Behavior prediction device

Abstract

The present invention relates to a behavior prediction device. The behavior prediction device performs distribution calculation processing, similarity level calculation processing, similarity level evaluation processing, and collision determination processing. In the distribution calculation processing, the probability density distribution related to the future position of the moving body is calculated for each behavior pattern using the position data and speed data of the moving body, their error data, and a plurality of behavior models set for each behavior pattern corresponding to the type of the moving body. In the similarity level calculation processing, the similarity level of the probability density distribution and the reference probability density distribution is calculated for each behavior pattern. In the similarity level evaluation processing, the similarity level is evaluated. The similarity level includes an out-of-range level that is lower than the similarity allowable range. In the similarity level evaluation processing, the behavior pattern with the out-of-range level is excluded from the object of the collision determination processing.

Description

Technical Field

[0001] The present invention relates to a device for predicting future behavior of a moving object around a vehicle detected by the vehicle. Background Art

[0002] Japanese Patent Laid-Open No. 2019-182093 discloses a device for predicting the future behavior of a moving body around a vehicle. The previous device uses a behavior prediction model to predict the future behavior of a moving body detected by a vehicle. The behavior prediction model is pre-set in association with the type of moving body, such as a four-wheeled vehicle, a two-wheeled vehicle, a pedestrian, or a bicycle. The prediction result obtained by the behavior prediction model is compared with the actual detection result of the moving body detected by the vehicle. When the prediction result deviates from the detection result, the previous device infers the reason. The reason for the deviation is also pre-set. The previous device also corrects the behavior prediction model currently in use or switches the behavior prediction model currently in use to a behavior prediction model corresponding to a type different from the behavior prediction model currently in use based on the inferred reason for the deviation.

[0003] Previous devices have focused on the reasons for the estimated deviation. Therefore, the number of behavior prediction models used in future predictions is consistent with the number of types of mobile bodies set in advance. Therefore, in the case where the prediction result deviates from the detection result, the behavior prediction model currently in use is corrected or switched to a behavior prediction model corresponding to another type.

[0004] Consider the case where the emphasis is placed on avoiding collisions with mobile objects. In this case, it is ideal to set a plurality of behavior prediction models for each type of mobile object. This is because, when the number of behavior prediction models is the same as the number of types of mobile objects, the behavior prediction model will be corrected or switched even when the reason for the divergence between the prediction result and the detection result is "an imminent collision situation". This will cause a delay in the prediction of the future behavior of the mobile object in the "imminent collision situation".

[0005] If the number of behavior prediction models is set to be multiple for each type of mobile body, it is possible to grasp in advance whether a vehicle is in a "collision-imminent situation" based on the prediction results obtained by any behavior prediction model. However, in this case, the prediction results obtained by any behavior prediction model may indicate a "collision-imminent situation" even though the vehicle is in a "collision-not-imminent situation" in reality. In this case, the vehicle's collision avoidance action may be started based on the erroneous prediction results. Alternatively, a collision warning may be issued to the driver of the vehicle based on the erroneous prediction results. Therefore, improvements from the perspective of practicality are sought. Summary of the invention

[0006] An object of the present invention is to provide a technique having high practicality when predicting future behaviors of moving objects around a vehicle using a plurality of behavior prediction models set for each type of moving object.

[0007] A first aspect of the present invention is a behavior prediction device that predicts future behavior of surrounding moving objects detected by a vehicle.

[0008] The behavior prediction device includes a database, a memory, and a processor.

[0009] The database stores a plurality of behavior models set for each behavior pattern corresponding to the type of the moving object.

[0010] The memory stores position data and speed data of the moving object and error data indicating an error between the position data and speed data.

[0011] The processor is configured to perform distribution calculation processing, similarity level calculation processing, similarity level evaluation processing, and collision determination processing.

[0012] In the distribution calculation process, the probability density distribution associated with the future position of the mobile body is calculated for each of the behavior patterns using the position data and speed data, the error data, and the plurality of behavior models.

[0013] In the similarity level calculation process, the similarity level between the probability density distribution and a reference probability density distribution is calculated for each of the behavior patterns. The reference probability density distribution represents the probability density distribution calculated using a behavior model corresponding to the current behavior of the moving body at a processing timing prior to the current processing timing of the distribution calculation process.

[0014] In the similarity level evaluation process, the similarity level is evaluated.

[0015] In the collision determination process, a collision between the vehicle and the moving object is determined using the probability density distribution.

[0016] The similar levels include out-of-range levels below a similar allowable range.

[0017] In the similarity level evaluation process, the processor excludes the behavior pattern having the out-of-range level from being a target of the collision determination process.

[0018] The second aspect of the present invention has the following features in addition to the first aspect.

[0019] The similarity level includes a low similarity level that exists within the similarity allowable range and is lower than a low similarity threshold value set within the similarity allowable range.

[0020] The processor is configured to also perform error data modification processing.

[0021] In the error data changing process, the error data is changed based on the similarity level.

[0022] In the error data changing process, the processor enlarges the error data to be used at the next processing timing of the distribution calculation process using a low similarity model, compared to the error data used at the current processing timing of the distribution calculation process. The low similarity model represents a behavior model corresponding to the behavior pattern having the low similarity level.

[0023] The third aspect of the present invention has the following features in addition to the first aspect or the second aspect.

[0024] The similarity level includes a high similarity level that exists within the similarity allowable range and is higher than a high similarity threshold value set within the similarity allowable range.

[0025] The processor is configured to also perform error data modification processing.

[0026] In the error data changing process, the error data is changed based on the similarity level.

[0027] In the error data changing process, the processor reduces the error data to be used at the next processing timing of the distribution calculation process using a high similarity model compared to the error data used at the current processing timing of the distribution calculation process. The high similarity model represents a behavior model corresponding to the behavior pattern with the high similarity level.

[0028] A fourth aspect of the present invention further has the following features in any one of the first to third aspects.

[0029] The similarity level includes a high similarity level that exists within the similarity allowable range and is higher than a high similarity threshold value set within the similarity allowable range.

[0030] The processor is configured to also perform a distributed recalculation process.

[0031] In the distribution recalculation process, the probability density distribution is recalculated using a high similarity model representing a behavior model corresponding to the behavior pattern having the high similarity level.

[0032] In the distribution recalculation process, the processor superimposes the probability density distribution calculated using the high similarity level at the current processing timing and the probability density distribution calculated using the high similarity level at a processing timing before the current processing timing.

