A method for estimating target recognition position error using average intersection-over-union ratio

Abstract

The present invention relates to the field of intelligent transportation technology, and discloses a method for estimating target recognition position errors by using an average intersection-over-union ratio. The method comprises the following steps: S1, firstly sensing the target position through a monitoring device, sensing the positions of pedestrians, non-motor vehicles, motor vehicles, and traffic events in a road environment, and locating the target through an orientation sensor. By imaging the target and extracting the position state information of the target through a high-performance computing unit, the perception data of the system to be tested is obtained; S2, after obtaining the perception data of the system to be tested, projecting the perspective relationship onto the image, that is, establishing a mapping relationship between the image coordinate system and the GPS coordinates, and the mapping relationship is expressed in the form of a perspective transformation matrix; the position error range obtained by the invention is anisotropic, and compared with the traditional scheme that can only obtain isotropic errors through statistical methods, the theory is more rigorous and the information is more sufficient.

Description

Technical Field

[0001] The present invention relates to the field of intelligent transportation technology, and more particularly to a method for estimating target recognition position errors by using average intersection-over-union (IoU) ratio. Background Art

[0002] As an important branch of the autonomous driving field, connected autonomous driving technology has developed rapidly in recent years. Benefiting from the development of new infrastructure such as smart cities and digital roads, solutions that provide digital auxiliary data information of roads to connected autonomous driving vehicles to improve their autonomous driving level have gradually become one of the key directions of industry development.

[0003] Intelligent roadside systems are deployed along roadsides. They use sensors like radar and cameras to perceive the road environment, vehicles, and pedestrians in real time, processing and distributing this information to autonomous vehicles within their area. The GPS locations of vehicles and pedestrians are crucial information for autonomous driving, and the GPS accuracy provided by these systems plays a key role in the various algorithms subsequently executed by autonomous vehicles.

[0004] Most of the existing technologies use high-precision maps and other sensors on the vehicle to measure the positioning accuracy of the sensor to be tested.

[0005] A technology has been proposed to address the GPS-perceived positioning errors of roadside monitoring equipment for vehicles, pedestrians, and other traffic participants. This approach often relies on traditional GPS positioning equipment measurement error estimation schemes. For systems like roadside monitoring equipment, where perception errors are strongly coupled to deployment location and parameter settings, traditional schemes offer low confidence in accuracy, high measurement costs, and real-time difficulty, requiring the use of expensive, high-precision measurement vehicles. Other methods require the measured vehicles to be autonomous and supported by high-definition map information, making them difficult to implement in practice. Furthermore, for large deployments of roadside monitoring equipment, individual measurements must be performed, which is labor-intensive and makes it difficult to guarantee measurement accuracy. Summary of the Invention

[0006] In order to overcome the above-mentioned defects of the prior art, the present invention provides a method for estimating the position error of target recognition by using the average intersection-over-union ratio to solve the problems existing in the above-mentioned background technology.

[0007] The present invention provides the following technical solution: a method for estimating target recognition position error using average intersection-over-union ratio, comprising the following steps:

[0008] S1. First, the target position is sensed by the monitoring equipment. By sensing the location of pedestrians, non-motor vehicles, motor vehicles, and traffic events in the road environment, the target is located by the orientation sensor. By imaging the target and extracting the target's position status information through the high-performance computing unit, the perception data of the system under test will be collected.

[0009] S2. After obtaining the sensory data of the system under test, project the perspective relationship onto the image, that is, establish a mapping relationship between the image coordinate system and the GPS coordinates. The mapping relationship is expressed in the form of a perspective transformation matrix, which is obtained through static camera calibration.

[0010] S3. Recording a video of an actual road scene with an image sensor, intercepting the video using certain rules to obtain multiple single-frame image data, annotating the target positions in the video, and forming the position truth data of the image coordinate system. The interception rule is generally fixed-interval interception, and the annotated position truth data set is obtained;

[0011] S4. Randomly extract a training set from the true value dataset of the labeled position, calculate the average IOU of all objects in all images, obtain multiple training set data IOUs, and calculate the average IOU of them;

[0012] S5. Mechanically extract the test set from the ground truth dataset of the labeled position, calculate the IOU of the test set, and compare its IOU value with the average IOU data of the training set to obtain the estimated IOU of the test set;

[0013] S6. Perform error statistics on the estimated test set IOU and the true test set IOU to determine whether the error statistics exceed the threshold. If so, record the final average IOU. If it exceeds the threshold, repeat the random sampling of training sets to calculate the average IOU to increase the training set.

