TWI646306B - Method for analyzing error and existence probability of multi-sensor fusion of obstacle detection - Google Patents

Abstract

The invention provides an error and detection probability analysis method applied to multi-sensor fusion, wherein an obstacle sensing step generates obstacle observation information, and an obstacle prediction step generates obstacle prediction information. The error model offline establishment step establishes a prior average error distribution function according to the sensor. The detection confidence establishment step establishes the pre-detection probability information according to the sensor. The tracking fusion step is performed to track the fusion method to fuse the information to generate a fusion error variation. The fusion error variation cumulative correction step corrects the cumulative value of the fusion error variation based on the prior detection probability information. In this way, through the pre-processed sensor error analysis combined with the detection confidence model to fuse the detection information between the plurality of sensors, the presence of obstacles with higher reliability can be obtained.

Description

Applied to multi-sensor fusion error and detection Probability analysis method

The invention relates to an error and detection probability analysis method, in particular to an error and detection probability analysis method applied to multi-sensor fusion.

At present, the functions of the vehicle computer are becoming more and more perfect. In order to improve the safety of driving and the future of autonomous driving, the reliability of the detection and classification of the obstacles in front of the vehicle is particularly important. For various objects such as cars, pedestrians, bicycles, and utility poles, the classification items are determined according to the system settings. In this way, the system can determine whether to prompt the brakes, automatically brake, or perform other actions according to the classification of the obstacles.

There are many types of sensors commonly used in vehicles to detect obstacles. Visual imaging systems and radar systems are commonly used. Visual imaging systems are used on vehicles to enhance object detection and other visual or positioning systems. Application, such a system uses a camera to capture an image and identify an object (ie, an obstacle) from the image, which may be within the road He is a vehicle, a pedestrian or even an object; as far as the radar system is used to detect objects in the road, the radar system uses radio waves to determine the distance, direction or speed of the object, and the radar transmitter emits radio wave pulses in its path. Any object within it will be bounced by a radio wave pulse, and in addition, a pulse reflected from the object will transmit a small fraction of the radio wave energy to the receiver, which is typically in the same position as the transmitter.

Although the above sensors can detect obstacles, their reliability is often insufficient, and the error of the detection result is likely to be too large, and the error of tracking the position of the obstacle often occurs frequently. It can be seen that there is a lack of error and detection probability analysis method for multi-sensor fusion in the market, so the relevant industry is seeking solutions.

Therefore, the object of the present invention is to provide an error and detection probability analysis method for multi-sensor fusion, which combines a plurality of sensors through a pre-processed sensor error analysis combined with a detection confidence model. The pre-detection of the probability information and the correction of the cumulative value of the fusion error can be used to obtain a more reliable obstacle existence judgment; the global satellite positioning real-time dynamic measurement device pre-establishes various environments, obstacles and vehicle condition conditions. The pre-existing average error distribution function, and dynamically correcting the tracking result by using the pre-average error distribution function, can generate the information of the merged obstacle with less error and higher credibility, and further, the step of establishing confidence by detecting confidence Pre-detecting the probability information to correct the cumulative value of the fusion error, as a basis for judging the existence of obstacles, can greatly increase the judgment The reliability of the broken result achieves the effect of real-time operation, and solves the problem that the sensing error of the prior art is too large and the reliability is too low.

According to an embodiment of the present invention, an error and detection probability analysis method for multi-sensor fusion is provided for determining an obstacle in a traveling direction of a vehicle. The error and detection probability analysis method applied to the multi-sensor fusion includes an obstacle sensing step, an obstacle prediction step, an error model offline establishing step, a detection confidence establishing step, a tracking fusion step, and a fusion error variation amount accumulation. a correction step, wherein the obstacle sensing step provides a plurality of sensors to sense the obstacles to generate a plurality of obstacle observation information respectively; the obstacle prediction step provides the processor to generate the plurality of obstacles according to the plurality of obstacle observation information respectively Obstacle prediction information. Furthermore, the error model offline establishing step uses the processor to respectively establish a plurality of pre-existing average error distribution functions according to the plurality of sensors; the detecting confidence establishing step is to use the processor to establish a plurality of pre-investigations respectively according to the plurality of sensors. Measuring the rate information; the tracking and fusion step is to use the processor to combine the obstacle observation method, the obstacle prediction information and the prior average error distribution function to generate a plurality of fusion error variations and a plurality of post-fusion obstacle information; The fusion error variation cumulative correction step is to use the processor to correct the cumulative value of the fusion error variation according to a plurality of prior detection probability information to determine whether an obstacle exists.

Therefore, the method of the present invention combines the pre-processed sensor error analysis with the detection confidence model to fuse the pre-detection probability information between the plurality of sensors and correct the cumulative value of the fusion error variation. Get a judgment of a more credible obstacle. In addition, through the detection letter The pre-detection probability information of the heart-building step is used to correct the cumulative value of the fusion error as a basis for judging the presence or absence of the obstacle, which can greatly increase the reliability of the judgment result, achieve the real-time operation effect, and solve the conventional technology. The problem of excessive sensing error and low reliability.

