |
| Jin et al. | A method for traffic signs detection, tracking and recognition | |
| Chen et al. | Context-aware lane marking detection on urban roads | |
| Syed et al. | Color edge enhancement based fuzzy segmentation of license plates | |
| Sushma et al. | Text detection in color images | |
| Wang et al. | Traffic lights recognition based on PCANet |
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 | ||
| GR01 | Patent grant |