[0014] S7, the final average IOU is calculated by the formula:

[0015]

[0016] Where S is the detection target area, τ is the IOU, d is the pixel error, k is the empirical coefficient, and the initial value is selected as 0.8. k can be changed by adding randomly sampled training sets. The estimated IOU of the actual runtime data can be obtained by calculation;

[0017] S8. By projecting the pixel error range onto the GPS coordinate system, an anisotropic position error range is obtained, that is, the position error ranges in different directions are different, thereby obtaining a target recognition position error result.

[0018] Furthermore, the position value of S1 is GPS information, which can be used to perform error correction on the target position information perceived by the roadside system of the intelligent transportation system. The position information of the target perceived by the roadside system is in GPS format, and the error format is a dispersion range.

[0019] Furthermore, the S2 is specifically a matching point group method, the number of matching point groups is greater than 4, and the geometric relationship cannot form a straight line. The GPS position and size of the target are projected into the image coordinate system through mapping to obtain the target position measurement value.

[0020] Furthermore, the position measurement values ​​obtained in S4-S5 are grouped into data sets according to time, and a certain amount of training sets and test sets are randomly selected, generally 80% of the training set and 20% of the test set.

[0021] Furthermore, the number of S3 image data is generally greater than 200, and the annotation method is to draw the minimum circumscribed rectangle of the target outline in the image. The image data should include the GPS position and size of the target.

[0022] Furthermore, the error estimation method has at least one camera as an image sensor to provide an image coordinate system reference, and the position of the target perceived by the roadside system needs to be projected onto the image coordinate system reference.

[0023] Furthermore, the annotated position true set adopts a static annotation method, manually annotating the true value of pixels in the image, and the true value is represented in the form of the minimum circumscribed rectangle of the target outline.

[0024] The technical effects and advantages of the present invention are as follows:

[0025] 1. The present invention uses manual annotation to perform true value annotation, which is lower in cost compared with the traditional solution of obtaining the true value of the position by using a true value car and high-precision map method.

[0026] 2. The present invention ensures GPS measurement accuracy through image static calibration, avoiding the separate calibration of multiple sensors such as radar. The implementation method is simple and the efficiency is greatly improved.

[0027] 3. The present invention achieves decoupling of target recognition error estimation and deployment mode, that is, the error estimation algorithm is weakly correlated with sensor deployment. Compared with traditional solutions, it has great advantages in large-scale deployment.

[0028] 4. The position error range obtained by the present invention is anisotropic. Compared with the traditional solution that can only obtain isotropic errors through statistical methods, the theory is more rigorous and the information is more sufficient. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 Schematic diagram of the error calculation process of the present invention.

[0030] Figure 2 Schematic diagram of IOU calculation of the present invention. DETAILED DESCRIPTION

[0031] The technical solutions of the present invention will be described clearly and completely below in conjunction with the drawings in the present invention. In addition, the forms of the various structures described in the following embodiments are merely examples. The method for estimating the position error of target recognition using the average intersection-over-union ratio involved in the present invention is not limited to the various structures described in the following embodiments. All other implementations obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0032] Reference Figure 1 and Figure 2 The present invention provides a method for estimating target recognition position error using average intersection-over-union ratio, comprising the following steps:

[0033] S1. First, the target position is sensed by the monitoring equipment. By sensing the location of pedestrians, non-motor vehicles, motor vehicles, and traffic events in the road environment, the target is located by the orientation sensor. By imaging the target and extracting the target's position status information through the high-performance computing unit, the perception data of the system under test will be collected.

[0034] S2. After obtaining the sensory data of the system under test, project the perspective relationship onto the image, that is, establish a mapping relationship between the image coordinate system and the GPS coordinates. The mapping relationship is expressed in the form of a perspective transformation matrix, which is obtained through static camera calibration.