Other implementations of the foregoing embodiments include: in the foregoing tracking fusion step, the tracking fusion method may be a Kalman Filter algorithm, and each of the merged obstacle information includes a merged obstacle position and a fusion Obstacle speed and the type of obstacle after fusion. Furthermore, the aforementioned one sensor may be a radar sensor (RADAR), and the other sensor is a camera. In addition, the obstruction observation information of the obstacle sensing step may include an observation position, and the error model offline establishing step may previously set an instant dynamic positioning module on the obstacle and drive the real dynamic positioning module to generate the complex number. The instantaneous dynamic positioning position, and then the processor receives and calculates the relative error between the instantaneous dynamic positioning position and the observation position to generate a prior average error distribution function. The sensor has a field of view (FOV), and the instantaneous dynamic positioning position and the observation position are both located in the field of view and correspond to the observation speed of the obstacle. In addition, in the foregoing fusion error change amount cumulative correction step, the processor may store a preset cumulative threshold value, and the processor compares the preset cumulative threshold value and the accumulated numerical value to determine whether an obstacle exists. When the accumulated value is less than or equal to the preset cumulative threshold, the obstacle is considered to exist; when the accumulated value is greater than the preset cumulative threshold, the obstacle is considered to be absent.

According to another embodiment of the present invention, an error and detection probability analysis method for multi-sensor fusion is provided, which is used to determine a vehicle. An obstacle in the direction of travel. The error and detection probability analysis method applied to the multi-sensor fusion includes an obstacle sensing step, an obstacle prediction step, an error model offline establishing step, and a tracking fusion step, wherein the obstacle sensing step provides a plurality of sensing steps The device respectively generates a plurality of obstacle observation information by sensing the obstacle, and the obstacle observation information includes an observation position and an observation speed; the obstacle prediction step provides a processor to generate a plurality of pieces according to the plurality of obstacle observation information respectively. Obstacle prediction information; the error model offline establishing step is to use the processor to establish a plurality of pre-existing average error distribution functions according to the plurality of sensors, and the error model offline establishing step is to set an instant dynamic positioning module on the obstacle beforehand and Driving the real-time dynamic positioning module to generate a plurality of real-time dynamic positioning positions (RTK-GPS), and then the processor receives and calculates the relative error between the instantaneous dynamic positioning position and the observed position to generate a prior average error distribution function; the tracking fusion step is processed Fusion tracking obstacles with a tracking fusion method Obstacle to predict in advance the average error information and distribution function to generate the plurality of information integration obstacles. In addition, the sensor has a field of view (FOV), and the instantaneous dynamic positioning position and the observation position are both located in the field of view and correspond to the observation speed of the obstacle.

Thereby, the method of the invention pre-establishes the prior average error distribution function under various environments, obstacles and vehicle conditions by the global satellite positioning real-time dynamic measuring device, and dynamically corrects the tracking result by using the function database, which can generate errors. Smaller and more reliable information on post-fusion obstacles.

Other implementations of the foregoing embodiments include: in the foregoing tracking fusion step, the tracking fusion method may be a Kalman Filter algorithm, and each of the merged obstacle information includes the position of the obstacle after the fusion and the speed of the obstacle after the fusion And the type of obstacles after fusion. One of the aforementioned sensors may be a radar sensor (RADAR) and the other sensor is a camera. In addition, the foregoing error and detection probability analysis method applied to the multi-sensor fusion may include a detection confidence establishing step, wherein the detecting confidence establishing step uses the processor to respectively establish a plurality of pre-detection according to the sensor. Probability information. In the tracking fusion step, the processor generates a plurality of fusion error variations by integrating the obstacle observation information, the obstacle prediction information, and the prior average error distribution function by a tracking fusion method. Furthermore, the error and detection probability analysis method applied to the multi-sensor fusion may include a fusion error variation cumulative correction step, and the fusion error variation cumulative correction step is performed by the processor according to the pre-detection probability information correction fusion. The cumulative value of the amount of error change to determine if an obstacle exists. In addition, in the foregoing fusion error change amount cumulative correction step, the processor stores a preset cumulative threshold value, and the processor compares the preset cumulative threshold value and the accumulated numerical value to determine whether an obstacle exists. When the accumulated value is less than or equal to the preset cumulative threshold, the obstacle is considered to exist; when the accumulated value is greater than the preset cumulative threshold, the obstacle is considered to be absent.

