TWI755669B - Homogeneous vehicle retrieval system and method - Google Patents

Abstract

The invention discloses homogeneous vehicle retrieval system and method. First, a vehicle to be retrieved is obtained from vehicle pictures with different perspectives to uniformly align directions and angles of the vehicles to be retrieved to achieve normalization, and feature points are extracted from the vehicles to be retrieved. Then, feature points of the vehicle to be retrieved are matched with feature points of a vehicle in vehicle pictures of a database to calculate the feature points of the vehicle to be retrieved, and the feature points of the vehicle in the vehicle pictures of the database are successfully matched number of the feature points as similarity score. Final, a voting mechanism of the similarity score is used to count number of votes of the vehicles in the vehicle pictures of the database, and the vehicle with the higher or highest number of votes is selected from the vehicle pictures as the same type or similar vehicle of the vehicle to be retrieved.

Description

Similar vehicle retrieval system and method

The present invention relates to a same-type vehicle retrieval technology, in particular to a same-type vehicle retrieval system and method.

For the detection of road traffic information, in recent years, many cities have gradually adopted technological law enforcement to replace the previous manpower support, which greatly reduces the need for police manpower to manually review surveillance images and use the naked eye to search for criminal vehicles in the past. It takes a lot of effort and labor to get out of the criminal vehicle.

In the process of tracing a vehicle, the license plate is usually the unique identity of the vehicle as the basis for tracking, but the license plate may be difficult to identify in some cases. For example, factors such as images with poor resolution, or the license plate being out of bounds or being cropped may cause the number of the license plate to be unclear, and even those with intentions will cover, dismantle, forge or replace the license plate in advance when committing a crime. Make it impossible for the police to track down the information of the license plate.

In addition to using the license plate to identify the vehicle, each vehicle can also distinguish the difference of a vehicle by the appearance and characteristics of the vehicle, so as to find a specific target or a vehicle of the same model, so the vehicle retrieval technology can be widely used in technology law enforcement and Smart transportation and other fields.

In addition, the vehicle retrieval technology is mainly for the same vehicle picture, to find all the vehicles of the same model in the scene images of different times, places, and viewing angles, but generally speaking, in hundreds of surveillance images, it is necessary to search for a series of external images. Identifying a particular vehicle among similar types of vehicles is quite time consuming and challenging.

Therefore, how to provide a novel or innovative same-type vehicle retrieval technology to improve the accuracy of vehicle identification of the same-type vehicle has become a major research topic for those skilled in the art.

The present invention provides a novel or innovative same-type vehicle retrieval system and method, which can improve the reliability of the same-type vehicle or similar vehicle retrieval, or quickly track the correct vehicle to improve the efficiency, thereby improving the vehicle identification accuracy rate of the same-type vehicle.

The same-type vehicle retrieval system of the present invention includes: an adaptive vehicle normalization module, which captures the vehicle to be retrieved in the vehicle pictures with different viewing angles, so as to compare the direction and angle of the vehicle to be retrieved in the vehicle pictures with different viewing angles. Uniform alignment to achieve normalization; a feature point extraction module, which extracts feature points of the vehicle to be retrieved from the complex vehicle images of different perspectives normalized by the adaptive vehicle normalization module; a recognition and similarity calculation module. group, which is to match the feature points of the vehicle to be retrieved extracted by the feature point extraction module with the feature points of the vehicle in the vehicle image in the database to calculate the feature points of the vehicle to be retrieved and the vehicle image in the database. The feature points of the two vehicles have the number of successfully paired feature points as the similarity score; and a voting mechanism module, which uses the similarity score calculated by the identification and similarity calculation module to perform a voting mechanism to count the database. The number of votes of the vehicle in the vehicle picture, and then select the vehicle with the higher or highest number of votes from the plurality of vehicle pictures in the database as the same type or similar vehicle of the vehicle to be retrieved vehicle.

The method for retrieving similar vehicles of the present invention includes: extracting vehicles to be retrieved in a plurality of vehicle pictures with different viewing angles by an adaptive vehicle normalization module, so as to unify the directions and angles of the vehicles to be retrieved in the plurality of vehicle pictures with different viewing angles Alignment to achieve normalization; a feature point extraction module extracts feature points of the vehicle to be retrieved from the complex vehicle images of different perspectives normalized by the adaptive vehicle normalization module; an identification and similarity calculation module is used. Match the feature points of the vehicle to be retrieved extracted by the feature point extraction module with the feature points of the vehicle in the vehicle image of the database to calculate the feature points of the vehicle to be retrieved and the vehicle in the vehicle image of the database. The two feature points have the number of successfully paired feature points as the similarity score; and a voting mechanism module performs a voting mechanism on the similarity score calculated by the identification and similarity calculation module to count the vehicle images in the database. The number of votes obtained by the vehicle, and then the vehicle with the higher or highest number of votes is selected from the plurality of vehicle pictures in the database as the vehicle of the same type or similar vehicle of the vehicle to be retrieved.