[0033] In the collision determination process, the processor performs determination of the behavior pattern having the high similarity level based on the superimposed probability density distribution obtained by the distribution recalculation process.

[0034] According to the first viewpoint, distribution calculation processing, similarity level calculation processing and similarity level evaluation processing are performed. In particular, according to the similarity level evaluation processing, behavior patterns with out-of-range levels can be excluded from the objects of collision determination processing. Therefore, it is possible to suppress the start of collision avoidance actions that are not originally necessary. Therefore, in the case of predicting behavior patterns using various behavior models, the instability of the driving state of the vehicle can be suppressed. In addition, it is also possible to suppress the issuance of collision warnings that are not originally necessary. Therefore, it is also possible to suppress the execution of collision warnings from causing discomfort to the driver. In summary, according to the first viewpoint, a technology with high practicality can be provided.

[0035] According to the second viewpoint, in addition to the similarity level evaluation process, an error data change process is also performed. According to the error data change process, the error data to be used at the next processing timing of the distribution calculation process using the low similarity model is changed to data larger than the error data used at the current processing timing of the distribution calculation process. If the error data becomes larger, the standard deviation of the probability density distribution calculated using the low similarity model at the next processing timing becomes larger. Therefore, the similarity level calculated based on the probability density distribution can be reduced so that it is lower than the allowable range. In this way, the behavior pattern that is evaluated as having a low similarity level at the current calculation timing can be evaluated as a behavior pattern with an out-of-range level at the next calculation timing. As a result, the behavior pattern with a relatively low similarity level can be excluded from the object of the collision determination process.

[0036] According to the third viewpoint, in addition to the similarity level evaluation process, an error data change process is also performed. According to the error data change process, the error data to be used at the next processing timing of the distribution calculation process using a high similarity level is changed to data that is smaller than the error data used at the current processing timing of the distribution calculation process. If the error data becomes smaller, the standard deviation of the probability density distribution calculated using the high similarity model at the next processing timing becomes smaller. Here, in the case where the behavior pattern of the mobile body does not change at each moment, it is expected that the similarity level of the behavior pattern will become a high similarity level. However, since the influence of the error data accumulates with the passage of time, the standard deviation of the probability density distribution expands with the passage of time. Therefore, even in the case where the behavior pattern does not change at each moment, it may be determined that a collision between the vehicle and the mobile body will occur at a position far enough away from the mobile body in the future. In this regard, if the standard deviation of the probability density distribution becomes smaller, such an undesirable situation can be suppressed.

[0037] According to the fourth viewpoint, in addition to the similarity level evaluation process, a distribution recalculation process is also performed. According to the distribution recalculation process, a probability density distribution calculated using a high similarity model at the current processing timing of the distribution calculation process and a probability density distribution calculated using a high similarity model at the processing timing before the current processing timing are superimposed. If superposition is performed, the standard deviation of the probability density distribution after superposition is smaller than the standard deviation of the probability density distribution before superposition. Therefore, according to the distribution recalculation process, the same effect as that obtained when the error data is changed to small data by the error data change process can be expected. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Hereinafter, features, advantages, and technical and industrial significance of exemplary embodiments of the present invention will be described with reference to the accompanying drawings, wherein like reference numerals denote like elements, and wherein:

[0039] Figure 1 This is a diagram showing an example of a situation that the embodiment assumes as a premise.

[0040] Figure 2 This is a diagram showing an example of the probability density distribution of the future position of the reference point of the moving object.

[0041] Figure 3 This is a diagram showing an example of the probability density distribution of the future position of the reference point of the moving object in the behavior pattern corresponding to the constant speed straight travel model (that is, constant speed straight travel).

[0042] Figure 4 This is a diagram showing an example of the probability density distribution of the future position of the reference point of the moving object in the behavior pattern (that is, right turn) corresponding to the right turn model.

[0043] Figure 5 It will be Figure 3 and Figure 4 The various orbits shown in the figure are arranged in such a way that future moments are aligned.

[0044] Figure 6 This is a diagram showing an example of the probability density distribution of the future position of the reference point of the moving object in the behavior pattern (that is, acceleration) corresponding to the acceleration model.

[0045] Figure 7 It will be Figure 3 and Figure 6 The various orbits shown in the figure are arranged in such a way that future moments are aligned.

[0046] Figure 8 is a diagram showing an example of a similar level.

[0047] Fig. 9 It is a diagram showing a configuration example of a behavior prediction device according to an embodiment.

[0048] Fig.10 This is a block diagram showing an example of the functional configuration of a control device associated with future behavior prediction processing.

[0049] Fig.11 is a flowchart showing the flow of similarity level evaluation processing.

[0050] Fig.12 This is a diagram showing an example of distribution recalculation processing. DETAILED DESCRIPTION

[0051] Hereinafter, the embodiments of the present invention will be described with reference to the accompanying drawings. It should be noted that in each figure, the same or corresponding parts are marked with the same reference numerals and their description is simplified or omitted. In addition, the present invention is not limited to the following embodiments and can be implemented in various schemes.

[0052] 1. Purpose of Implementation Methods

[0053] 1-1. Behavioral Model

[0054] Figure 1 1 is a diagram showing an example of a situation that is considered as a premise of the implementation method. Figure 1 , a vehicle VH moving on a lane L1 is depicted. The vehicle VH is equipped with a behavior prediction device 10 of the embodiment. The vehicle VH is, for example, a car using an internal combustion engine such as a diesel engine or a gasoline engine as a power source, an electric car using an electric motor as a power source, or a hybrid car having an internal combustion engine and an electric motor. The electric motor is driven by a battery such as a secondary battery, a hydrogen fuel cell, a metal fuel cell, or an alcohol fuel cell.

[0055] exist Figure 1 Also depicted is a moving body B1 moving in lane L2 and a moving body B2 moving on the roadside strip (or shoulder) of lane L1. Moving bodies B1 and B2 are examples of surrounding moving bodies B detected by vehicle VH. As moving body B1, four-wheeled vehicles, two-wheeled vehicles, and bicycles can be exemplified. Four-wheeled vehicles include large vehicles such as buses and small vehicles such as passenger cars. The moving direction of moving body B1 is opposite to the traveling direction of vehicle VH. As moving body B2, pedestrians and animals can be exemplified. The moving direction of moving body B2 is the same as the traveling direction of vehicle VH.

