A multi-sensor obstacle presence probability fusion method and vehicle

Abstract

The present invention discloses a multi-sensor obstacle presence probability fusion method and a vehicle. The fusion module receives the obstacle detection results sent by multiple upstream sensors after detection, performs matching and association, and then updates the tracking of the obstacle presence probability based on a probability model. The false alarm rate and missed detection rate of the preset sensors are considered, and the sensor characteristics, environmental factors, obstacle status, and distance are comprehensively considered to improve the vehicle's obstacle detection and tracking performance in dynamic scenarios, enhance the robustness and adaptability, and improve the real-time performance and response ability of the ego-vehicle system.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent driving, and particularly relates to a multi-sensor obstacle presence probability fusion method and a vehicle. Background Art

[0002] The perception module in an intelligent driving system is a core component, mainly responsible for collecting and understanding the environmental information around the vehicle. Since different types of sensors have their own advantages and limitations, a single sensor is difficult to handle complex and changing driving scenarios. Therefore, modern intelligent driving systems usually integrate multiple sensors, such as lidar (LiDAR), millimeter-wave radar, cameras, and ultrasonic sensors, etc.

[0003] With the rapid development of deep learning technology, the intelligent driving system can directly output obstacle detection results relying on a deep learning network. However, due to hardware limitations, a single deep learning network is difficult to efficiently process all types of sensor information and thus generate a unified obstacle detection result. Therefore, the system usually uses multiple networks to process the information of different sensors and then outputs multiple obstacle detection results. At this time, the fusion module performs result-level fusion processing based on the detection results of multiple sensors, and through fusion tracking technology, generates accurate, stable, and unified obstacle information.

[0004] The fusion module generally maintains a set of obstacle track lists. First, through an association algorithm, the module matches the detection results of different sensors with the existing tracks in the track list. If the detection result is associated with an existing track, the track is updated; if it is not associated with any existing track, a new target is initialized and added to the track list. Finally, the result provided by the fusion module to the downstream module is the high-quality tracks screened from the track list.

[0005] However, not all obstacle tracks can be directly sent to the downstream module. Obstacles in a track do not necessarily mean they actually exist; they may also be misdetections by sensors. Obstacles not in the track list do not necessarily mean they do not exist; they may also be missed detections by sensors. Therefore, it is necessary to model the existence of obstacles in the track in order to better express the existence and manage the life cycle. In traditional methods, the tracking count is usually used as an indicator, and a counter is used to evaluate the existence of the track. Specifically, a counter is set. Whenever a track is successfully associated with a certain detection result, the counter increases; conversely, if it fails to be associated with a new detection result, the counter decreases. When the counter exceeds the set threshold, it is considered to exist and can be sent to the downstream; otherwise, it is considered to disappear and the track is deleted. In this way, the system can dynamically evaluate and decide whether to send it to the downstream module or delete the track. However, this technique has the following disadvantages:

[0006] 1. Sensor differences: Different types of sensors (such as lidar, millimeter-wave radar, cameras, etc.) have different detection capabilities and error characteristics. Simply relying on the number of tracking times cannot fully reflect these differences, which may lead to underestimation of some high-quality but low-detection-frequency tracks.

[0007] 2. Environmental impact: Different sensors have different sensitivities to environmental changes. For example, cameras may perform poorly at night or under strong light conditions, while millimeter-wave radar can still maintain good performance in rainy or snowy weather. Counting methods based on fixed thresholds are difficult to adapt to these changes, which may affect the accuracy of tracking quality assessment.

[0008] 3. Adaptability to dynamic scenarios: In complex and dynamic driving environments, the motion patterns and appearance frequencies of targets may change rapidly. Fixed technical thresholds may not be able to flexibly respond to these changes, resulting in misjudgments or missed detections.

[0009] 4. Uncertainty in the initial stage: Newly initialized tracks may exhibit significant uncertainty in the early stage due to insufficient data. Simply relying on the number of tracking times may prematurely trust these tracks and affect the overall performance of the system. Summary of the Invention

[0010] In view of the above problems, the main objective of the present invention is to design a multi-sensor obstacle presence probability fusion method and vehicle, which comprehensively combines the detection results of multiple sensors, and manages the life cycle of obstacles through the presence probability value of obstacles and the threshold of probability values during the modeling and fusion process, so as to solve the problems of sensor differences, environmental factors, and adaptability to dynamic scenarios.