In order to make the above-mentioned features and advantages of the present invention more obvious and easy to understand, the following specific embodiments are given and described in detail with the accompanying drawings. Additional features and advantages of the present invention will be set forth in part in the description that follows, and in part will be apparent from the description, or may be learned by practice of the invention. The features and advantages of the invention will be realized and attained by means of the elements and combinations particularly pointed out in the scope of the patent application. It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not intended to limit the scope of the invention as claimed.

1‧‧‧Identical vehicle retrieval system

10‧‧‧Adaptive Vehicle Normalization Module

20‧‧‧Feature point extraction module

30‧‧‧Recognition and similarity calculation module

40‧‧‧Voting Mechanism Module

50‧‧‧Reasonable vehicle selection module

60‧‧‧Database

A, B‧‧‧vehicle pictures

C‧‧‧Results for same or similar vehicles

Steps S1 to S7‧‧‧

Figure 1 is a schematic diagram of the structure of the same-type vehicle retrieval system of the present invention;

Figure 2 is a schematic diagram of a vehicle picture stored in the database of the present invention;

Fig. 3 is a schematic flow chart of the method for retrieving similar vehicles of the present invention;

FIG. 4 is a schematic diagram of a plurality of vehicle pictures from different viewing angles of the vehicle to be retrieved in the present invention;

Fig. 5 is a schematic diagram of removing the background of the vehicle from the vehicle picture in the present invention;

Fig. 6 to Fig. 8 are schematic diagrams of extracting the vehicle contour and aligning the direction and angle of the vehicle uniformly in the present invention;

Fig. 9 is a schematic diagram of extracting feature points of vehicles in the present invention;

FIG. 10 is a schematic diagram of matching the feature points of the vehicle to be retrieved with the feature points of the vehicle in the database according to the present invention;

FIG. 11 is a schematic diagram of comparing the vehicle pictures of different viewing angles of the vehicle to be retrieved with the different vehicle pictures of the database in a many-to-many manner according to the present invention;

Fig. 12 is a schematic diagram of selecting the vehicle with the highest number of votes as the vehicle of the same type to be retrieved in the present invention; and

FIG. 13 is a schematic diagram of the result of masking the license plates of a plurality of vehicle pictures to identify the same type of vehicles or similar vehicles according to the present invention.

The embodiments of the present invention are described below with specific specific embodiments. Those skilled in the art can understand other advantages and effects of the present invention from the contents disclosed in this specification, and can also be implemented by other different specific equivalent embodiments. or apply.

FIG. 1 is a schematic diagram of the structure of the same-type vehicle retrieval system 1 of the present invention, and the same The main technical contents of the model vehicle retrieval system 1 are as follows, and the rest of the contents are the same as those in Fig. 2 to Fig. 13, and will not be repeated here. As shown in FIG. 1, the same-type vehicle retrieval system 1 may include an adaptive vehicle normalization module 10, a feature point extraction module 20, an identification and similarity calculation module 30, and a voting mechanism that communicate or connect with each other. The module 40 , a reasonable vehicle selection module 50 and a database 60 .

For example, the adaptive vehicle normalization module 10 can be a software adaptive vehicle normalization program, etc., and the feature point extraction module 20 can be a software feature point extraction program, a hardware extractor, etc. The group 30 may be a software identification and similarity calculation program, a hardware identifier or calculator, etc., the voting mechanism module 40 may be a software voting mechanism program, etc., and the reasonable vehicle selection module 50 may be a software reasonable vehicle selection Programs, hardware selectors, etc. The database 60 can be or represent various data storage media, such as hard disk (cloud drive), server (data server), storage, memory, and the like. The vehicle may be an automobile, a locomotive, a tour bus, a truck, a van, or the like. However, the present invention is not limited to this.

[1] Adaptive vehicle normalization module 10: It automatically extracts an optimized vehicle image from, for example, a monitoring image (monitoring screen) or a vehicle image A provided by the user, so as to improve subsequent identification and similarity calculation The recognition rate of the module 30. That is, the adaptive vehicle normalization module 10 can use the deep learning technology to accurately crop a plurality of vehicle pictures A with different viewing angles to remove the background of the vehicle (the vehicle to be retrieved) in the vehicle picture A, and the deep learning technology can be, for example, Darknet's Yolo (You Only Look Once; you only need to look once) algorithm can accurately detect and cut out multiple vehicle pictures A for the vehicle position in the surveillance image (monitoring screen), as shown in Figures 4 to 5 as shown in the figure.