[0056] exist Figure 1 In the example shown, the state (position and speed) of the moving body B is detected by the vehicle VH. In the embodiment, the detected state and the behavior model MB of the moving body B are used to predict the future behavior of the moving body B. The behavior model MB is set in advance corresponding to various behavior patterns that the moving body B can take. Each behavior pattern can be called a candidate for the future behavior of the moving body B. The behavior pattern is set for each type of the moving body B. The behavior pattern is set based on the moving direction of the moving body B.

[0057] As the behavior model MB1 of the moving body B1, there can be exemplified a constant speed straight-moving model Mh, a deceleration model Mi, an acceleration model Mj, a right turn model Mk, a left turn model Ml, and a shaking model Mm. The behavior model MB1 also includes a model corresponding to a behavior pattern that determines the driving position of the moving body B1, such as moving at a position close to the boundary between lanes L1 and L2. The deceleration model Mi and the acceleration model Mj also include models that assume various accelerations, such as a slow model and a fast model. The behavior model MB2 of the moving body B2 is set with the same purpose as that of the behavior model MB1.

[0058] exist Figure 1 The track CP_M indicated by the dotted line in FIG schematically shows the future track of the mobile body B predicted using a certain behavior model MB. As already explained, the behavior model MB corresponds to various behavior patterns that the mobile body B can take. Therefore, there are actually multiple tracks CP_M. It should be noted that Figure 1 The trajectory CP_M depicted in FIG. 1 matches the future trajectory of the moving bodies B1 and B2 predicted using the constant speed straight travel model Mh, the deceleration model Mi, or the acceleration model Mj.

[0059] 1-2. Probability density distribution

[0060] The detection state of the moving object B obtained by the vehicle VH includes an error ε. The error ε includes at least one of a position error and a speed error. When the detection state includes the error ε, the future position of the moving object B is represented by a probability density distribution with the position on the track CP_M as the mode value. Figure 2 This is a diagram showing an example of a probability density distribution DRB of the future position of the reference point RB of the moving body B. It should be noted that the reference point RB is set at an arbitrary position of the moving body B. Consider the case where the time t = t0 is set as the current time. In this case, it can be said that the position of the reference point RB at the future time t = t1 (that is, the future position) exists in the interval of 1σ to 3σ (σ: standard deviation) of the probability density distribution DRB (t1).

[0061] In the embodiment, the error ε is applied to the prediction of the future behavior of the moving body B. In this prediction, the probability density distribution of the position of the reference point RB at future times t1, t2, t3, ... when the time t=t0 is set as the current time is calculated.

[0062] Figure 3 : is a diagram showing an example of a probability density distribution DRB1 of the future position of the reference point RB1 in the behavior pattern corresponding to the constant speed straight driving model Mh (that is, constant speed straight driving). Figure 3 In the upper layer (i) of , probability density distributions DRB1(t1), DRB1(t2), and DRB1(t3) of future times t=t1, t2, and t3 are depicted when time t=t0 is set as the current time. Figure 3 In the lower layer (ii), the probability density distributions DRB1(t2), DRB1(t3) and DRB1(t4) of the future moments t=t2, t3 and t4 when the moment t=t1 is set as the current moment are depicted. It should be noted that Figure 3 As shown in FIG. 1 , the interval of 1σ to 3σ of the probability density distribution DRB gradually expands with the passage of time. The reason for this is that the error ε accumulates with the passage of time.

[0063] Figure 4 1 is a diagram showing an example of the probability density distribution of the future position of the reference point RB1 in the behavior pattern (ie, right turn) corresponding to the right turn model Mk. Figure 4 The probability density distribution DRB1 at time t1 to t4 shown in the upper layer (i) and the lower layer (ii) is the same as that at Figure 3 The time t1 to t4 shown in FIG. 1 are the same time. The reason why the interval of 1σ to 3σ of the probability density distribution DRB expands with the passage of time is as described above. It should be noted that for the sake of convenience, Figure 3 and Figure 4In , a combination of a behavior model MB used for calculation of the probability density distribution DRB1 and a time at which the calculation is performed (that is, the current time) is used, so that the track CP_M is represented as "CP_Mh(t0)", "CP_Mk(t1)", etc.

[0064] 1-3. Collision determination

[0065] In the embodiment, the probability density distribution DRB is used to determine whether a collision between the vehicle VH and the moving body B will occur in the future (collision determination). Figure 3 and Figure 4 , reference points RVH(t0) to RVH(t4) at respective times of the reference point RVH of the vehicle VH are plotted.

[0066] In collision judgment, first, it is determined whether the intersection condition between the vehicle VH and the moving body B is satisfied. In the judgment of the intersection condition, first, the future time is aligned between the reference point RVH and the reference point RB1 with the current time as the reference. Then, at any future time, it is determined whether the reference point RVH exists in the interval of 1σ to 3σ of the probability density distribution DRB1 (hereinafter also referred to as the "determination interval"). The determination interval can also be limited to the interval of 1σ to 2σ. Then, when it is determined that the reference point RVH exists in the determination interval, it is determined that the intersection condition is satisfied. Figure 4 The upper layer (i) of meets the situation where the intersection condition is satisfied at the future time t3.

[0067] In the collision determination, it is also determined whether the conditions related to the remaining collision time TTC (Time To Collision: collision time) (hereinafter also referred to as "TTC conditions") are met. For example, when the interval from time t0 to time t3 is less than the time TTC, the TTC condition is met. When the intersection condition and the TTC condition are met, it is determined that a collision between the vehicle VH and the moving body B will occur in the future. It should be noted that in this case, the collision avoidance action of the vehicle VH (specifically, at least one of the actions of deceleration and steering) is started. Alternatively, a process is performed to warn the driver of an imminent collision with the moving body B.

[0068] As already explained, Figure 4 The upper layer (i) of meets the situation where the intersection condition is satisfied. Therefore, if the TTC condition is satisfied in addition to the intersection condition, a collision avoidance action is started or a warning process is performed. The problem here is that the actual behavior of the moving body B is not a right turn. Figure 4 The lower layer (ii) of shows that the moving body B moves straight at a constant speed from time t0 to time t1. That is, Figure 4The lower layer (ii) shows that the actual behavior of the moving object B from time t0 to time t1 is not a right turn.