[0011] To achieve the above objective, the present invention adopts the following technical solutions:

[0012] A multi-sensor obstacle presence probability fusion method, which is applied to a vehicle. The vehicle includes multiple sensors that detect the same area; based on a fusion module, the detection results of multiple sensors are fused and output. A set of obstacle track lists is maintained inside the fusion module, and this track list is defined as the system track.

[0013] The fusion method includes the following steps:

[0014] The fusion module receives the obstacle detection results sent by multiple upstream sensors after detection.

[0015] Through an association matching algorithm, the detection results of each sensor are associated and matched with the obstacles in the track list; for the detection results of the obstacles successfully associated with the obstacles in the system track, the tracking of the obstacles in the system track is updated; for the detection results that are not successfully associated with the obstacles in the system track, a new target is initialized and added to the track list.

[0016] Generate a new system track based on initialization and tracking updates, filter within the new system track, and output unified obstacle results under multiple sensors to the downstream.

[0017] As a further description of the present invention, the obstacle detection results sent after the above sensor detections include position and shape;

[0018] The tracking update of the obstacle includes position update, shape update, and existence probability update, and the initialization of a new target includes position initialization, shape initialization, and existence probability initialization;

[0019] Among them, the above existence probability initialization and existence probability update are performed by constructing a probability model.

[0020] As a further description of the present invention, constructing a probability model includes the following steps:

[0021] Define the detection result of a single sensor for an obstacle as a binary value :

[0022] ;

[0023] Define the true existence of the obstacle as a binary value :

[0024] ;

[0025] Determine the prior probability: and are respectively and the prior probabilities of occurrence, and ;

[0026] Define the relevant probability of the detection result:

[0027] False alarm probability ;

[0028] Miss detection probability ;

[0029] Detection probability ;

[0030] Correct rejection probability ;

[0031] Calculate the posterior probability according to Bayes' theorem, and the expression is:

[0032] ;

[0033] Among them, is the prior probability of the true existence of the obstacle, is the probability of the sensor detection result; is the likelihood probability, indicating that the obstacle actually exists as in the case of, the sensor detects as the probability; is the posterior probability, indicating that when the current sensor detection result is , the probability that the obstacle actually exists as ;

[0034] When , the obstacle actually exists, and the expression is:

[0035] ;

[0036] Combined with Bayes' theorem:

[0037] ;

[0038] Taking out the likelihood probability, the likelihood ratio is:

[0039] ;

[0040] According to the ratio of the likelihood ratio to the prior probability to determine the existence of the obstacle;

[0041] In the case of multiple sensors, there are detection results , then the likelihood ratio under multiple sensors is:

[0042] ;

[0043] ;

[0044] ;

[0045] ;

[0046] Among them, is the number of sensors that detect the target, is the th sensor correctly detects the obstacle when the obstacle actually exists, is the th sensor fails to detect the obstacle when the obstacle actually exists, is the th sensor wrongly detects the obstacle when the obstacle actually does not exist;

[0047] Taking the logarithm of the above likelihood ratio, we get:

[0048] ;

[0049] For each sensor, based on the detection result , define the weight , and the expression is:

[0050] ;

[0051] Then:

[0052] ;

[0053] ;

[0054] Assume , then the above is simplified to judge whether is greater than 0;

[0055] , then the obstacle exists;

[0056] , then the obstacle does not exist.

[0057] As a further description of the present invention, the obstacles within the system track are set with a probability attribute, and the preset value of the existence probability of the obstacle is set to , and the upper limit value is set to ;

[0058] Before the above existence probability initialization and existence probability update, preset the false alarm rate and missed detection rate of the sensors for different sensors, different types of obstacles, and different environments;

[0059] For the detection result of each sensor, calculate the weight value for updating the existence probability of the obstacle according to the preset false alarm rate and missed detection rate , and the expression is:

[0060] .

[0061] As a further description of the present invention, the existence probability initialization includes the following steps:

[0062] Determine the source and type of the detection result according to the detection result;

[0063] Calculate the log ratio according to the preset false alarm rate and missed detection rate ;

[0064] Initialize the existence probability , and the expression is:

[0065] ;

[0066] The calculated is saved to the system track as the existence probability of the new detected target.

[0067] As a further description of the present invention, after the process of the association matching algorithm, for the detection of obstacles associated with the system track, a detection list is maintained to save all detections of the obstacle system track.