In addition, since the angles and directions of the cropped plural vehicle pictures A are not the same, so The adaptive vehicle normalization module 10 can align the plurality of vehicle pictures A uniformly. In other words, the adaptive vehicle normalization module 10 can use an edge detection algorithm (such as the Sobel edge detection algorithm or the Canny edge detection algorithm) to convert the background-removed vehicles in the vehicle pictures A from multiple different viewing angles. (To retrieve the vehicle) perform edge detection to detect the vehicle outline, and then obtain the horizontal and vertical directions of the vehicle (to retrieve the vehicle) through the vehicle outline. The adaptive vehicle normalization module 10 can also use the Hough transform method to obtain the longest horizontal line (front or rear) and longest vertical line (the left or right side of the vehicle body) of the vehicle (the vehicle to be retrieved) through the vehicle outline to calculate the longest The angle between the horizontal line and the longest vertical line, and then use the angle between the longest horizontal line and the longest vertical line to calculate the direction and angle of the vehicle (the vehicle to be retrieved), so as to align multiple vehicle pictures A from different perspectives, as shown in Section 6 Figure to Figure 8.

[2] Feature point extraction module 20 : extracts feature points from the vehicle normalized by the adaptive vehicle normalization module 10 , as shown in FIG. 9 . This feature point has the invariant characteristics of the vehicle's position, scale, and rotation, and also maintains a certain degree of stability against the vehicle's viewing angle change, mirror transformation, and noise, and is suitable for fast, rapid, exact match.

[3] Recognition and similarity calculation module 30: to identify and calculate the similarity between the vehicle to be retrieved and the vehicle group in the vehicle picture B of the database 60, and the identification and similarity calculation module 30 can The feature points of the retrieved vehicle are matched with the feature points of the vehicle in the vehicle picture B of the database 60 , as shown in FIG. 10 . Furthermore, the matching of the feature points by the identification and similarity calculation module 30 is performed by calculating the multi-dimensional feature points ( Such as the Euclidean distance of 128-dimensional feature points), when the Euclidean distance between the two feature points is smaller, the similarity between the two feature points is higher, and when the two feature points are When the Euclidean distance of the points is less than the set threshold, it can be determined that the two feature points are paired successfully. At the same time, the identification and similarity calculation module 30 can calculate the number of successfully paired feature points, and use the number of successfully matched feature points as the similarity score. When the similarity score is greater than the similarity threshold value (for example, the set similarity threshold value), the identification and similarity calculation module 30 can preliminarily determine that the vehicle to be retrieved and the vehicle in the vehicle picture B of the database 60 are paired successfully.

[4] Voting mechanism module 40 : compares a plurality of vehicle pictures A with different viewing angles of the vehicle to be retrieved and a plurality of different vehicle pictures B in the database 60 in a many-to-many manner (eg, batch comparison), and The voting mechanism module 40 can perform a voting mechanism on the similarity scores calculated by the identification and similarity calculation module 30 to count the number of votes for the vehicles in the vehicle pictures B of the database 60 , and then obtain the number of votes from the plurality of vehicle pictures in the database 60 . In B, the vehicle with the highest or highest number of votes is selected as the same type of vehicle or similar vehicle of the vehicle to be retrieved, so as to improve the reliability of the retrieval of the same type of vehicle or similar vehicle, as shown in Figure 11.

[5] Reasonable vehicle selection module 50: It selects the vehicle with the highest number of votes from the results of the voting mechanism of the voting mechanism module 40 as the vehicle of the same type or similar vehicle of the vehicle to be retrieved, so as to select the vehicle of the same type as the vehicle to be retrieved. or similar vehicle result C is displayed on the display (not shown), as shown in Figures 12 to 13.

Therefore, the present invention provides a plurality of vehicle pictures A, and adds an adaptive vehicle normalization module 10 to automatically crop the vehicle pictures A and align the directions of the vehicles uniformly, so that the multi-directional or multi-perspective vehicle pictures A need not be manually selected. , that is, the optimized vehicle picture A can be extracted. In addition, the present invention can be matched with the feature point extraction module 20, the identification and similarity calculation module 30, the voting mechanism module 40, and the reasonable vehicle selection module 50 to select the vehicle with the highest or highest number of votes as the vehicle to be retrieved. Result C for the same or similar vehicles, thus boosting the same or similar vehicles Reliability of retrieval.