100, 100a‧‧‧Error and detection probability analysis system for multi-sensor fusion

110‧‧‧ Vehicles

120‧‧‧ obstacles

130‧‧‧Grid

200‧‧‧ sensor

300, 300a‧‧‧ processor

S12, S21‧‧‧ obstacle sensing steps

S14, S22‧‧‧ obstacle prediction steps

S16, S23‧‧‧ error model offline establishment steps

S18, S24‧‧‧ tracking fusion steps

310‧‧‧ obstacle sensing module

320‧‧‧ obstacle prediction module

330‧‧‧ Error model offline module

332‧‧‧Instant Dynamic Positioning Module

340, 340a‧‧‧ Tracking Fusion Module

342‧‧‧Confusion error variation

350‧‧‧Detection confidence building module

352‧‧‧Pre-detection probability information

360‧‧‧Fused Error Variation Accumulation Correction Module

362‧‧‧ Status information

370‧‧‧ Collision time calculation module

400, 400a‧‧‧Application of multi-sensor fusion error and detection probability analysis method

S25‧‧‧Steps to detect confidence

S26‧‧‧Combined error variation Correction step

S27‧‧‧ Collision time calculation steps

(x, y, v) ‧ ‧ obstacle observation information

(x, y) ‧ ‧ observation position

V‧‧‧ observation speed

(x', y', v') ‧ ‧ obstacle prediction information

(x', y') ‧ ‧ predicted position

V'‧‧‧ forecast speed

f (x, y, v) ‧ ‧ ante average error distribution function

(x", y", v") ‧ ‧ post-fusion obstacle information

(x", y") ‧ ‧ position of the obstacle after fusion

v"‧‧‧ obstacle speed after fusion

T1~T30‧‧‧Time

FOV‧‧ Vision

FIG. 1 is a block diagram showing an error and detection probability analysis system applied to multi-sensor fusion according to an embodiment of the present invention.

FIG. 2 is a schematic flow chart showing an error of the multi-sensor fusion and a detection probability analysis method according to an embodiment of the present invention.

FIG. 3 is a schematic diagram showing the measurement environment of an obstacle-equipped instant dynamic positioning module according to an embodiment of the present invention.

Fig. 4A is a diagram showing the dynamic tracking result of the sensor of the radar model in the offline setting step of the error model in Fig. 2.

Fig. 4B is a diagram showing the error distribution of the obstacle observation speed of 20kph in the offline establishment step of the error model in Fig. 2.

Fig. 4C is a diagram showing the error distribution of the obstacle observation speed of 60 kph in the offline establishment step of the error model in Fig. 2.

The 5A~5C diagram shows that the error model offline establishment step in Fig. 2 dynamically corrects the tracking result with the detection error between multiple sensors in different time periods.

FIG. 6 is a block diagram showing an error and detection probability analysis system applied to multi-sensor fusion according to another embodiment of the present invention.

FIG. 7 is a flow chart showing a method for analyzing errors and detecting probability of multi-sensor fusion according to another embodiment of the present invention.

Fig. 8A is a graph showing the amount of fusion error variation of the fusion error variation cumulative correction step in Fig. 7.

Fig. 8B is a graph showing the cumulative value of the amount of change in the fusion error of Fig. 8A.

Hereinafter, a plurality of embodiments of the present invention will be described with reference to the drawings. For the sake of clarity, many practical details will be explained in the following description. However, it should be understood that these practical details are not intended to limit the invention. That is, in some embodiments of the invention, these practical details are not necessary. In addition, some of the conventional structures and elements are illustrated in the drawings in a simplified schematic manner, and the repeated elements may be represented by the same reference numerals.

Please refer to FIG. 1 and FIG. 3 together. FIG. 1 is a block diagram showing an error and detection probability analysis system 100 applied to multi-sensor fusion according to an embodiment of the present invention. The error and detection probability analysis system 100 includes a plurality of sensors 200 and a processor 300. FIG. 3 is a schematic diagram of a measurement environment according to an embodiment of the present invention, wherein the obstacle 120 is equipped with an instant dynamic positioning module 332, and a vehicle 110 including a plurality of sensors 200 and a processor 300. The error and detection probability analysis system 100 for multi-sensor fusion is used to determine the obstacle 120 in the direction of travel of the vehicle 110.

A plurality of sensors 200 are disposed on the vehicle 110, and the plurality of sensors 200 can be in different configurations. The number of the plurality of sensors 200 in this embodiment is two, and one of the sensors 200 may be a radar sensor (RADAR), and the other sensor 200 may be a camera. A radar sensor (RADAR) is used to sense the position and speed of the obstacle 120, and the camera is used to sense the position of the obstacle 120 and identify the type of the obstacle 120, but is not limited to the above number and the type of the sensor 200.