[0068] As a further description of the present invention, the existence probability update includes the following steps:

[0069] An obstacle within the system track , define its existence probability before update as ;

[0070] Traverse multiple sensors one by one according to the following steps:

[0071] S1: For each sensor , based on the position, speed, and shape of the obstacle compare with the field of view of the sensor to determine whether the obstacle is within the FOV of the sensor;

[0072] S2: Obtain the preset false alarm rate and miss detection rate corresponding to the sensor ;

[0073] S3: If the obstacle is detected within the FOV of the sensor , and the obstacle is in the detection list, then calculate the weight value according to the preset false alarm rate and miss detection rate under this sensor , and accumulate it into the total weight;

[0074] If the obstacle is not in the detection list , then calculate the weight value according to the preset false alarm rate and false detection rate under this sensor

[0075] S4: If the obstacle is not detected within the FOV of the sensor , then no calculation is performed and 0 is returned;

[0076] After traversing all sensors, add the accumulated result to the existence probability before update to obtain the updated existence probability , and the expression is:

[0077] ;

[0078] Among them, is the total number of sensors, is the number of sensors that detect obstacles ; is the number not within the FOV of the corresponding sensor;

[0079] Save the detection results of the above sensors into the detection list.

[0080] As a further description of the present invention, the generation of a new system track includes the following steps:

[0081] Set the deletion threshold of the obstacle existence probability ;

[0082] After initialization and tracking update, traverse the existence probabilities of all obstacles in the system track at this time , compare it with the deletion threshold , when , delete the obstacle;

[0083] After the traversal, the system track is used as the new system track.

[0084] As a further description of the present invention, the screening and distribution of obstacles in the new system track include the following steps:

[0085] Based on the upper limit value of the obstacle existence probability and the deletion threshold of the existence probability Set the threshold , ;

[0086] Compare the existence probability of the obstacle in the new system track with the threshold , when , it is considered that the existence state of the current obstacle is reliable, and output the information of the obstacle to the downstream for processing or decision-making; when , select to retain the track of the obstacle but further observe.

[0087] A vehicle includes a plurality of sensors for detecting the same area and a fusion module. The fusion module executes the above-mentioned obstacle existence probability fusion method, and through the obstacle existence probability fusion method, obstacles are tracked under multiple sensors.

[0088] Compared with the prior art, the technical effect of the present invention is:

[0089] The present invention provides a multi-sensor obstacle presence probability fusion method and a vehicle. The fusion module receives the obstacle detection results sent by multiple upstream sensors. After matching and correlation, the presence probability of the obstacle is tracked and updated based on a probability model. The false alarm rate and missed detection rate of the preset sensors are considered, and the sensor characteristics, environmental factors, obstacle state, and distance are comprehensively considered to improve the detection and tracking performance of the vehicle for obstacles in a dynamic scenario, enhance the robustness and adaptability, and improve the real-time performance and response ability of the vehicle system. BRIEF DESCRIPTION OF THE DRAWINGS

[0090] Figure 1 It is a schematic diagram of the probability fusion method of the present invention;

[0091] Figure 2 It is a block diagram of information transmission of the present invention;

[0092] Figure 3 It is a schematic diagram of the presence probability initialization process of the present invention;

[0093] Figure 4 It is a schematic diagram of the presence probability update process under a single sensor of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0094] The present invention will be described in detail below with reference to the accompanying drawings:

[0095] In an embodiment of the present invention, a multi-sensor obstacle presence probability fusion method is disclosed. Referring to Figures 1-4 as shown, for the detection of obstacles, only the existence is described in this embodiment, and no discussion is made on its detailed detection results (such as length, width, height, etc.).

[0096] Specifically, in this embodiment, the method is applied to a vehicle. The vehicle includes multiple sensors that detect the same area. After fusing the detection results of the multiple sensors based on a fusion module, the fusion module internally maintains a set of obstacle track lists, and this track list is defined as a system track;

[0097] The fusion method includes the following steps:

[0098] The fusion module receives the obstacle detection results sent by multiple upstream sensors;

[0099] Through an association matching algorithm, the detection results of each sensor are associated and matched with the obstacles in the track list. For the detection results of the obstacles successfully associated with the system track, the tracking update of the obstacles in the system track is performed. For the detection results of the obstacles not successfully associated with the system track, a new target is initialized and added to the track list;

[0100] Generate a new system track based on initialization and tracking updates, filter within the new system track, and output unified obstacle results under multiple sensors to the downstream.