Therefore, the present invention can accurately retrieve the same type of vehicles or similar vehicles that have appeared in the surveillance image (monitoring screen) of the road, and can still quickly retrieve a large number of vehicles from a large number of vehicles even when the license plate is masked, dismantled, forged or replaced. Identify the suspected vehicle by finding the similar vehicle with the highest number of votes in the image. At the same time, the present invention can assist the police or related agencies to find out the same type of vehicle or similar vehicle of the vehicle to be retrieved from many vehicle pictures, so as to quickly track the correct vehicle to improve efficiency.

FIG. 2 is a schematic diagram of the vehicle picture B stored in the database 60 of FIG. 1 of the present invention. For example, the database 60 can collect 100 vehicles from the surveillance images (monitoring images) of a certain intersection, and each vehicle takes 2 vehicle pictures B from different perspectives, so that the database 60 stores a total of 200 vehicle pictures B for the actual search.

Fig. 3 is a schematic flowchart of the method for retrieving similar vehicles of the present invention, Fig. 4 is a schematic diagram of a plurality of vehicle pictures from different viewing angles of the vehicle to be retrieved in the present invention, and Fig. 5 is the background of the present invention to remove the vehicle from the vehicle picture Fig. 6 to Fig. 8 are schematic diagrams of extracting the vehicle outline and uniformly aligning the direction and angle of the vehicle in the present invention, Fig. 9 is a schematic diagram of extracting the feature points of the vehicle in the present invention, and Fig. 10 is the present invention. A schematic diagram of matching the feature points of the vehicle to be retrieved with the feature points of the vehicle in the database, FIG. 11 is a many-to-many way of comparing the vehicle pictures of the vehicle to be retrieved from different perspectives with the different vehicle pictures in the database in a many-to-many manner. Fig. 12 is a schematic diagram of the vehicle with the highest number of votes being selected as a vehicle of the same type to be retrieved in the present invention, and Fig. 13 is a result of the present invention masking the license plates of a plurality of vehicle pictures to identify a vehicle of the same type or a similar vehicle schematic diagram.

Meanwhile, the method for retrieving similar vehicles in FIG. 3 may include the following steps S21 to S27 , which will be described with reference to FIGS. 1 and 4 to 13 .

As shown in step S1 in Fig. 3 and Fig. 4, first obtain a plurality of (such as 5) vehicle pictures A (see Fig. 1) with different viewing angles of the vehicle to be retrieved, and then combine the plural (such as 5) vehicle pictures A with Upload to the same type vehicle retrieval system 1 to prepare for retrieval.

As shown in step S2 in Fig. 3 and Fig. 5, the adaptive vehicle normalization module 10 in Fig. 1 can use deep learning technology, such as Darknet's Yolo (you only need to see it once) algorithm or further apply CoCo ( Common Objects in Context; common objects in context) training set to detect the vehicle position in multiple (such as 5) vehicle pictures A from different perspectives, so as to accurately extract the vehicle position from the vehicle picture A according to the vehicle position through deep learning technology Cut out the vehicle (to be retrieved) and remove the background of the vehicle.

As shown in step S3 of FIG. 3 and FIGS. 6 to 8 , the adaptive vehicle normalization module 10 can use an edge detection algorithm (such as the Sobel edge detection algorithm or the Canny edge detection algorithm), Perform edge detection on the vehicle with the background removed (the vehicle to be retrieved) in the vehicle image A to detect the vehicle outline, and use the Hough transform method to obtain the longest horizontal line (front or rear) and longest vertical line of the vehicle through the vehicle outline (the left or right side of the vehicle body), and then use the angle between the longest horizontal line and the longest vertical line to calculate the direction and angle of the vehicle (the vehicle to be retrieved), so that the plural numbers (such as 5) are different according to the direction and angle of the vehicle to be retrieved. The vehicle image A of the viewing angle is uniformly aligned.

In detail, in FIGS. 6 to 7 , the adaptive vehicle normalization module 10 can capture the vehicle to be retrieved in the vehicle pictures A with different viewing angles, so as to convert the desired vehicle in the vehicle pictures A with different viewing angles. The orientation and angle of the retrieved vehicle are uniformly aligned to achieve normalization. In addition, since the vehicle is basically an object similar to a cuboid, the adaptive vehicle normalization module 10 can use an edge detection algorithm (such as the Sobel edge detection algorithm or the Canny edge detection algorithm) to obtain the vehicle outline, and Using the Hough transform method to obtain the longest horizontal line and The longest vertical line, and then calculate the direction of the vehicle (the vehicle to be retrieved) based on the slope of the longest vertical line, and then use the mirror method to unify the direction of the vehicle in one direction (such as right or left).