The processor 300 is disposed on the vehicle 110 and connected to the plurality of sensors 200, and the processor 300 can be an electronic control unit (ECU) for the vehicle, a microprocessor or other electronic computing processor, and the like. The device 300 includes an obstacle sensing module 310, an obstacle prediction module 320, an error model offline establishing module 330, and a tracking fusion module 340. The obstacle sensing module 310 is sensed by using the plurality of sensors 200. The obstacle 120 signal generates a plurality of obstacle observation information (x, y, v), and each obstacle observation information (x, y, v) includes an observation position (x, y) and an observation speed v, wherein the observation position ( x, y) represents the position obtained by the obstacle 120 being sensed, and the observation speed v represents the moving speed obtained by the obstacle 120 being sensed; and the obstacle prediction module 320 generates the plural according to the plurality of sensors 200 respectively. Obstacle prediction information (x', y', v'), wherein the predicted position (x', y') represents the predicted position of the obstacle 120, and the predicted speed v' represents the predicted obstacle 120. The speed of movement. In addition, the error model offline establishing module 330 establishes a plurality of prior average error distribution functions f (x, y, v) according to the plurality of sensors 200 in different test scenarios, and the prior average error distribution function f (x, y) , v) is the average function of the error between the obstacle observation information (x, y, v) and the actual information of the actual obstacle 120, which is used to dynamically correct obstacle observation information (x, y, v) and obstacle prediction The error between the information (x', y', v'), for example: Figure 4B and Figure 4C respectively show the error distribution of two different test scenarios, wherein the obstacle observation speed of Figure 4B is 20kph, The obstacle observation speed in Fig. 4B is 60kph. In the two test scenarios, two different pre-average error distribution functions f (x, y, v) can be obtained. The tracking fusion module 340 integrates obstacle observation information (x, y, v), obstacle prediction information (x', y', v') and the prior average error distribution function f (x, y) through a tracking fusion method. , v) and generate a plurality of post-fusion obstacle information (x", y", v"), wherein the tracking fusion method is Kalman Filter (Kalman Filter), the position of the obstacle after fusion (x", y") Representing the position of the obstacle 120 after fusion, the speed of the obstacle after v" represents the moving speed of the obstacle 120 after fusion. In other words, the tracking fusion module 340 is connected to the obstacle sensing module 310, the obstacle prediction module 320, and the error model offline establishing module 330. Since the present embodiment is two sensors 200, namely a radar sensor and a camera, each sensor 200 has corresponding obstacle observation information (x, y, v) and obstacle prediction information (x' , y', v') and the pre-existing average error distribution function f (x, y, v), thereby generating post-fusion obstacle information (x", y", v") with less error and higher reliability.

Please refer to FIG. 1 to FIG. 4C. FIG. 2 is a schematic flow chart of the error and detection probability analysis method 400 applied to the multi-sensor fusion according to an embodiment of the present invention. FIG. 4A is a diagram showing the dynamic tracking result of the sensor 200 in the error model offline establishing step S16 in FIG. 2 as a radar sensor. 4B is a diagram showing the error distribution of the obstacle 120 in the error map offline creation step S16 in FIG. 2 when the observation speed v is 20 kph. Fig. 4C is a diagram showing the error distribution of the obstacle 120 in the error model offline creation step S16 in Fig. 2 at an observation speed v of 60 kph. As shown, the multi-sensor fusion error and detection probability analysis method 400 is used to determine the obstacle 120 in the direction of travel of the vehicle 110; this is applied to multi-sensor fusion error and detection probability. Analytical method 400 includes an obstacle sensing step S12, an obstacle prediction step S14, an error model offline establishment step S16, and a tracking fusion step S18.

The obstacle sensing step S12 provides a plurality of sensors 200 to sense the obstacles 120 to generate a plurality of obstacle observation information (x, y, v), that is, the obstacle sensing step S12 utilizes sensing. After the obstacle 200 is sensed by the device 200, the system generates a plurality of obstacle observation information (x, y, v) through the obstacle sensing module 310.

The obstacle prediction step S14 is to provide the processor 300 to generate a plurality of obstacle prediction information (x', y', v') according to the plurality of sensors 200, respectively. In detail, after the obstacle prediction step S14 senses the obstacle 120 by the sensor 200, the system generates a plurality of obstacle prediction information (x', y according to the plurality of sensors 200 through the obstacle prediction module 320. ',v').