[0101] It should be noted that the above-mentioned association matching algorithm includes but is not limited to the Hungarian matching, and also includes other algorithms that can realize the association matching between the sensor detection results and the obstacles in the track list.

[0102] In this embodiment, the above-mentioned fusion method will be described in detail as follows:

[0103] The obstacle detection results sent after the above sensor detections include but are not limited to the position and shape; the tracking update of the obstacle includes position update, shape update, and existence probability update, and the initialization of a new target includes position initialization, shape initialization, and existence probability initialization; among them, the above-mentioned existence probability initialization and existence probability update are carried out by constructing a probability model.

[0104] In this embodiment, for the fusion detection algorithm based on the maximum a posteriori probability, the steps for constructing the probability model are as follows:

[0105] Define the detection result of a single sensor for an obstacle as a binary value :

[0106] ;

[0107] Define the true existence of the obstacle as a binary value :

[0108] ;

[0109] Determine the prior probability: and are respectively and the prior probabilities of occurrence, and .

[0110] At this time, four situations will occur:

[0111] Two correct situations:

[0112] is true, and the detection result , correct rejection, there is actually nothing and it is not detected either;

[0113] is true, and the detection result 1, correct detection, there is actually something and it is detected at the same time;

[0114] Two wrong situations:

[0115] is true, the detection result , is the first type of error: false alarm (there is actually nothing, but it is detected);

[0116] is true, the detection result , is the second type of error: missed detection (there is actually something, but it is not detected).

[0117] Define the relevant probabilities of the detection result:

[0118] False alarm probability ;

[0119] Missed detection probability ;

[0120] Detection probability ;

[0121] Correct rejection probability ;

[0122] Above, according to Bayes' theorem, calculate the posterior probability, the expression is:

[0123] ;

[0124] Among them, is the prior probability of the true existence of the obstacle, is the probability of the sensor detection result; is the likelihood probability, indicating that the true existence of the obstacle is in the case of, the probability that the sensor detects as ; is the posterior probability, indicating that when the current sensor detection result is the true existence of the obstacle is the probability of.

[0125] When judging the binary characteristic of existence, usually take the larger probability value. When , it is considered that the obstacle truly exists, otherwise the obstacle actually does not exist, the expression is:

[0126] ;

[0127] Combined with Bayes' theorem formula:

[0128] ;

[0129] Factor out the likelihood probability, and the likelihood ratio is:

[0130] ;

[0131] According to the ratio of the likelihood ratio to the prior probability Determine the existence of obstacles.

[0132] Under multi-sensors, assuming there are N sensors, then there are detection results Then the likelihood ratio under multi-sensors is:

[0133] ;

[0134] ;

[0135] ;

[0136] ;

[0137] Among them, is the number of sensors that detect the target, is the probability that the -th sensor correctly detects the obstacle when the obstacle actually exists, is the probability that the -th sensor fails to detect the obstacle when the obstacle actually exists, is the probability that the -th sensor wrongly detects the obstacle when the obstacle actually does not exist, represents the -th sensor;

[0138] Taking the logarithm of the above likelihood ratio, we can get:

[0139] ;

[0140] For each sensor, according to the detection result , define the weight , and the expression is:

[0141] ;

[0142] Then:

[0143] ;

[0144] It can be changed to:

[0145] ;

[0146] It should be noted that since the prior probability of the obstacle is the true existence of the obstacle, generally each takes 0.5, that is, there is a 0.5 probability that a certain obstacle actually exists and a 0.5 probability that it does not exist. That is, assuming , then the above is simplified to judge Is it greater than 0; , then the obstacle exists; , then the obstacle does not exist.

[0147] Through the above content, each detection result under multiple sensors is specifically modeled as a probability model associated with the false alarm and missed detection situations of the corresponding sensors , and this model is used to better characterize the existence of obstacles.

[0148] In this embodiment, the specific implementation of the above existence probability initialization and existence probability update is as follows:

[0149] Set the existence probability attribute for the obstacles in the system track, and the preset value of the existence probability of the obstacles is set to , and the upper limit value is set to .

[0150] The detection results sent from upstream are multi-dimensional, and the existence probability only cares about the indicator of whether there is a detection. For different sources, types, and environmental conditions (such as day, night, etc.) of the detection, the corresponding false alarm rates and missed detection rates of the sensor detections are also different. Different false alarm rates and missed detection rates are set for different sensors, different types of obstacles, and different environments according to empirical values (which can also be obtained from the indicators of model training).