[0069] However, in Figure 4 In the case of the upper layer (i) of FIG. 1 , when the TTC condition is satisfied between time t0 and time t1, a collision avoidance action or a warning process is started. Moreover, the collision avoidance action or the warning process is actually unnecessary. Performing an unnecessary collision avoidance action will destabilize the driving state of the vehicle VH. Performing an unnecessary warning process will cause discomfort to the driver.

[0070] 1-3-1. Similar level

[0071] Therefore, in the embodiment, in order to limit the behavior pattern for collision determination (determination of intersection conditions), "similarity level RLv" is applied to the probability density distribution DRB calculated using the behavior model MB. The similarity level RLv uses the reference probability density distribution SDRB as a comparison reference. The reference probability density distribution SDRB is a past probability density distribution DRB calculated using the behavior model MB corresponding to the current behavior of the mobile body B. For the comparison between the probability density distribution DRB and the reference probability density distribution SDRB, after aligning the future moments of these distributions DRB and SDRB, KLD (Kullback-Leiber divergence) is calculated. KLD is an indicator for measuring the difference between two different probability density distributions, and is suitable as an element of the similarity level RLv.

[0072] Figure 5 It will be Figure 3 and Figure 4 The tracks CP_Mk(t0), CP_Mh(t0), CP_Mk(t1), and CP_Mh(t1) shown in FIG are arranged so as to align the future time. As already explained, in Figure 3 and Figure 4 In the example, the actual behavior pattern from time t0 to t1 is straight forward at a constant speed. That is, if time t=t1 is set as the current time, then at least the behavior of the moving body B from time t0 to t1 can be considered as the current behavior of the moving body B. Therefore, Figure 5 The current behavior in the example shown is “straight ahead at a constant speed.” Furthermore, the reference probability density distribution SDRB is the probability density distributions DRB(t2) and DRB(t3) predicted at time t0 using the straight ahead at a constant speed model Mh.

[0073] It should be noted that the reason why the probability density distribution DRB(t1) is not included in the baseline probability density distribution SDRB is because the time t=t1 is set as the current time. If the alignment is performed under the condition of time t=t2, the difference between the baseline probability density distribution SDRB(t2) and the probability density distribution DRB(t2) calculated using the constant speed straight-line model Mh at time t1 is only 1σ to 3σ. The difference is also the same under the condition of time t=t3. Therefore, the KLD for the probability density distributions DRB(t2) and DRB(t3) are both high scores, and the similarity level RLv using them as elements is also high.

[0074] On the other hand, the KLD between the reference probability density distribution SDRB (t2) and the probability density distribution DRB (t2) calculated at time t1 using the right turn model Mk becomes a low score. The reason is that in addition to the size of 1σ to 3σ, the position in the width direction of the lane L2 is also different. In addition, the KLD between the reference probability density distribution SDRB (t3) and the probability density distribution DRB (t3) calculated at time t1 using the right turn model Mk becomes an even lower score. As a result, the similarity level RLv of these elements also becomes a low level.

[0075] Here, use Figure 6 and Figure 7 , an explanation is given of the similarity level RLv of the probability density distribution DRB calculated using the acceleration model Mj. Figure 6 The probability density distribution DRB1 at time t1 to t4 shown in the upper layer (i) and the lower layer (ii) is the same as that at Figure 3 and Figure 4 The times t1 to t4 shown in FIG. 1 are the same times. Figure 7 The trajectories CP_Mh(t0) and CP_Mh(t1) shown are similar to those in Figure 5 The trajectories CP_Mh(t0) and CP_Mh(t1) shown in φ are the same. The reason why the interval of 1σ to 3σ of the probability density distribution DRB expands with the passage of time is as described above.

[0076] like Figure 7 As shown, the KLD between the reference probability density distribution SDRB (t2) and the probability density distribution DRB (t2) calculated using the acceleration model Mj at time t1 becomes a relatively high score. The reason is that the difference between these distributions is only the size of 1σ~3σ and the slight position deviation of the moving direction of the moving body B1. In addition, the KLD between the reference probability density distribution SDRB (t3) and the probability density distribution DRB (t3) calculated using the acceleration model Mj at time t1 also becomes a relatively high score. As a result, the similarity level RLv using them as elements also becomes a relatively high level.

[0077] 1-3-2. Evaluation of similar levels

[0078] In the embodiment, the similarity level RLv is calculated for all behavior patterns at the current time t. Then, once the similarity level RLv for all behavior patterns is calculated, a behavior pattern is selected based on the similarity level RLv, which is the behavior pattern for collision determination (determination of intersection condition). Figure 8 is a diagram showing an example of a similar level RLv. Figure 8 As shown, an allowable range AR is set in the similarity level RLv. The allowable range AR is set so that the behavior pattern with a low similarity level RLv is not set as the object of collision determination (determination of intersection condition). That is, the behavior pattern with a similarity level RLv lower than the allowable range AR is not set as the object of collision determination (determination of intersection condition).

[0079] In this way, if the behavior pattern below the allowable range AR is excluded, the implementation of the determination of the intersection condition of the behavior pattern can be canceled. Therefore, the start of the collision avoidance action that is not originally necessary can be suppressed. Therefore, when the prediction of the behavior pattern using various behavior models MB is performed, the instability of the driving state of the vehicle VH can be suppressed. In addition, the execution of the warning processing that is not originally necessary can also be suppressed. Therefore, the execution of the warning processing can also be suppressed from causing discomfort to the driver. Moreover, the load of the determination processing of the intersection condition can also be reduced, thereby suppressing the reduction in the responsiveness of the collision determination.

[0080] A high similarity threshold TH and a low similarity threshold TL are also set in the allowable range AR. These thresholds are set to determine whether to change (enlarge or reduce) the error ε to be used in the next calculation timing of the probability density distribution DRB. "The next calculation timing" refers to, for example, the calculation timing corresponding to the future moment t=t2 when the moment t=t1 is set to the current moment. In an embodiment, for a behavior pattern in which the similarity level RLv is higher than the high similarity threshold TH, the error ε used in the next calculation timing of the probability density distribution DRB using the behavior model MB corresponding to the behavior pattern is reduced. Conversely, for a behavior pattern in which the similarity level RLv is lower than the low similarity threshold TL, the error ε is increased.