The error model offline establishing step S16 is to use the processor 300 to respectively establish a plurality of prior average error distribution functions f (x, y, v) according to the sensor 200. In detail, the error model offline establishing step S16 is to pre-set an instant dynamic positioning module 332 on the obstacle 120 and drive to generate a plurality of instantaneous dynamic positioning positions (such as the symbol "○" in FIG. 4A and the 4B, 4C. The horizontal axis of the graph is shown). Then, the error model offline establishing module 330 of the processor 300 receives and calculates the relative error between the instantaneous dynamic positioning position and the observation position (x, y) of the obstacle observation information (x, y, v) to generate a prior average error distribution function. f (x, y, v), where the observation position (x, y) is as indicated by the symbol "X" in Fig. 4A. The black points in Figures 4B and 4C are the difference between the observation position (x, y) of the radar sensor and the instantaneous dynamic positioning position of the global positioning system RTK-GPS in different test situations. The difference is used to establish an average error distribution function f (x, y, v) beforehand, and the prior average error distribution function f (x, y, v) varies with the state of the vehicle condition of the preceding vehicle obstacle 120. The ex ante average error distribution function f (x, y, v) is shown by the trend line in Figures 4B and 4C. In addition, the sensor 200 has a field of view (FOV), and both the instantaneous dynamic positioning position and the observation position (x, y) are located in the field of view and correspond to the observation speed v of the obstacle 120. The processor 300 can form a plurality of grids 130 within the field of view to determine the location of the obstacles 120. The real-time dynamic positioning module 332 of the present embodiment is a global satellite positioning real-time dynamic measuring device (RTK-GPS), so the real-time dynamic positioning module 332 establishes the module 330 offline through the global positioning system (GPS) signal connection error model, and The instantaneous dynamic positioning position measured by the real-time dynamic positioning module 332 is regarded as the correct information of the actual obstacle 120. The radar sensor 200 senses that the plurality of observation positions (x, y) generated by the obstacle 120 are compared with the instantaneous dynamic positioning position, and the difference between the two can establish an average error distribution function beforehand. For the database of f (x, y, v), the pre-existing mean error distribution function f (x, y, v) can be used as a reference for correction. In other words, when the radar sensor 200 is used in the future, the observation position (x, y) can be appropriately corrected by a pre-established pre-averaged error distribution function f (x, y, v) to obtain a more accurate fusion. The rear obstacle position (x", y"), the post-fusion obstacle position (x", y") is indicated by the symbol "△" in Fig. 4A. The details of this modification will be explained below, that is, the tracking fusion step S18.

The tracking and fusion step S18 uses the processor 300 to fuse obstacle observation information (x, y, v), obstacle prediction information (x', y', v') and the prior average error distribution function f (x) by a tracking fusion method. , y, v) and generate a plurality of post-fusion obstacle information (x", y", v"). In detail, the tracking fusion method is a Kalman filter algorithm, and the Kalman filter algorithm is used to track the fusion module. 340 is performed, and each of the post-fusion obstacle information (x", y", v") includes the post-fusion obstacle position (x", y") and the post-fusion obstacle speed v". In other embodiments, the fusion The posterior obstacle information (x", y", v") may also include a post-fusion obstacle species, which may be a pedestrian, a car, or other type of obstacle 120. Therefore, the error and detection probability analysis method 400 applied to the multi-sensor fusion of the present invention utilizes the error analysis of the pre-processed sensor 200, and pre-establishes various environments and obstacles through the global satellite positioning instantaneous dynamic measurement device. 120 and the pre-existing average error distribution function f (x, y, v) under the condition of the vehicle 110, and dynamically correcting the tracking result by the prior average error distribution function f (x, y, v), which can generate less error and be credible A higher degree of post-fusion obstacle information (x", y", v").

Please refer to the figures 1, 2, 3 and 5A~5C together. The 5A~5C diagram shows the error of the error model in the second figure. Step S16 is used to detect errors between multiple sensors 200 in different time. Dynamically correct tracking results. As shown, the symbol "X" represents the observed position (x, y) sensed by the camera, and the symbol "△" represents the tracking position (observed, predicted, and corrected by the error model) of the camera's observed position (x, y). The symbol "*" represents the observed position (x, y) sensed by the radar sensor, and the symbol "○" represents the tracking position (x, y) of the radar sensor (observed, predicted, and error model) Correction), as for the symbol "□", the dynamic correction tracking result after the camera and the radar sensor are synchronized. There are 30 different times in Figures 5A~5C The tracking results of T1~T30, these times T1~T30 are equally spaced from each other and occur sequentially. In general, the use of the radar sensor to sense the distance and position of the obstacle 120 is more accurate than the sensing information of the camera, and the present invention dynamically corrects the error through the detection error between the plurality of different sensors 200. And get more reliable tracking results.

Please refer to FIG. 1 and FIG. 6 together. FIG. 6 is a block diagram showing an error and detection probability analysis system 100a applied to multi-sensor fusion according to another embodiment of the present invention. As shown, the multi-sensor fusion error and detection probability analysis system 100a is used to determine the obstacle 120 in the direction of travel of the vehicle 110 and includes a plurality of sensors 200 and a processor 300a. The processor 300a includes an obstacle sensing module 310, an obstacle prediction module 320, an error model offline establishing module 330, a tracking fusion module 340a, a detection confidence setting module 350, and a fusion error variation cumulative correction mode. The group 360 and the collision time calculation module 370; the sensor 200, the obstacle sensing module 310, the obstacle prediction module 320, and the error model offline creation module 330 are the same as the corresponding blocks in FIG. Narration. In particular, the processor 300a further includes a tracking fusion module 340a, a detection confidence establishment module 350, a fusion error variation cumulative correction module 360, and a collision time calculation module 370.