[0151] Specifically, before the above existence probability update and existence probability initialization, preset the false alarm rate and missed detection rate of the sensor for different sensors, different types of obstacles, and different environments;

[0152] For the detection result of each sensor, calculate the weight value for updating the existence probability of the obstacle according to the preset false alarm rate and missed detection rate , and the expression is:

[0153] .

[0154] In this embodiment, after the process of the association matching algorithm, for the detections that are not associated with any obstacles in the system track, it is necessary to initialize their existence probabilities. As Figure 3 shown, the existence probability initialization includes the following steps:

[0155] Determine the source and type of the detection result according to the detection result;

[0156] Calculate the log ratio according to the preset false alarm rate and missed detection rate ;

[0157] Initialize the existence probability , and the expression is:

[0158] ;

[0159] Save the calculated to the system track as the existence probability of the new detected target.

[0160] In this embodiment, after the process of the association matching algorithm, for the detection of obstacles associated with the system track, a detection list is maintained to save all detections of the obstacle system track. For example, for an obstacle in the system track , before the existence probability update is , among a total of sensors, sensors detect it, then save the detections of these sensors to the list for subsequent confirmation.

[0161] Specifically, as Figure 4 shown, the update of the existence probability of the above-mentioned obstacle includes the following steps:

[0162] For an obstacle in the system track , define its existence probability before update as ;

[0163] Traverse multiple sensors for one-by-one processing according to the following steps:

[0164] S1: For each sensor , compare the information such as the position, speed, and shape of the obstacle with the field of view of the sensor to determine whether the obstacle is within the FOV of the sensor;

[0165] S2: Obtain the preset false alarm rate and missed detection rate corresponding to the sensor ;

[0166] S3: If the obstacle is detected within the FOV of the sensor , it means that the sensor can theoretically observe it. If the obstacle is in the detection list at this time, it means that the sensor has normally detected the obstacle. Then, calculate the weight value according to the preset false alarm rate and missed detection rate under this sensor , and accumulate it into the total weight; if the obstacle is not in the detection list at this time, it means that the obstacle has not been normally detected. Then, calculate the weight value according to the preset false alarm rate and false detection rate under this sensor , and accumulate it into the total weight;

[0167] S4: If the obstacle is not detected within the FOV of the sensor , it means that the obstacle is not observable under this sensor. Then, no calculation is performed at this time, and 0 is returned;

[0168] Repeat the above steps for each sensor. After traversing all sensors, add the accumulated result to the existence probability before update to obtain the updated existence probability . The expression is:

[0169] ;

[0170] where is the total number of sensors, is the number of sensors that detect the obstacle , and is the number of those not within the corresponding sensor's FOV;

[0171] Save the detection results of the above sensors into the detection list.

[0172] It should be noted that the above existence probability is the existence probability after multiple association updates after initialization , and it is the current existence probability of the obstacle.

[0173] In this embodiment, according to the characteristics of the ln function, false alarm rate, and miss detection rate, if there is a real obstacle, the existence probability increases positively; if there is no obstacle, the existence probability decreases inversely.

[0174] In this embodiment, the generation of a new system track includes the following steps:

[0175] Set the deletion threshold of the obstacle existence probability ;

[0176] After initialization and tracking update, traverse the existence probabilities of all obstacles in the system track at this time , compare them with the deletion threshold . When , delete this obstacle;

[0177] After the traversal, the system track is used as the new system track.

[0178] In this embodiment, the screening and distribution of obstacles in the new system track include the following steps:

[0179] Based on the upper limit value of the obstacle existence probability and the deletion threshold of the existence probability Set a threshold , ;

[0180] Compare the probability of the presence of an obstacle in the new system track with the threshold . When , it is considered that the presence state of the current obstacle is reliable, and the information of the obstacle is output to the downstream for processing or decision-making; when , select to retain the track of the obstacle but further observe.

[0181] It should be noted that the above threshold parameters can be dynamically adjusted according to the state of the host vehicle, the state of the obstacle, etc. In addition, when an obstacle is stably tracked, its probability of presence cannot increase infinitely, and the upper limit value is ; when the obstacle disappears and is not detected, its probability of presence cannot decrease infinitely, and the set deletion threshold is . Those less than this value are directly deleted from the system track.