In addition, in Fig. 8, the adaptive vehicle normalization module 10 can obtain the included angles of all the vehicles through the longest vertical line and the longest horizontal line of the multiple (eg five) vehicles processed by the mirroring method, and then calculate all the vehicles. Average the included angles of the vehicles, and then make a matrix fine-tuning for multiple vehicles (such as five vehicles) at the average angle, so that the angles of all vehicles are uniformly aligned.

As shown in step S4 of FIG. 3 and FIG. 9 , the feature point extraction module 20 can extract feature points of the vehicle to be retrieved from the plurality of vehicle images A with different viewing angles normalized by the adaptive vehicle normalization module 10 . That is, the feature point extraction module 20 can extract unique and stable feature points (such as as many as Fig. 9) from the vehicle whose background has been removed (the vehicle to be retrieved) through the Speeded Up Robust Features (SURF) algorithm. A circle ○ represents a plurality of feature points), and a large number of stable feature points can also be obtained from vehicles whose background has been removed (vehicles to be retrieved) by adjusting the Hessian matrix threshold. For example, when the threshold value of the Hessian matrix is set to 400, the number of feature points that can be extracted from each vehicle is about 100 to 150 points.

As shown in step S5 of FIG. 3 and FIG. 10 , the identification and similarity calculation module 30 can compare the feature points of the vehicle to be retrieved extracted by the feature point extraction module 20 and the vehicle in the vehicle picture B of the database 60 Match the two feature points of the vehicle to be retrieved to calculate the Euclidean distance between the feature points of the vehicle to be retrieved and the feature points of the vehicle in the vehicle picture B of the database 60 to find the feature points that are paired successfully, and match them successfully. The number of feature points is taken as the similarity score. Next, the identification and similarity calculation module 30 can determine whether the similarity score is higher than the similarity threshold to preliminarily determine whether the vehicle to be retrieved and the vehicle in the vehicle picture B of the database 60 are the same type or similar vehicle. For example, if the similarity threshold is set to 20, the conditions for successful vehicle identification are: When the similarity score of the two vehicles (that is, the vehicle to be retrieved and the vehicle in the vehicle picture B in the database 60 ) is greater than 20, the identification and similarity calculation module 30 can preliminarily determine the vehicle in the vehicle picture B in the database 60 . The vehicle may be of the same type or similar to the vehicle being retrieved.

As shown in step S6 in FIG. 3 and FIG. 11 , the voting mechanism module 40 can combine a plurality of (eg, 5) vehicle pictures A with different viewing angles of the vehicle to be retrieved and a plurality of different vehicle pictures B in the database 60 in pairs The comparison is performed in multiple ways, and the above step S5 is performed in batches, and the voting mechanism module 40 can perform a voting mechanism on the similarity scores calculated by the identification and similarity calculation module 30 to count the number of vehicle images in the database 60 . According to the number of votes of the vehicle, the vehicle with the higher or highest vote number is selected from the plurality of vehicle pictures B in the database 60 as the same type or similar vehicle of the vehicle to be retrieved, thereby improving the retrieval reliability of the same type or similar vehicle.

As shown in step S7 of FIG. 3 and FIGS. 12 to 13, the vehicle with a higher number of votes selected by the voting mechanism module 40 from step S6 may be a vehicle of the same type or similar to the vehicle to be retrieved, so it is reasonable The vehicle selection module 50 can select the vehicle with the highest number of votes from the results of the voting mechanism of the voting mechanism module 40 as the final vehicle of the same type or similar vehicle to be retrieved, so as to compare the results of the same type of vehicle or similar vehicle of the vehicle to be retrieved. C is displayed on the display (not shown).

For example, in Figure 12, multiple (eg, 5) vehicle pictures A from different perspectives have been successfully identified as one of the vehicles in the database 60, which means that the vehicle in the database 60 is the same type of vehicle to be retrieved or similar vehicles. In addition, in Fig. 13, the license plates of the above-mentioned plural (for example, 5) vehicle pictures A are actually masked to simulate the actual situation that a person with intentions covers and removes the license plates. Through the present invention, the vehicle of the same type can still be successfully identified. or result C for similar vehicles.

To sum up, the system and method for retrieving similar vehicles of the present invention can at least have the following features: advantage or technical efficacy.