[0081] If the error ε is reduced, 1σ to 3σ of the probability density distribution DRB calculated at the next calculation timing becomes smaller. In the case where the behavior pattern of the mobile body B does not change at each moment, it is expected that the similarity level RLv of the behavior pattern will be higher than the high similarity threshold TH. However, since the interval of 1σ to 3σ of the probability density distribution DRB expands with the passage of time, even in such a case, the intersection condition may be satisfied at a position far enough away from the mobile body B. In this regard, if 1σ to 3σ of the probability density distribution DRB becomes smaller, such an undesirable situation can be suppressed. This helps to improve the accuracy of collision judgment.

[0082] If the error ε is enlarged, 1σ to 3σ of the probability density distribution DRB calculated at the next calculation timing becomes larger. If 1σ to 3σ of the probability density distribution DRB becomes larger, the similarity level RLv is expected to decrease. The behavior pattern in which the similarity level RLv is lower than the low similarity threshold TL can sometimes be classified as a behavior pattern that is lower than the allowable range AR. At this point, if the similarity level RLv decreases, the possibility of being lower than the allowable range AR increases. Thus, the implementation of the judgment of the cross condition regarding the behavior pattern in which the similarity level RLv is lower than the low similarity threshold TL can be canceled.

[0083] Hereinafter, the behavior prediction device according to the embodiment will be described in detail.

[0084] 2. Behavior prediction device

[0085] 2-1. Example of configuration

[0086] Fig. 9 1 is a diagram showing an example of the configuration of the behavior prediction device 10 according to the embodiment. Fig. 9 As shown, the behavior prediction device 10 includes an external information acquisition device 1 , an internal information acquisition device 2 , a behavior model DB (database) 3 , an HMI (Human Machine Interface) unit 4 , a travel device 5 , and a control device 6 .

[0087] The external information acquisition device 1 detects the conditions around the vehicle VH. As the external information acquisition device 1, a radar sensor and a camera can be exemplified. The radar sensor uses radio waves (for example, millimeter waves) or light to detect objects around the vehicle VH. The objects include fixed objects and moving objects. As fixed objects, guardrails and buildings can be exemplified. The moving objects include moving bodies B such as four-wheeled vehicles, two-wheeled vehicles, and pedestrians. The camera captures the external conditions of the vehicle VH. The camera is, for example, mounted on the back of the front windshield. The external information acquisition device 1 sends the detected external information to the control device 6.

[0088] The internal information acquisition device 2 detects the driving state of the vehicle VH. As the internal information acquisition device 2, a vehicle speed sensor, an acceleration sensor, and a yaw rate sensor can be exemplified. The vehicle speed sensor detects the driving speed of the vehicle VH. The acceleration sensor detects the acceleration of the vehicle VH. The yaw rate sensor detects the yaw rate of the center of gravity of the vehicle VH around the vertical axis. The internal information acquisition device 2 sends the detected internal information to the control device 6.

[0089] The behavior model DB 3 stores the behavior model MB. As described above, the behavior model MB is set in advance in correspondence with various behavior patterns that the moving body B can take.

[0090] The HMI unit 4 is an interface for providing information to the driver (or remote operator) of the vehicle VH and receiving information from the driver. The HMI unit 4 includes, for example, an input device, a display device, a speaker, and a microphone. As the input device, a touch panel, a keyboard, a switch, and a button can be exemplified. The information provided to the driver includes the driving state of the vehicle VH, reminders, and warnings. The display device and the speaker are used to provide information to the driver. The input device and the microphone are used to receive information from the driver.

[0091] The travel device 5 includes various actuators. The various actuators include a drive actuator, a brake actuator, and a steering actuator. The drive actuator drives the vehicle VH. The brake actuator applies a braking force to the vehicle VH. The steering actuator steers the tires of the vehicle VH.

[0092] The control device 6 is a microcomputer having at least a processor 7 and a memory 8. The control device 6 receives various information (for example, external information and internal information) via an input / output interface. The information received by the control device 6 is stored in the memory 8. As the memory 8, a volatile memory and a non-volatile memory can be exemplified. The information stored in the memory 8 includes position data x, speed data v, position error data εx, and speed error data εv of the mobile body B. The processor 7 performs various processes based on the data and programs stored in the memory 8. The various processes include a process for predicting the future behavior of the mobile body B (hereinafter also referred to as "future behavior prediction processing"). The functional structure of the control device 6 associated with the future behavior prediction processing is described below.

[0093] 2-2. Example of functional configuration of control device

[0094] Fig.10 1 is a block diagram showing an example of the functional configuration of the control device 6 associated with the future behavior prediction process. Fig.10As shown, the control device 6 includes a model extraction unit 61, a distribution calculation unit 62, a distribution history storage unit 63, a reference distribution extraction unit 64, a similarity level calculation unit 65, a similarity level evaluation unit 66, a distribution recalculation unit 67, a collision determination unit 68, and an error data change unit 69. It should be noted that these functional blocks are implemented by the processor 7 executing a program stored in the memory 8.

[0095] The model extraction unit 61 determines the type of the moving body B based on the external information. The type of the moving body B is determined, for example, by analyzing the captured image from the camera. The type of the moving body B can also be determined based on the position data x and the speed data v of the moving body B. The model extraction unit 61 extracts the behavior model MB corresponding to the determined type from the behavior model DB3. The extracted behavior model MB is all the models set according to the type of the moving body B. The model extraction unit 61 sends the extracted behavior model MB to the distribution calculation unit 62.

[0096] The model extraction unit 61 may also determine the object ID of the moving body B. The object ID is determined by tracking processing based on the position data x and velocity data v of the moving body B. In the tracking process, when an object different from the object currently being detected is detected, a new object ID is assigned to the object. When an object that is considered to be the same as the object currently being detected is detected as a result of tracking, the object is given an object ID that is the same as the already assigned object ID. When the new object ID is assigned to the moving body B, the model extraction unit 61 extracts the behavior model MB corresponding to the type of the moving body B. Then, while the moving body B continues to be detected, the initially extracted behavior model MB is sent to the distribution calculation unit 62.

[0097] The model extraction unit 61 may also cancel the transmission of a part of the behavior model MB based on external information before transmitting the extracted behavior model MB to the distribution calculation unit 62. For example, when it is detected from the road surface markings of the lane L2 that a U-turn in the lane L2 is prohibited, the model extraction unit 61 may cancel the transmission of the behavior model MB corresponding to the U-turn. In addition, for example, when it is recognized that the red light of the traffic light in front of the moving body B is on, the model extraction unit 61 may cancel the transmission of the behavior model MB corresponding to constant speed straight driving and acceleration. In this way, the transmission of a part of the behavior model MB may be canceled according to traffic regulations, traffic conditions, etc.