[0182] A vehicle, which includes multiple sensors and a fusion module for detecting the same area, and the fusion module executes the above-mentioned obstacle presence probability fusion method, and through the obstacle presence probability fusion method, obstacles are tracked under multiple sensors.

[0183] Through the above content, the present invention discloses an existence probability fusion method and a vehicle. Compared with the prior art, the present invention has the following advantages:

[0184] 1. The existence probability fusion method of the present invention adopts a fusion detection algorithm of maximum a posteriori probability to design the obstacle existence probability strategy in the fusion process, providing a more accurate obstacle existence evaluation;

[0185] 2. The existence probability fusion method of the present invention associates the obstacle existence probability with the obstacle state, sensor characteristics, etc., can perform sub-divided modeling, increases the flexibility and adaptability of the model, improves the robustness, and has strong adaptability to dynamic scenarios;

[0186] 3. The existence probability fusion method of the present invention manages the life cycle of obstacles in the system track during the post-fusion process through the set threshold, which helps to improve the accuracy and efficiency of obstacle tracking. By setting different release threshold parameters, it can adapt to complex and dynamic driving environments, meet the rapidly changing requirements, and enhance the real-time performance and response ability of the system;

[0187] 4. The existence probability fusion method of the present invention fully considers sensor differences and environmental impacts, improves the accuracy of detection, and improves the robustness and adaptability of the system;

[0188] 5. The existence probability fusion method of the present invention, and the design of false alarm rate and missed detection rate parameters are obtained from the detection indexes of the upstream model, which simplifies the process of parameter acquisition.

[0189] The above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Any other modifications or equivalent replacements made by those of ordinary skill in the art to the technical solutions of the present invention shall be covered within the scope of the claims of the present invention as long as they do not depart from the spirit and scope of the technical solutions of the present invention.

Claims

1. A multi-sensor obstacle presence probability fusion method, characterized by: The method is applied in a vehicle, the vehicle comprising a plurality of sensors for detecting the same area; The fusion module is based on which a fusion module performs fusion of detection results of multiple sensors and outputs the fusion of the detection results. The fusion module internally maintains a set of obstacle track lists, which are defined as system tracks. The fusion method comprises the following steps: The fusion module receives obstacle detection results sent down by multiple upstream sensors, including location and shape; Through the association matching algorithm, the detection results of each sensor are associated and matched with the obstacles in the track list; According to the detection results of obstacles successfully associated with the system track, the obstacles within the system track are tracked and updated; For the detection results that are not successfully associated with obstacles in the system track, a new target is initialized and added to the track list; The tracking update of obstacles includes position update, shape update, and existence probability update. The initialization of new targets includes position initialization, shape initialization, and existence probability initialization. Among them, the above existence probability initialization and existence probability update are performed by constructing a probability model. The probability model is based on a fusion detection algorithm of maximum a posteriori probability. The detection result of a single sensor and the real existence of an obstacle are first defined as binary values, the prior probability is determined, and the false alarm probability, missed detection probability, detection probability and correct rejection probability are defined at the same time. The posterior probability is then calculated based on the Bayesian theorem, and the ratio of the posterior probability is converted into the ratio of the comparative likelihood ratio and the prior probability to determine the existence of the obstacle under a single sensor. Under multi-sensor, calculate the logarithm of the multi-sensor likelihood ratio and define the existence probability weight of a single sensor. Define the prior probability of the existence and non-existence of an obstacle as 0.

5. Then, compare the existence probability weight sum with 0 to determine whether the obstacle exists. If the existence probability weight sum is greater than 0, the obstacle exists; if the existence probability weight sum is less than 0, the obstacle does not exist. Generate a new system track based on initialization and tracking update, filter within the new system track, and output unified obstacle results under multiple sensors to the downstream.