1. The present invention provides a plurality of vehicle pictures, and adds an adaptive vehicle normalization module to automatically cut the vehicle pictures and align the direction of the vehicle uniformly. It can extract the multi-directional or multi-view vehicle pictures without manual selection. Generate optimized vehicle images.

2. The present invention can be matched with a feature point extraction module, an identification and similarity calculation module, a voting mechanism module, and a reasonable vehicle selection module, so as to select the vehicle with the highest or highest number of votes as the vehicle of the same type or similar to the vehicle to be retrieved. vehicle results, and improve the reliability of the retrieval of the same or similar vehicles.

3. The present invention can accurately retrieve, for example, the same type of vehicle or similar vehicle that has appeared in the surveillance image, and can quickly find the highest number of votes from a large number of vehicle pictures even when the license plate is masked, dismantled, forged or replaced, etc. similar vehicles to identify suspected vehicles.

4. The present invention can assist the police or related agencies to find out the same type of vehicle or similar vehicle of the vehicle to be retrieved from many vehicle pictures, so as to quickly track the correct vehicle to improve efficiency.

5. The present invention is an important technology of an intelligent security protection system, and can effectively reduce the problems of excessive dependence on manpower or huge cost in the traditional security protection field.

6. The present invention can be used by the police or related institutions in scientific and technological law enforcement, and can accurately and effectively control public security and safety protection, so as to greatly save police force or law enforcement manpower.

7. The industries that the present invention may be applied to are, for example, transportation, vehicles, or public security-related industries, and the products that may be applied are, for example, smart security protection systems or technology law enforcement systems.

The above-mentioned embodiments are only illustrative of the principles, features and effects of the present invention, and are not intended to limit the applicable scope of the present invention. Modifications and changes are made to the implementation form. any use of this Equivalent changes and modifications made by the disclosed contents of the invention shall still be covered by the scope of the patent application. Therefore, the scope of protection of the right of the present invention should be listed in the scope of the patent application.

1‧‧‧Identical vehicle retrieval system

10‧‧‧Adaptive Vehicle Normalization Module

20‧‧‧Feature point extraction module

30‧‧‧Recognition and similarity calculation module

40‧‧‧Voting Mechanism Module

50‧‧‧Reasonable vehicle selection module

60‧‧‧Database

A, B‧‧‧vehicle pictures

C‧‧‧Results for same or similar vehicles

Claims (16)