[0098] The distribution calculation unit 62 performs distribution calculation processing. In the distribution calculation processing, the probability density distribution DRB is calculated using the behavior model MB received from the model extraction unit 61, the position data x and speed data v of the mobile body B, and their error data εx and εv. The moment when the position data x and the speed data v are obtained is set as the current moment, and the probability density distribution DRB is calculated in association with the future moment that has advanced by a specified time scale (for example, 1 second) from the current moment. The number of probability density distributions DRB calculated in one processing is not particularly limited. However, the more the number of probability density distributions DRB increases, the greater the calculation load of the processor 7. Therefore, it is ideal that the number of calculations for each processing is set in advance according to the processing power of the processor 7.

[0099] In principle, the error data εx and εv use pre-set values. However, when correction data εx* and εv* related to a specific behavior model MB* are received from the error data changing unit 69, these correction data are used as an exception. Specifically, when the behavior model MB received from the model extracting unit 61 includes a specific behavior model MB*, the probability density distribution DRB is calculated using the behavior model MB*, the position data x and the velocity data v of the moving body B, and the correction data εx* and εv*.

[0100] The distribution calculation section 62 sends the probability density distribution DRB obtained through the distribution calculation process to the similarity level calculation section 65 .

[0101] The distribution history storage unit 63 stores the probability density distribution DRB calculated by the distribution calculation unit 62 in the memory 8. The probability density distribution DRB is stored in combination with the object ID, the behavior model MB, and the time (hh:mm:ss:fff) when the probability density distribution DRB is calculated. This data set constitutes the history of the probability density distribution DRB.

[0102] The reference distribution extraction unit 64 extracts the reference probability density distribution SDRB. The distribution history storage unit 63 stores the probability density distribution DRB calculated at the past time points traced back from the current time point at a predetermined time scale. The reference distribution extraction unit 64 estimates the current behavior of the mobile body B based on the external information from the last calculation timing of the probability density distribution DRB to the current calculation timing. "This calculation timing" refers to the calculation timing corresponding to the current time point. "Last calculation timing" refers to the calculation timing corresponding to the past time point t=t0 when, for example, the time point t=t1 is set as the current time point.

[0103] The reference distribution extraction unit 64 uses the estimated current behavior of the moving body B as a search key to extract the reference probability density distribution SDRB from the history stored in the distribution history storage unit 63. It should be noted that the search key includes the object ID of the moving body B and the time t corresponding to the last calculation timing of the probability density distribution DRB. The reference distribution extraction unit 64 sends the extracted reference probability density distribution SDRB to the similarity level calculation unit 65.

[0104] The similarity level calculation unit 65 performs similarity level calculation processing. In the similarity level calculation processing, the similarity level RLv between the probability density distribution DRB received from the distribution calculation unit 62 and the reference probability density distribution SDRB received from the reference distribution extraction unit 64 is calculated. When calculating the similarity level RLv, after aligning the future moments of these distributions DRB and SDRB, KLD is calculated. KLD is calculated for each future moment t.

[0105] For example, if time t=t0 is set as the current time, the similarity level RLv is calculated by the following equation (1) using KLD at future times t=t1, t2, ..., tn.

[0106] RLv=w1·KLD(t1)+w2·KLD(t2)+……+wn·KLD(tn)……(1)

[0107] In equation (1), w1, w2, and wn are weight coefficients that satisfy the following equation (2).

[0108] w1+w2+……+wn=1……(2)

[0109] The similarity level calculation section 65 sends the calculated similarity level RLv to the similarity level evaluation section 66 .

[0110] The similarity level evaluation unit 66 performs similarity level evaluation processing. Fig.11 This similarity level evaluation process will be described. Fig.11 7 is a flowchart showing the flow of the similarity level evaluation process performed by the processor 7. Each time the similarity level RLv of all the behavior patterns at the current moment is calculated, the process is repeated. Fig.11 It is to be noted that, of course, when transmission of a part of the behavior model MB is canceled in accordance with traffic regulations or the like, the behavior pattern corresponding to the canceled behavior model MB is not included in “all behavior patterns at the current time”.

[0111] exist Fig.11In the routine shown, first, it is determined whether a behavior pattern having a similar level RLv (hereinafter, also referred to as "out-of-range level RNA") lower than the permissible range AR is included (step S11). Figure 8 If it is determined that a behavior pattern having an out-of-range level RNA is included, the probability density distribution DRB calculated using the behavior model MB corresponding to the behavior pattern is discarded (step S12).

[0112] If the result of the determination in step S11 is negative, it is determined whether a behavior pattern having a similarity level RLv satisfying RLv≥TH (hereinafter, also referred to as a "high similarity level RLH") is included (step S13). Figure 8 As described in . In the case where it is determined that a behavior pattern having a high similarity level RLH is included, the probability density distribution DRB calculated using the behavior model MB corresponding to the behavior pattern (hereinafter also referred to as "high similarity model MBH") is sent to the distribution recalculation unit 67 (step S14). In the process of step S14, the information of the high similarity model MBH is also sent to the error data change unit 69.

[0113] If the result of the determination in step S13 is negative, it is determined whether a behavior pattern having a similarity level RLv (hereinafter, also referred to as a "low similarity level RLL") satisfying RLv≤TL is included (step S15). Figure 8 As described in . When it is determined that a behavior pattern having a low similarity level RLL is included, the probability density distribution DRB calculated using the behavior model MB corresponding to the behavior pattern (hereinafter also referred to as "low similarity model MBL") is sent to the collision determination unit 68 (step S16). In the process of step S16, the information of the low similarity model MBL is also sent to the error data change unit 69.

[0114] When the determination result of step S15 is negative, it is assumed that a behavior pattern having a similarity level RLv (hereinafter also referred to as "medium similarity level RLM") that satisfies TL<RLv<TH remains. Therefore, in this case, the probability density distribution DRB calculated using the behavior model MB corresponding to the remaining behavior pattern is sent to the collision determination unit 68 (step S17). It should be noted that it is also assumed that there is no behavior pattern having a medium similarity level RLM left. Therefore, it is also possible to determine whether a behavior pattern having a medium similarity level RLM is included before processing step S17. Then, the processing of step S17 can be performed when the determination result is positive.