[0035] S3. Recording a video of an actual road scene with an image sensor, intercepting the video using certain rules to obtain multiple single-frame image data, annotating the target positions in the video, and forming the position truth data of the image coordinate system. The interception rule is generally fixed-interval interception, and the annotated position truth data set is obtained;

[0036] S4. Randomly extract a training set from the true value dataset of the labeled position, calculate the average IOU of all objects in all images, obtain multiple training set data IOUs, and calculate the average IOU of them;

[0037] S5. Mechanically extract the test set from the ground truth dataset of the labeled position, calculate the IOU of the test set, and compare its IOU value with the average IOU data of the training set to obtain the estimated IOU of the test set;

[0038] S6. Perform error statistics on the estimated test set IOU and the true test set IOU to determine whether the error statistics exceed the threshold. If so, record the final average IOU. If it exceeds the threshold, repeat the random sampling of training sets to calculate the average IOU to increase the training set.

[0039] S7, the final average IOU is calculated by the formula:

[0040]

[0041] Where S is the detection target area, τ is the IOU, d is the pixel error, k is the empirical coefficient, and the initial value is selected as 0.8. k can be changed by adding randomly sampled training sets. The estimated IOU of the actual runtime data can be obtained by calculation;

[0042] S8. By projecting the pixel error range onto the GPS coordinate system, an anisotropic position error range is obtained, that is, the position error ranges in different directions are different, thereby obtaining a target recognition position error result.

[0043] Among them, the position value of S1 is GPS (longitude, latitude) information, which can be used to correct the error of the target position information perceived by the intelligent transportation system (ITS) roadside system. The position information of the target perceived by the roadside system is in GPS format, and the error format is a dispersion range.

[0044] Among them, S2 is specifically a (pixel-GPS) matching point group method. The number of matching point groups is greater than 4, and the geometric relationship cannot form a straight line. The GPS position and size of the target are projected into the image coordinate system through mapping to obtain the target position measurement value.

[0045] The position measurement values ​​obtained by S4-S5 are grouped into time pairs to form a data set, and a certain amount of training set and test set are randomly selected, generally 80% of the training set and 20% of the test set.

[0046] Among them, the number of S3 image data is generally greater than 200. The annotation method is to draw the minimum circumscribed rectangle of the target outline in the image. The image data should include the GPS position and size of the target.

[0047] Among them, the error estimation method has at least one camera as an image sensor to provide an image coordinate system reference, and the position of the target perceived by the roadside system needs to be projected to the image coordinate system reference.

[0048] Among them, the true set of annotated positions adopts a static annotation method, manually annotating the true value of pixels in the image, and the true value is expressed in the form of the minimum circumscribed rectangle of the target contour.

[0049] Working principle of the present invention:

[0050] A roadside monitoring device can sense the target position and extract the target's position status information through a high-performance computing unit. This function is implemented by a perception program deployed in the roadside system. The system includes an image sensor, and the image coordinate system of the image formed by the image sensor can establish a mapping relationship with the GPS coordinate system. The mapping relationship is expressed in the form of a perspective transformation matrix. The perspective transformation matrix is ​​obtained through static calibration of the camera. The image sensor is used to record the actual road scene video, and multiple single-frame image data are obtained by intercepting the video. The target position is manually annotated to form the position true value data of the image coordinate system. The interception rule is generally fixed interval interception. The number of image data is generally greater than 200. The annotation method is to draw the minimum circumscribed rectangle of the target outline in the image.

[0051] At the same time, by synchronously storing the perception results of the roadside equipment at the corresponding time of the video, the perception result data corresponding to the image data is obtained. This data should include the GPS position and size of the target. The obtained target GPS position and size are projected into the image coordinate system to obtain the target position measurement value. Then, the obtained position measurement value is grouped by time to form a data set. A certain amount of training set and test set are randomly selected, generally 80% for training set and 20% for test set. The intersection over union (IOU) of each target in the obtained training set is calculated, and the IoU estimation is performed on the obtained test set. The error statistics are performed between the obtained result and the true value IoU of the test set.

[0052] By determining that the error statistic is less than the threshold, the final average IoU is recorded. Otherwise, the training set is increased, and the mapping relationship between the target image position box and the target pixel position is established according to certain rules. The center point of the bottom edge of the minimum circumscribed rectangle of the contour in the image can be selected. The average IoU obtained from the actual perception data is used to calculate the estimated IoU, and the pixel error range is calculated using the estimated IoU. See the formula for calculating the anisotropic position error range. Anisotropy means that the position error range in different directions is different. By projecting the pixel error range into the GPS coordinate system, the anisotropic position error range is obtained. Anisotropy means that the position error range in different directions is different, thereby obtaining a specific error measurement result. The above method is used to obtain a more accurate target recognition position error estimation method.