A vehicle retrieval system of the same type, comprising: an adaptive vehicle normalization module, which captures a vehicle to be retrieved in a plurality of vehicle pictures with different viewing angles, so as to match the direction of the vehicle to be retrieved in the vehicle pictures of the plurality of different viewing angles with the direction of the vehicle to be retrieved. The angles are uniformly aligned to achieve normalization; a feature point extraction module extracts feature points of the vehicle to be retrieved in the plurality of vehicle images of different viewing angles normalized by the adaptive vehicle normalization module; an identification and The similarity calculation module matches the feature points of the vehicle to be retrieved extracted by the feature point extraction module with the feature points of the vehicle in the vehicle pictures in the database to calculate the feature points of the vehicle to be retrieved The number of feature points that are successfully paired with the feature points of the vehicle in the vehicle pictures in the database is used as a similarity score; and a voting mechanism module is used for the identification and similarity calculation module. The similarity score uses a voting mechanism to count the number of votes of the vehicle in the vehicle pictures in the database, and then select the vehicle with the higher or highest number of votes from the plurality of vehicle pictures in the database as the vehicle of the same type of the vehicle to be retrieved or similar vehicles, wherein the feature point extraction module extracts unique and stable feature points from the vehicle to be retrieved from which the background has been removed by accelerating the robust feature algorithm, or by adjusting the Hessian matrix threshold value from the removed vehicle. A large number of stable feature points are obtained from the vehicle to be retrieved in the background. The same-type vehicle retrieval system as described in claim 1, wherein the adaptive vehicle normalization module further utilizes deep learning technology to crop the plurality of vehicle images from different viewing angles to remove the vehicle to be retrieved in the vehicle image background, and through the edge detection algorithm, edge detection is performed on the vehicle to be retrieved from which the background has been removed from the plurality of vehicle images from different viewing angles. The vehicle contour is detected, and then the horizontal and vertical directions of the vehicle to be retrieved are obtained from the vehicle contour. The same-type vehicle retrieval system as described in claim 2, wherein the adaptive vehicle normalization module utilizes the Yolo (you only need to look once) algorithm in the deep learning technology or further applies CoCo (in the context of general object) training set to detect the vehicle position in the plurality of vehicle images from different viewing angles, so as to cut out the vehicle to be retrieved from the vehicle image through the deep learning technology according to the vehicle position and remove the vehicle to be retrieved The background of the vehicle. The same-type vehicle retrieval system as described in claim 2, wherein the adaptive vehicle normalization module further utilizes the Hough transform method to obtain the longest horizontal line and the longest vertical line of the vehicle to be retrieved through the vehicle outline to calculate Calculate the angle between the longest horizontal line and the longest vertical line, and then use the angle between the longest horizontal line and the longest vertical line to calculate the direction and angle of the vehicle to be retrieved. The plurality of vehicle images from different viewing angles are uniformly aligned. The same-type vehicle retrieval system as described in claim 1, wherein the identification and similarity calculation module further calculates both the feature points of the vehicle to be retrieved and the feature points of the vehicle in the vehicle pictures in the database The Euclidean distance between them is used to find the feature points that are paired successfully, and the number of successfully paired feature points is used as the similarity score, and then it is judged whether the similarity score is higher than the similarity threshold value to preliminarily determine the desired retrieval. Whether the vehicle and the vehicle in the database's vehicle pictures are of the same type or a similar vehicle. A same-type vehicle retrieval system, comprising: an adaptive vehicle normalization module, which captures a vehicle to be retrieved in a plurality of vehicle images with different viewing angles, so as to uniformly align the directions and angles of the vehicle to be retrieved in the vehicle images of different viewing angles to achieve normalization; a feature point extraction module, which extracts feature points of the vehicle to be retrieved in the plurality of vehicle images from different viewing angles normalized by the adaptive vehicle normalization module; an identification and similarity calculation module, which is The feature points of the vehicle to be retrieved extracted by the feature point extraction module are matched with the feature points of the vehicle in the vehicle pictures in the database to calculate the feature points of the vehicle to be retrieved and the vehicle pictures in the database. The number of the feature points of the vehicle that are paired successfully is used as a similarity score; and a voting mechanism module is used to perform a voting mechanism on the similarity score calculated by the identification and similarity calculation module to count The number of votes of the vehicle in the vehicle pictures in the database is obtained, and the vehicle with the higher or highest votes is selected from the plurality of vehicle pictures in the database as the vehicle of the same type or similar vehicle of the vehicle to be retrieved, wherein the identification and the similarity calculation module further calculates the Euclidean distance between the feature point of the vehicle to be retrieved and the feature point of the vehicle in the vehicle picture in the database to find the paired feature point, so as to match the The number of successful feature points is used as the similarity score, and then it is determined whether the similarity score is higher than the similarity threshold value to preliminarily determine whether the vehicle to be retrieved and the vehicle in the vehicle picture in the database are the same type of vehicle or similar vehicle. The same-type vehicle retrieval system as described in claim 1 or 6, wherein the voting mechanism module further compares the plurality of vehicle pictures from different viewing angles with the plurality of different vehicle pictures in the database in a many-to-many manner , and then the identification and the similarity calculated by the similarity calculation module The similarity score performs the voting mechanism to count the number of votes for the vehicle in the vehicle pictures in the database. The same-type vehicle retrieval system as described in item 1 or 6 of the scope of the patent application further includes a reasonable vehicle selection module, which selects the vehicle with the highest number of votes as the vehicle from the results of the voting mechanism of the voting mechanism module. The vehicle of the same type or the similar vehicle of the vehicle to be retrieved is displayed on the display with the result of the vehicle of the same type or the similar vehicle of the vehicle to be retrieved. A method for retrieving a vehicle