2. The multi-sensor obstacle existence probability fusion method according to claim 1, characterized in that: Building a probability model includes the following steps: Define the detection result of a single sensor for an obstacle as a binary value : ； Define the real existence of an obstacle as a binary value : ； Determine the prior probability: and They are and The prior probability of occurrence, and ; Define the associated probabilities of the test results: False alarm probability ; Missed detection probability ; Probability of detection ; Correct rejection probability ; According to Bayes' theorem, the posterior probability is calculated as follows: ； in, is the prior probability of the real existence of the obstacle, is the probability of the sensor detection result; is the likelihood probability, indicating that the obstacle really exists. In this case, the sensor detects The probability of is the posterior probability, indicating that the current sensor detection result is When the obstacle actually exists The probability of when When , the obstacle really exists, and the expression is: ； Combined with Bayes' theorem: ； Taking the likelihood probability into account, we can get the likelihood ratio for: ； According to the ratio of likelihood ratio to prior probability Determine the presence of obstacles; Under multi-sensor, there are Test results , then the likelihood ratio under multi-sensor is: ； ； ； ； in, is the number of sensors that detect the target, For the The probability that a sensor correctly detects an obstacle when the obstacle actually exists, For the The probability that a sensor fails to detect an obstacle when the obstacle actually exists, For the The probability that a sensor falsely detects an obstacle when the obstacle does not actually exist; Taking the logarithm of the above likelihood ratio, we get: ； For each sensor, according to the detection results , define the weight , the expression is: ； but: ； ； Assumptions , then it is simplified to judge Is it greater than 0? , then the obstacle exists; , then the obstacle does not exist.

3. The multi-sensor obstacle existence probability fusion method according to claim 2, characterized in that: The obstacle in the system track is set with a probability attribute. The default value of the obstacle probability is set to , the upper limit is set to ; Before the above existence probability initialization and existence probability update, the false alarm rate and missed detection rate of the sensor are preset for different sensors, different types of obstacles, and different environments; For each sensor’s detection result, calculate the weight value of the sensor’s probability update for obstacle presence based on the preset false alarm rate and missed detection rate. , the expression is: 。 4. The multi-sensor obstacle existence probability fusion method according to claim 3 is characterized by: The existence probability initialization includes the following steps: Based on the test results, determine the source and type of the test results; Calculate the logarithmic ratio according to the preset false alarm rate and missed detection rate ; Initialize existence probability , the expression is: ； The calculated Save it to the system track as the existence probability of the new target after initialization.

5. A multi-sensor obstacle existence probability fusion method according to claim 3 or 4, characterized in that: After the process of associating the matching algorithm, for the detection associated with the obstacle in the system track, a detection list is maintained to save all the detections of the obstacle system track.

6. The multi-sensor obstacle existence probability fusion method according to claim 5, characterized in that: The existence probability update includes the following steps: Obstacles within the system track , define its existence probability before updating as ; Follow the steps below to traverse multiple sensors and process them one by one: S1: For each sensor , based on obstacles Position, speed, shape and sensors Compare the visual range to determine obstacles Whether it is within the sensor FOV; S2: Get sensor The corresponding preset false alarm rate and missed detection rate; S3: If the obstacle In the sensor The obstacle is detected within the FOV and is in the detection list. , then according to the sensor Under the preset false alarm rate and missed detection rate, calculate the weight value , and add it to the total weight; The obstacle is not in the detection list. , then according to the sensor Under the preset false alarm rate and false detection rate, calculate the weight value , and add it to the total weight; S4: If the obstacle In the sensor If it is not detected within the FOV, no calculation is performed and 0 is returned; After traversing all sensors, the accumulated results are compared with the existence probability before the update Add together to get the updated existence probability , the expression is: ； in, is the total number of sensors, To detect obstacles The number of sensors, is the number of items that are not in the corresponding sensor FOV; The above The detection results of each sensor are saved in the detection list.

7. The multi-sensor obstacle existence probability fusion method according to claim 6, characterized in that: The new system track generation includes the following steps: Set the obstacle probability deletion threshold ; After initialization and tracking update, the existence probability of all obstacles in the system track is traversed. , and compare it with the deletion threshold For comparison, , delete the obstacle; After the traversal is completed, the system track is used as the new system track.

8. The multi-sensor obstacle existence probability fusion method according to claim 7, characterized in that: The obstacle screening and issuance in the new system track includes the following steps: Upper limit based on obstacle existence probability Deletion threshold with existence probability Setting the Threshold , ; The probability of obstacles existing in the new system track With threshold For comparison, When , the existence of the current obstacle is considered reliable, and the information of the obstacle is output to the downstream for processing or decision-making; If the obstacle is found, the user chooses to keep the track of the obstacle but observe it further.

9. A vehicle, characterized in that: The invention comprises a plurality of sensors and a fusion module for detecting the same area, wherein the obstacle existence probability fusion method according to any one of claims 1 to 8 is executed by the fusion module, and the obstacle existence probability fusion method is used to track obstacles under multiple sensors.

Images