[0115] Back to Fig.10, the description of the configuration example of the control device 6 is continued. The distribution recalculation unit 67 performs distribution recalculation processing. Fig.12 This distribution recalculation process will be described. Fig.12 is a diagram showing an example of distribution recalculation processing. The object of the distribution recalculation processing is the probability density distribution DRB calculated using the high similarity model MBH. Fig.12 In the description, the case where the constant speed straight-moving model Mh conforms to the high similarity model MBH is taken as an example for explanation.

[0116] exist Fig.12 In the example shown, Figure 3 The tracks CP_Mh(t0) and CP_Mh(t1) shown in are arranged so as to align the future time points. Figure 3 Different, in Fig.12 The probability density distribution DRB(t4) predicted at time t4 is added. In the distribution recalculation process, the probability density distribution DRB1 of the future time t2, t3 and t4 when the time t=t1 is set as the current time is superimposed. The objects to be superimposed are the probability density distribution DRB1 calculated using the high similarity model MBH at time t=t1 and the time before the current time (in Fig.12 The example shown is the probability density distribution DRB1 at time t=t0. The superposition is performed for each future time. It should be noted that the probability density distribution DRB1(t1) at time t=t1, which is the current time, is not the object of superposition.

[0117] Fig.12 The difference between the probability density distribution DRB1(t2)~DRB1(t4) shown in the upper layer and DRB1(t2)~DRB1(t4) shown in the lower layer is only 1σ~3σ. Therefore, if the probability density distribution DRB1 of the upper layer and the lower layer is superimposed for each future moment, the 1σ~3σ of the probability density distribution DRB1* after superposition is smaller than the 1σ~3σ of the probability density distribution DRB1 before superposition. In this way, if superposition is performed, the degree of deviation of the future position of the moving body B becomes smaller.

[0118] The distribution calculation unit 62 sends the superimposed probability density distribution DRB* obtained by the distribution recalculation process to the collision determination unit 68 .

[0119] The collision determination unit 68 performs a collision determination process. In the collision determination process, the probability density distribution DRB received from the similarity level evaluation unit 66 or the distribution recalculation unit 67 is used to determine whether the intersection condition and the TTC condition are satisfied. In the determination process of the intersection condition, the future time is aligned between the reference point RVH and the reference point RB with the current time as a reference. Then, at any future time, it is determined whether the reference point RVH exists in the determination interval. Then, when it is determined that the reference point RVH exists in the determination interval, it is determined that the intersection condition is satisfied.

[0120] In the TTC condition determination process, it is determined whether the interval from the current time to the future time determined to satisfy the intersection condition is less than or equal to the time TTC. If it is determined that the interval is less than or equal to the time TTC, it is determined that the TTC condition is satisfied.

[0121] In the collision determination process, when it is determined that the intersection condition and the TTC condition are satisfied, it is determined that a collision between the vehicle VH and the moving body B will occur in the future. When it is determined that a collision will occur, the collision determination unit 68 outputs a control command for starting a collision avoidance action of the vehicle VH to the travel device 5. When the control command is output, at least one of the brake actuator and the steering actuator is controlled by the travel device 5. The collision determination unit 68 also outputs a warning signal indicating that a collision with the moving body B is imminent to the HMI unit 4.

[0122] The error data changing unit 69 performs error data changing processing. In the error data changing processing, the information of the high similarity model MBH and the low similarity model MBL received from the similarity level evaluation unit 66 is referred to. When the information of the high similarity model MBH is received, the error ε used in the next calculation timing of the probability density distribution DRB using the high similarity model MBH is changed. The error ε* after the change represents a value smaller than the error ε before the change. That is, the error ε is reduced. The error ε is also changed when the information of the low similarity model MBL is received. The error ε* after the change represents a value larger than the error ε before the change. That is, the error ε is increased.

[0123] 3. Effect

[0124] According to the behavior prediction device of the embodiment described above, distribution calculation processing, reference distribution calculation processing, similarity level calculation processing and similarity level evaluation processing are performed. In particular, according to the similarity level evaluation processing, the behavior pattern with the out-of-range level RNA can be excluded from the object of the collision determination processing. Therefore, it is possible to suppress the start of collision avoidance actions that are not originally necessary. Therefore, in the case of prediction of behavior patterns using various behavior models MB, the instability of the driving state of the vehicle VH can be suppressed. In addition, it is also possible to suppress the performance of warning processing that is not originally necessary. Therefore, it is also possible to suppress the execution of the warning processing from causing discomfort to the driver. Moreover, it is also possible to reduce the load of the determination processing of the intersection condition, thereby suppressing the reduction in the responsiveness of the collision determination.

[0125] Furthermore, according to the behavior prediction device of the embodiment, in addition to the similarity level evaluation process, an error data change process is also performed. According to the error data change process, the error ε regarding the behavior pattern with a low similarity level RLL is increased. If the error ε is increased, 1σ to 3σ of the probability density distribution DRB calculated using the low similarity model MBL at the next calculation timing becomes larger. Therefore, the similarity level RLv calculated based on the probability density distribution DRB can be reduced so as to be lower than the allowable range AR. In this way, the behavior pattern that is evaluated as having a low similarity level RLL at the current calculation timing is evaluated as a behavior pattern with an out-of-range level RNA at the next calculation timing. Thus, the behavior pattern with a relatively low similarity level RLv can be excluded from the object of the collision determination process.

[0126] In addition, according to the error data change processing, the error ε regarding the behavior pattern with a high similarity level RLH is reduced. If the error ε is reduced, 1σ~3σ of the probability density distribution DRB calculated using the high similarity model MBH at the next calculation timing becomes smaller. In the case where the behavior pattern of the mobile body B does not change at each moment, it is expected that the similarity level RLv of the behavior pattern will become a high similarity level RLH. However, since the interval of 1σ~3σ of the probability density distribution DRB expands with the passage of time, even in such a case, the intersection condition may be satisfied at a position far enough away from the mobile body B. In this regard, if 1σ~3σ of the probability density distribution DRB becomes smaller, such an undesirable situation can be suppressed. This helps to improve the accuracy of collision judgment.