of the same type, comprising: extracting a vehicle to be retrieved in a plurality of vehicle pictures of different viewing angles by an adaptive vehicle normalization module, so as to obtain the direction and angle of the vehicle to be retrieved in the vehicle pictures of the plurality of different viewing angles Uniform alignment to achieve normalization; a feature point extraction module extracts feature points of the vehicle to be retrieved in the plurality of vehicle images of different viewing angles normalized by the adaptive vehicle normalization module; The degree calculation module matches the feature points of the vehicle to be retrieved extracted by the feature point extraction module with the feature points of the vehicle in the vehicle pictures in the database to calculate the feature points of the vehicle to be retrieved and the data. The number of the feature points of the vehicle in the vehicle pictures in the library that are paired successfully is used as the similarity score; and a voting mechanism module performs the identification and the similarity score calculated by the similarity calculation module. A voting mechanism is used to count the number of votes of the vehicle in the vehicle pictures in the database, and then select the vehicle with the higher or highest number of votes from the plurality of vehicle pictures in the database as the vehicle of the same type or a similar vehicle of the vehicle to be retrieved, The feature point extraction module further extracts unique and stable feature points from the vehicle to be retrieved from which the background has been removed through an accelerated robust feature algorithm, or adjusts the Hessian matrix threshold value to extract unique and stable feature points from the vehicle from which the background has been removed. To retrieve a large number of stable feature points in the vehicle. The method for retrieving similar vehicles as described in item 9 of the scope of the patent application further comprises that the adaptive vehicle normalization module uses deep learning technology to crop the plurality of vehicle images from different viewing angles to remove the vehicle to be retrieved in the vehicle image. background, and through the edge detection algorithm, the vehicle to be retrieved whose background has been removed from the plurality of vehicle images from different viewing angles performs edge detection to detect the vehicle outline, and then obtains the vehicle outline through the vehicle outline. Horizontal and vertical directions. The same-type vehicle retrieval method as described in item 10 of the scope of application, further comprising using the Yolo (you only need to look once) algorithm in the deep learning technology by the adaptive vehicle normalization module or further applying CoCo (in the context of general object) training set to detect the vehicle position in the plurality of vehicle images from different viewing angles, so as to cut out the vehicle to be retrieved from the vehicle image through the deep learning technology according to the vehicle position and remove the vehicle to be retrieved The background of the vehicle. The method for retrieving similar vehicles as described in item 10 of the scope of application, further comprising obtaining the longest horizontal line and the longest vertical line of the vehicle to be retrieved by using the Hough transform method by the adaptive vehicle normalization module to calculate the Calculate the angle between the longest horizontal line and the longest vertical line, and then use the angle between the longest horizontal line and the longest vertical line to calculate the direction and angle of the vehicle to be retrieved. The plurality of vehicle images from different viewing angles are uniformly aligned. The method for retrieving similar vehicles as described in item 9 of the scope of application, further comprising calculating, by the identification and similarity calculation module, the feature points of the vehicle to be retrieved and the vehicle map of the database The Euclidean distance between the feature points of the vehicle in the film is used to find the feature points that are paired successfully, and the number of successfully paired feature points is used as the similarity score, and then it is judged whether the similarity score is higher than The similarity threshold is used to preliminarily determine whether the vehicle to be retrieved and the vehicle in the vehicle picture in the database are the same type of vehicle or a similar vehicle. A method for retrieving a vehicle of the same type, comprising: extracting a vehicle to be retrieved in a plurality of vehicle pictures of different viewing angles by an adaptive vehicle normalization module, so as to obtain the direction and angle of the vehicle to be retrieved in the vehicle pictures of the plurality of different viewing angles Uniform alignment to achieve normalization; a feature point extraction module extracts feature points of the vehicle to be retrieved in the plurality of vehicle images of different viewing angles normalized by the adaptive vehicle normalization module; The degree calculation module matches the feature points of the vehicle to be retrieved extracted by the feature point extraction module with the feature points of the vehicle in the vehicle pictures in the database to calculate the feature points of the vehicle to be retrieved and the data. The number of the feature points of the vehicle in the vehicle pictures in the library that are paired successfully is used as the similarity score; and a voting mechanism module performs the identification and the similarity score calculated by the similarity calculation module. A voting mechanism is used to count the number of votes of the vehicle in the vehicle pictures in the database, and then select the vehicle with the higher or highest number of votes from the plurality of vehicle pictures in the database as the vehicle of the same type or a similar vehicle of the vehicle to be retrieved, Wherein, the identification and similarity calculation module further calculates the Euclidean distance between the feature point of the vehicle to be retrieved and the feature point of the vehicle in the vehicle picture in the database to find the paired feature point, Take the number of successfully paired feature points as the similarity score, and then judge Whether the similarity score is higher than a similarity threshold is used to preliminarily determine whether the vehicle to be retrieved and the vehicle in the vehicle picture in the database are the same type of vehicle or a similar vehicle. The method for retrieving similar vehicles as described in item 9 or 14 of the scope of application, further comprising: comparing the plurality of vehicle pictures from different viewing angles with the plurality of different vehicle pictures in the database in a many-to-many manner by the voting mechanism module , and then perform the voting mechanism on the similarity score calculated by the identification and similarity calculation module to count the number of votes for the vehicle in the vehicle pictures in the database. The method for retrieving similar vehicles as described in item 9 or 14 of the scope of application, further comprising selecting the vehicle with the highest number of votes from the results of the voting mechanism of the voting mechanism module by a reasonable vehicle selection module as the desired vehicle Searching for the same type of vehicle or similar vehicle of the vehicle to display the result of the same type of vehicle or similar vehicle of the vehicle to be searched on the display.

Images