[0053] Finally, a few points should be explained: First, in the description of this application, it should be noted that, unless otherwise specified or limited, the terms "mounted," "connected," and "connected" should be understood in a broad sense, and may refer to mechanical or electrical connections, internal communication between two components, or direct connection. "Up," "down," "left," and "right" are only used to indicate relative positional relationships. When the absolute positions of the objects being described change, the relative positional relationships may also change.

[0054] Secondly: The drawings of the embodiments disclosed in the present invention only involve structures related to the embodiments disclosed in the present invention. Other structures may refer to conventional designs. The same embodiment and different embodiments of the present invention may be combined with each other without conflict.

[0055] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for estimating target recognition position error using average intersection-over-union ratio, characterized in that: The following steps are involved: S1. First, the target position is sensed by the monitoring equipment. By sensing the location of pedestrians, non-motor vehicles, motor vehicles, and traffic events in the road environment, the target is located by the orientation sensor. By imaging the target and extracting the target's position status information through the high-performance computing unit, the perception data of the system under test will be collected. S2. After obtaining the sensory data of the system under test, project the perspective relationship onto the image, that is, establish a mapping relationship between the image coordinate system and the GPS coordinates. The mapping relationship is expressed in the form of a perspective transformation matrix, which is obtained through static camera calibration. S3. Recording a video of an actual road scene using an image sensor, intercepting the video using a certain rule to obtain multiple single-frame image data, annotating the target positions in the video, and forming the position truth data of the image coordinate system. The interception rule is fixed-interval interception, and a labeled position truth data set is obtained; S4. Randomly extract a training set from the true value dataset of the labeled position, calculate the average IOU of all objects in all images, obtain multiple training set data IOUs, and calculate the average IOU of them; S5. Mechanically extract the test set from the ground truth dataset, calculate the IOU of the test set, and then compare its IOU value with the average IOU data of the training set to obtain the estimated IOU of the test set; S6. Perform error statistics on the estimated test set IOU and the true test set IOU to determine whether the error statistics exceed the threshold. If so, record the final average IOU. If it exceeds the threshold, repeat the random sampling of training sets to calculate the average IOU to increase the training set. S7. Calculate the final average IOU using the formula: Where S is the detection target area, τ is the final average IOU, d is the pixel error, k is the empirical coefficient, and the initial value is selected as 0.

8. k can be changed by adding randomly sampled training sets, and the estimated IOU of the actual runtime data is calculated; S8. By projecting the pixel error range onto the GPS coordinate system, an anisotropic position error range is obtained, that is, the position error ranges in different directions are different, thereby obtaining a target recognition position error result.

2. The method for estimating target recognition position error using average intersection-over-union (IoU) according to claim 1, wherein: The position value of S1 is GPS information, which can be used to correct the target position information perceived by the roadside system of the intelligent transportation system. The position information of the target perceived by the roadside system is in GPS format, and the error format is a dispersion range.

3. The method for estimating target recognition position error using average intersection-over-union (IoU) according to claim 1, wherein: The S2 is specifically a pixel-GPS matching point group method, the number of matching point groups is greater than 4, and the geometric relationship cannot form a straight line.

4. The method for estimating target recognition position error using average intersection-over-union (IoU) according to claim 1, wherein: The position measurement values ​​obtained in S4-S5 are grouped into pairs according to time to form a data set, and a certain amount of training set and test set are randomly selected, with 80% of the training set and 20% of the test set.

5. The method for estimating target recognition position error using average intersection-over-union (IoU) according to claim 1, wherein: The number of S3 image data is greater than 200, and the annotation method is to draw the minimum circumscribed rectangle of the target outline in the image.

6. The method for estimating target recognition position error using average intersection-over-union (IoU) according to claim 1, wherein: The error estimation method has at least one camera as an image sensor to provide an image coordinate system reference, and the position of the target perceived by the roadside system needs to be projected onto the image coordinate system reference.

7. The method for estimating target recognition position error using average intersection-over-union (IoU) according to claim 1, wherein: The annotated position true set adopts a static annotation method, and manually annotates the true value of pixels in the image, and the true value is represented in the form of the minimum circumscribed rectangle of the target outline.

Images