[0127] Furthermore, according to the behavior prediction device of the embodiment, in addition to the similarity level evaluation process, a distribution recalculation process is also performed. According to the distribution recalculation process, the probability density distribution DRB calculated using the high similarity model MBH at the current moment is superimposed with the probability density distribution DRB calculated using the high similarity model MBH at the moment before the current moment. If superposition is performed, 1σ to 3σ of the probability density distribution DRB1* after superposition is smaller than 1σ to 3σ of the probability density distribution DRB1 before superposition. Therefore, according to the distribution recalculation process, the same effect as that obtained when the error ε is reduced by the error data change process can be expected.

Claims

1. A behavior prediction device for predicting future behavior of a surrounding moving object detected by a vehicle, the behavior prediction device comprising: a database storing a plurality of behavior models set for each behavior pattern corresponding to the type of the moving body; a memory storing position data and speed data of the moving body and error data representing an error of the position data and speed data; as well as processor, The processor is configured to: distribution calculation processing of calculating a probability density distribution associated with the future position of the mobile body for each of the behavior patterns using the position data and speed data, the error data, and the plurality of behavior models; a similarity level calculation process of calculating, for each of the behavior patterns, a similarity level between the probability density distribution and a reference probability density distribution, wherein the reference probability density distribution represents the probability density distribution calculated using a behavior model corresponding to the current behavior of the mobile body at a processing timing prior to the current processing timing of the distribution calculation process; A similarity level evaluation process is performed to evaluate the similarity level; an error data changing process of changing the error data based on the similarity level; and collision determination processing, using the probability density distribution to determine a collision between the vehicle and the moving object, When it is determined that the vehicle and the moving object are likely to collide, at least one of a brake actuator and a steering actuator is controlled by a travel device so that the vehicle performs a collision avoidance action. The similar levels include levels outside the range that are lower than the similar permissible range, The similarity level includes a low similarity level that exists within the similarity tolerance range and is lower than a low similarity threshold value set within the similarity tolerance range, In the similarity level evaluation process, the processor excludes the behavior pattern having the out-of-range level from being a target of the collision determination process, In the error data changing process, the processor enlarges the error data to be used at the next processing timing of the distribution calculation process using the low similarity model compared to the error data used at the current processing timing of the distribution calculation process, The low similarity model represents a behavior model corresponding to the behavior pattern having the low similarity level.

2. The behavior prediction device according to claim 1, characterized in that: The similarity level includes a high similarity level that exists within the similarity allowable range and is higher than a high similarity threshold value set within the similarity allowable range, the processor is configured to further perform a distribution recalculation process of recalculating the probability density distribution using a high similarity model representing a behavior model corresponding to the behavior pattern having the high similarity level, In the distribution recalculation processing, the processor superimposes the probability density distribution calculated using the high similarity level at the current processing timing and the probability density distribution calculated using the high similarity level at the processing timing before the current processing timing, In the collision determination process, the processor performs determination of the behavior pattern having the high similarity level based on the superimposed probability density distribution obtained by the distribution recalculation process.

3. A behavior prediction device for predicting future behavior of a surrounding moving object detected by a vehicle, the behavior prediction device comprising: a database storing a plurality of behavior models set for each behavior pattern corresponding to the type of the moving body; a memory storing position data and speed data of the moving body and error data representing an error of the position data and speed data; as well as processor, The processor is configured to: distribution calculation processing of calculating a probability density distribution associated with the future position of the mobile body for each of the behavior patterns using the position data and speed data, the error data, and the plurality of behavior models; a similarity level calculation process of calculating, for each of the behavior patterns, a similarity level between the probability density distribution and a reference probability density distribution, wherein the reference probability density distribution represents the probability density distribution calculated using a behavior model corresponding to the current behavior of the mobile body at a processing timing prior to the current processing timing of the distribution calculation process; A similarity level evaluation process, evaluating the similarity level; an error data changing process of changing the error data based on the similarity level; and collision determination processing, using the probability density distribution to determine a collision between the vehicle and the moving object, When it is determined that the vehicle and the moving object are likely to collide, at least one of a brake actuator and a steering actuator is controlled by a travel device so that the vehicle performs a collision avoidance action. The similar levels include levels outside the range that are lower than the similar permissible range, The similarity level includes a high similarity level that exists within the similarity allowable range and is higher than a high similarity threshold value set within the similarity allowable range, In the similarity level evaluation process, the processor excludes the behavior pattern having the out-of-range level from being a target of the collision determination process, In the error data changing process, the processor reduces the error data to be used at the next processing timing of the distribution calculation process using the high similarity model compared to the error data used at the current processing timing of the distribution calculation process, The high similarity model represents a behavior model corresponding to the behavior pattern having the high similarity level.

4. A behavior prediction device for predicting future behavior of a surrounding moving object detected by a vehicle, the behavior prediction device comprising: a database storing a plurality of behavior models set for each behavior pattern corresponding to the type of the moving body; a memory storing position data and speed data of the moving body and error data representing an error of the position data and speed data; as well as processor, The processor is configured to: distribution calculation processing of calculating a probability density distribution associated with the future position of the mobile body for each of the behavior patterns using the position data and speed data, the error data, and the plurality of behavior models; a similarity level calculation process of calculating, for each of the behavior patterns, a similarity level between the probability density distribution and a reference probability density distribution, wherein the reference probability density distribution represents the probability density distribution calculated using a behavior model corresponding to the current behavior of the mobile body at a processing timing prior to the current processing timing of the distribution calculation process; a similarity level evaluation process of evaluating the similarity level; and collision determination processing, using the probability density distribution to determine a collision between the vehicle and the moving object, When it is determined that the vehicle and the moving object are likely to collide, at least one of a brake actuator and a steering actuator is controlled by a travel device so that the vehicle performs a collision avoidance action. The similar levels include levels outside the range that are lower than the similar permissible range, In the similarity level evaluation process, the processor excludes the behavior pattern having the out-of-range level from being a target of the collision determination process, The similarity level includes a high similarity level that exists within the similarity allowable range and is higher than a high similarity threshold value set within the similarity allowable range, the processor is configured to further perform a distribution recalculation process of recalculating the probability density distribution using a high similarity model representing a behavior model corresponding to the behavior pattern having the high similarity level, In the distribution recalculation processing, the processor superimposes the probability density distribution calculated using the high similarity level at the current processing timing and the probability density distribution calculated using the high similarity level at the processing timing before the current processing timing, In the collision determination process, the processor performs determination of the behavior pattern having the high similarity level based on the superimposed probability density distribution obtained by the distribution recalculation process.

Images