Tracking fusion module 340a to track fusion method fusion obstacle observation information (x, y, v), obstacle prediction information (x', y', v') and prior average error distribution function f (x, y, v) The plurality of fusion error variations 342 and the plurality of merged obstacle information (x", y", v") are generated. The tracking fusion method of the embodiment is a Kalman filter algorithm and has two types of sensors 200. The two sensors 200 are a radar sensor and a camera, respectively.

The detection confidence establishing module 350 signal connection sensor 200 and the error model offline establishing module 330, and the detection confidence establishing module 350 respectively establish a plurality of pre-detection probability information 352 according to different sensors 200, The pre-detection probability information 352 represents the probability that the signal detected by the sensor 200 is true (true) or false (False), and can also be regarded as the confidence of detection.

The fusion error variation cumulative correction module 360 signal connection tracking fusion module 340a and the detection confidence establishment module 350, and the fusion error variation cumulative correction module 360 establishes the pre-detection probability of the module 350 according to the detection confidence level. The information 352 is used to correct the accumulated value of the fusion error variation 342 of the tracking fusion module 340a to determine whether the obstacle 120 exists. In detail, the fusion error variation cumulative correction module 360 stores a preset cumulative threshold value, and the fusion error variation cumulative correction module 360 compares the preset cumulative threshold value with the cumulative value to determine whether the obstacle 120 exists. When the accumulated value is less than or equal to the preset cumulative threshold, the obstacle 120 is considered to exist; conversely, when the accumulated value is greater than the preset cumulative threshold, the obstacle 120 is considered to be absent. Furthermore, the fusion error variation cumulative correction module 360 corrects the cumulative value of the fusion error variation 342 according to the prior detection probability information 352, and outputs the obstacle 120 corresponding to the merged obstacle information (x", y", v" Status information 362.

The collision time calculation module 370 receives the merged obstacle information (x", y", v") and the presence status information 362 to calculate the collision time of the vehicle 110 and the obstacle 120, and the collision time can be used as the determination parameter of the automatic driving. Thereby, the system of the present invention combines the detection confidence model with the pre-processed sensor 200 error analysis to fuse the pre-detection probability information 352 between the plurality of sensors 200 and correct the fusion error variation 342. The accumulated value can be judged by a more reliable obstacle 120. In addition, the pre-average error distribution function f (x) of various environments, obstacles 120 and vehicle 110 conditions is pre-established by the global satellite positioning real-time dynamic measuring device. , y, v), and dynamically correct the tracking result by the pre-existing average error distribution function f (x, y, v), which can generate the information of the post-fusion obstacle (x", y" with less error and higher reliability. , v"). Furthermore, the cumulative value of the fusion error variation 342 is corrected by the prior detection probability information 352 as a basis for determining the presence or absence of the obstacle 120, which can greatly increase the reliability of the determination result and solve the sensing error of the prior art. Too big and too low reliability.

Please refer to FIGS. 5A-5C, 6, 7, 8A and 8B. FIG. 7 is a schematic flow chart of the error and detection probability analysis method 400a applied to the multi-sensor fusion according to another embodiment of the present invention. . Fig. 8A is a diagram showing the fusion error variation amount 342 of the fusion error variation amount cumulative correction step S26 in Fig. 7. Fig. 8B is a diagram showing the cumulative value of the fusion error variation amount 342 of Fig. 8A. As shown in FIG. 7, the error and detection probability analysis method 400a for multi-sensor fusion includes an obstacle sensing step S21, an obstacle prediction step S22, an error model offline establishing step S23, and a tracking fusion step S24. The detection confidence degree establishing step S25, the fusion error change amount cumulative correction step S26, and the collision time calculation step S27 are performed. abovementioned The obstacle sensing step S21, the obstacle prediction step S22, and the error model offline establishing step S23 are the same as the obstacle sensing step S12, the obstacle prediction step S14, and the error model offline establishing step S16 of FIG. 2, and will not be described again. In particular, the error and detection probability analysis method 400a applied to the multi-sensor fusion includes a tracking fusion step S24, a detection confidence establishment step S25, a fusion error variation amount accumulation correction step S26, and a collision time calculation step S27.

The tracking and fusion step S24 utilizes the error model offline establishment module 330 of the processor 300 to integrate the obstacle observation information (x, y, v) and the obstacle prediction information (x', y', v') with a tracking fusion method and The pre-existing average error distribution function f (x, y, v) generates a plurality of fusion error variations 342 and a plurality of post-fusion obstacle information (x", y", v"). The tracking fusion method of this embodiment is Carl. The Man filter algorithm has two types of sensors 200, which are respectively a radar sensor and a camera.

The detection confidence establishing step S25 uses the detection confidence setting module 350 of the processor 300 to respectively establish a plurality of pre-detection probability information 352 according to different sensors 200. Moreover, the error model offline establishing step S23 has a certain correlation with the detection confidence establishing step S25, and the variation of the prior average error distribution function f (x, y, v) is larger (that is, the sensor 200 is less sensitive) Accurate), the lower the detection confidence represented by the pre-recovery probability information 352; in other words, the smaller the variation of the prior mean error distribution function f (x, y, v) (ie, the sensor 200 sense) The more accurate the measurement, the higher the detection confidence that the pre-detection probability information 352 represents, that is, the higher the reliability of the sensor 200.

The fusion error variation cumulative correction step S26 is based on the fusion error variation cumulative correction module 360 of the processor 300 to correct the fusion error variation 342 of the tracking fusion step S24 according to the prior detection probability information 352 of the detection confidence establishment step S25. The values are accumulated to determine if the obstacle 120 is present. In detail, in the fusion error change amount cumulative correction step S26, the fusion error change amount cumulative correction module 360 of the processor 300 stores a preset cumulative threshold value, and the fusion error variation amount cumulative correction module 360 compares the presets. The cumulative threshold value and the cumulative value are used to determine whether the obstacle 120 exists; when the accumulated value is less than or equal to the preset cumulative threshold, the obstacle 120 is considered to exist; otherwise, when the accumulated value is greater than the preset cumulative threshold, the obstacle 120 is considered not to exist. Furthermore, the fusion error variation cumulative correction module 360 corrects the cumulative value of the fusion error variation 342 according to the prior detection probability information 352, and outputs the obstacle 120 corresponding to the merged obstacle information (x", y", v" The presence information 362. For example, in the fifth, eighth, and eighth diagrams, the two types of sensors 200 of the time T19 to T24 cannot obtain the observed position (x, y) of the obstacle 120. In the two possible situations, the first possible condition is that the two sensors 200 fail simultaneously, that is, both the radar sensor and the camera fail, and the sensor 200 has a lower detection probability information 352 (for example) : The pre-detection probability information 352 is 30%); that is, the detection confidence of the sensor 200 is low. The second possible condition is that the two sensors 200 are normal and the obstacle 120 is disturbed by noise. When the observation position (x, y) is changed, the pre-detection probability information 352 of the sensor 200 is higher (for example, the pre-detection probability information 352 is 90%); that is, the sensor 200 detects Test confidence is higher. The fusion error variation 342 of Fig. 8A is a positive value at times T19 to T24, which represents that the error persists. At the same time, the cumulative value of the fusion error variation 342 of Fig. 8B is continuously accumulated at times T19 to T24, resulting in an increasing cumulative value. Assuming that the system sets the preset cumulative threshold to 2, it can be seen from Fig. 8B that the accumulated value exceeds 2, which means that "the sensor 200 cannot sense the obstacle 120 (the first possible condition)" or "the obstacle 120 does not." There is a (second possible condition), and the actual situation of which is the possible condition needs to be comprehensively judged together with the pre-detection probability information 352 (ie, the detection confidence) of the sensor 200. When the cumulative value is less than or equal to 2, it means that the obstacle 120 exists.

The collision time calculation step S27 uses the collision time calculation module 370 of the processor 300 to receive the post-fusion obstacle information (x", y", v") and the presence status information 362 to calculate the collision time of the vehicle 110 and the obstacle 120. The collision time can be used as a judgment parameter of the automatic driving. Thereby, the error and detection probability analysis method 400a of the present invention applied to the multi-sensor fusion method uses the pre-detection probability information 352 to correct the accumulated value of the fusion error variation 342. As the basis for judging the presence or absence of the obstacle 120, not only can the reliability of the judgment result be greatly increased, but also the Autonomous Emergency Braking System (AEB) and the Autonomous Driving System (ADS) can be effectively applied. on.

It can be seen from the above embodiments that the present invention has the following advantages: First, through the pre-processed sensor error analysis combined with the detection confidence model, the information of the pre-detection probability between the multiple sensors is integrated and repaired. The cumulative value of the amount of error variation is positively obtained, and a more reliable obstacle existence judgment can be obtained, and an instant operation result can be obtained. Secondly, the global average positioning error distribution function under various environmental, obstacle and vehicle conditions is pre-established by the global satellite positioning real-time dynamic measuring device, and the tracking result is dynamically corrected by the pre-average error distribution function, which can generate less error. And more reliable information on post-integration obstacles. Thirdly, the cumulative value of the fusion error is corrected by detecting the probability of the confidence detection step to detect the cumulative value of the fusion error, which can greatly increase the reliability of the judgment result and solve the conventional technology. The problem of excessive sensing error and low reliability.

Although the present invention has been disclosed in the above embodiments, it is not intended to limit the present invention, and the present invention can be modified and modified without departing from the spirit and scope of the present invention. The scope is subject to the definition of the scope of the patent application attached.

Claims (10)

An error and detection probability analysis method for multi-sensor fusion is used to determine an obstacle in a traveling direction of a vehicle, and the error and detection probability analysis method for multi-sensor fusion includes the following steps An obstacle sensing step is to provide a plurality of sensors to sense the obstacle to respectively generate a plurality of obstacle observation information; and an obstacle prediction step is to provide a processor to generate a plurality of the respective pieces according to the obstacle observation information Obstacle prediction information; an error model offline establishing step, the processor is used to establish a complex average error distribution function according to the sensors; a detection confidence establishing step is performed by the processor according to the sensors Establishing a plurality of pre-detection probability information respectively; a tracking fusion step is to use the processor to combine the obstacle observation information, the obstacle prediction information, and the average error distribution function to generate a complex fusion error by using a tracking fusion method The amount of change and the information of the obstacle after the fusion; and a cumulative error correction step of the fusion error The processor corrects the accumulated value of the fusion error according to the information of the pre-detection probability to determine whether the obstacle exists; wherein the obstacle observation information includes an observation position and An observation speed; wherein the error model offline establishing step pre-sets an instant dynamic positioning module on the obstacle and drives the instant dynamic positioning module to generate a plurality of instantaneous dynamic positioning positions, and then the processor receives and calculates the some The average error distribution function is generated by the relative error between the instantaneous dynamic positioning position and the observed positions; wherein the sensor has a field of view, and the instantaneous dynamic positioning position and the observation positions are located in the field of view and correspond to The speed of observation of the obstacle. The error and detection probability analysis method for multi-sensor fusion as described in claim 1, wherein in the tracking fusion step, the tracking fusion method is a Kalman filter algorithm, and each of the The post-fusion obstacle information includes a post-fusion obstacle position, a post-fusion obstacle speed, and a post-fusion obstacle type. For example, the error and the detection probability analysis method applied to the multi-sensor fusion according to the first aspect of the patent application, wherein the sensor is a radar sensor, and the other sensor is a camera. The error and detection probability analysis method for multi-sensor fusion as described in claim 1, wherein the processor stores a preset cumulative threshold value in the fusion error variation cumulative correction step, And determining, by the processor, whether the obstacle exists according to the preset cumulative threshold value and the accumulated value; wherein, when the accumulated value is less than or equal to the preset cumulative threshold, the obstacle is regarded as being present; When the accumulated value is greater than the preset cumulative threshold, the obstacle is considered to be absent. An error and detection probability analysis method for multi-sensor fusion is used to determine an obstacle in a traveling direction of a vehicle, and the error and detection probability analysis method for multi-sensor fusion includes the following steps An obstacle sensing step is to provide a plurality of sensors to sense the obstacle and respectively generate a plurality of obstacle observation information, wherein the obstacle observation information includes an observation position and an observation speed; and an obstacle prediction step, Providing a processor to generate complex obstacle prediction information according to the obstacle observation information; an error model offline establishing step, the processor is used to establish a complex average error distribution function according to the sensors, and the error model The offline establishing step pre-sets an instant dynamic positioning module on the obstacle and drives the instant dynamic positioning module to generate a plurality of instantaneous dynamic positioning positions, and then the processor receives and calculates the instantaneous dynamic positioning positions and the observations. The relative error distribution function is generated by the relative error of the position; and a tracking fusion step is performed The processor combines the obstacle observation information, the obstacle prediction information, and the average error distribution function to generate complex fusion obstacle information by using a tracking fusion method; wherein each of the sensors has a field of view, and the The instantaneous dynamic positioning position and the observation positions are both located in the field of view and correspond to the observation speed of the obstacle. The error and detection probability analysis method for multi-sensor fusion as described in claim 5, wherein in the tracking fusion step, the tracking fusion method is a Kalman filter algorithm, and each of the The post-fusion obstacle information includes a post-fusion obstacle position, a post-fusion obstacle speed, and a post-fusion obstacle type. As described in claim 5, the error and the detection probability analysis method applied to the multi-sensor fusion, one of the sensors is a radar sensor, and the other sensor is a camera. The method for analyzing the error of the multi-sensor fusion and the detection probability analysis method as described in claim 5, further comprising: a step of establishing confidence detection, which is respectively established by the processor according to the sensors The plurality of pre-detection probability information; wherein, in the tracking fusion step, the processor combines the obstacle observation information, the obstacle prediction information, and the average error distribution function to generate a complex fusion error by using a tracking fusion method The amount of change. The error and detection probability analysis method for multi-sensor fusion as described in claim 8 of the patent application scope includes: a fusion error variation cumulative correction step, which is based on the pre-detection probability of the processor. The information corrects the accumulated value of the fusion error amount to determine whether the obstacle exists. The error and detection probability analysis method for multi-sensor fusion as described in claim 9 wherein the processor stores a preset cumulative threshold value in the fusion error variation cumulative correction step. And determining, by the processor, whether the obstacle exists according to the preset cumulative threshold value and the accumulated value; wherein, when the accumulated value is less than or equal to the preset cumulative threshold, the obstacle is regarded as being present; When the accumulated value is greater than the preset cumulative threshold, the obstacle is considered to be absent.