Behavior prediction device, vehicle, and behavior prediction method

Abstract

To provide a behavior prediction device and a behavior prediction method for reducing a behavior prediction amount, which can be universally applied irrespective of the presence of a change factor of an object behavior.SOLUTION: This behavior prediction device comprises: an object prediction unit that detects an object to be predicted on the basis of environment information from an environment sensor; and a behavior prediction unit that predicts a behavior of the object. The behavior prediction unit includes a behavior pattern identification part that identifies behavior patterns which the object can take on the basis of at least one of a state of the object and a peripheral environment of the object, an output condition setting part that sets a prediction output condition for each of the identified behavior patterns, and a prediction processing part that outputs a prediction result for each of the behavior patterns on the basis of the set output condition.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to a behavior prediction device, a vehicle, and a behavior prediction method for predicting the behavior of others in order to realize a preventive safety function and an autonomous driving of a vehicle.

Background Art

[0002] In order to realize a preventive safety function and an autonomous driving of a vehicle, it is effective to predict the future behavior of surrounding objects detected by in-vehicle sensors. When predicting the behavior of surrounding objects, it may be necessary to consider the interaction between objects or predict a plurality of possible behaviors. While these behavior prediction methods can achieve high-precision prediction, they require complex calculations and have a high calculation load, which is a problem.

[0003] As a measure for reducing the calculation load during the behavior prediction of surrounding objects, the behavior prediction method of Patent Document 1 is known. For example, in the abstract of the same document, it is described as a problem that "behavior prediction is performed with high accuracy and the calculation load is reduced", and as a solution, "based on an object around the host vehicle or the driving environment around the host vehicle, it is determined whether there is a change factor leading to a change in the behavior of the prediction object. Based on the determination result of whether there is a change factor and the information obtained from the objects around the host vehicle, the behavior of the prediction object is predicted. When it is determined that there is no change factor, a short-term prediction is performed to predict the behavior of the prediction object based only on the information obtained from the prediction object. When it is determined that there is a change factor, a long-term prediction is performed to predict the behavior of the prediction object based on the information obtained from the prediction object and the information obtained from the surrounding objects around the prediction object."

[0004] As described above, Patent Document 1 discloses a method for reducing the calculation load in an environment where there is no change factor by determining the presence or absence of a change factor in the behavior of an object (such as the presence of other intersecting objects), performing a long-term prediction with a high calculation load when there is a change factor, and performing a short-term prediction with a low calculation load when there is no change factor. Thereby, compared with a behavior prediction method that always performs a prediction with a high calculation load, the scene for performing a high-precision behavior prediction is limited, and the overall calculation load can be reduced.

Prior Art Documents

Patent Documents

[0005]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0006] However, in the long-term prediction in Patent Document 1, in paragraph 0081 of the document, "In calculating the route candidates, the prediction processing unit 16 sets the lane center lines La1 and Lb2 at the center of the lane, respectively. In the case of the straight-ahead route candidate Rb2, the prediction processing unit 16 calculates the route candidate Rb2 so as to be parallel to the lane center line La2 based on the traveling position of the current other vehicle Vb1. On the other hand, in the case of the lane change route candidate Rb1, the prediction processing unit 16 calculates the traveling distance that the other vehicle Vb1 travels until the lane change is completed based on the time required until the lane change is completed and the traveling speed of the current other vehicle Vb1. The time required for the lane change is a value set in advance from, for example, the driving tendency of a general driver, and is, for example, 3 seconds. The prediction processing unit 16 sets the position when the lane change is completed on the lane center line La1 based on the calculated traveling distance. Then, the prediction processing unit 16 connects the current traveling position and the position when the lane change is completed with an arbitrary curve, for example, a clothoid curve, and sets this as the route candidate Rb1." As described in this way, a large number of route candidates such as the route candidates Rb1 to Rb4 and Rc1 in FIG. 6 of the same document are predicted by a method that requires a considerable computational load. Therefore, the effect of reducing the computational load by the technology of Patent Document 1 can only be obtained in an environment where short-term prediction can be performed without changing factors.

[0007] Therefore, in view of such problems, an object of the present invention is to provide a behavior prediction device and a behavior prediction method for reducing the amount of behavior prediction calculation that can be generally applied regardless of the presence or absence of factors that change the behavior of an object.

Means for Solving the Problems

[0008] An action prediction device includes an object detection unit that detects a prediction target object based on external information from an external sensor, and an action prediction unit that predicts the action of the prediction target object. The action prediction unit includes an action pattern specifying unit that specifies an action pattern that the prediction target object can execute based on at least one of the state of the prediction target object or the surrounding environment of the prediction target object, an output condition setting unit that sets an output condition for prediction for each of the specified action patterns, and a prediction processing unit that outputs a prediction result for each of the action patterns based on the set output condition.

Advantages of the Invention

[0009] According to the present invention, by dynamically changing the output conditions of prediction according to the action patterns that an object can take, it is possible to reduce the amount of calculation.

Brief Description of the Drawings

[0010]

Figure 1

Figure 2

Figure 3A

Figure 3B

Figure 3C

Figure 4

Figure 5

Figure 6

Figure 7

Figure 8

Embodiment for Carrying out the Invention

[0011] Hereinafter, the behavior prediction device 1 according to an embodiment of the present invention will be described in detail.

[0012] <Configuration of Behavior Prediction Device 1> FIG. 1 is a functional block diagram showing the configuration of the behavior prediction device 1 of this embodiment. As shown in FIG. 1, the behavior prediction device 1 has an object detection unit 10, a peripheral information management unit 11, and a behavior prediction unit 12. Further, the behavior prediction unit 12 has a behavior pattern specifying unit 12a, an output condition setting unit 12b, and a prediction processing unit 12c.

[0013] The behavior prediction device 1 may be applied to a vehicle having a preventive safety function or an automatic driving function, or may be applied to a device other than a vehicle such as a surveillance camera. Hereinafter, it is assumed that the behavior prediction device 1 is mounted on a vehicle. In the following description, the vehicle on which the behavior prediction device 1 is mounted may be referred to as the host vehicle.

[0014] FIG. 8 shows a hardware configuration diagram of a vehicle 9 equipped with the behavior prediction device 1 of this embodiment. The vehicle 9 includes a behavior prediction device 1, a vehicle control device 2, an external sensor 3, a display 4, a communication device 5, and a GPS receiver 6, and is mutually connected by an in-vehicle network 7.

[0015] The behavior prediction device 1 includes a CPU 1A which is a central processing unit, a ROM 1B which is a read-only storage device, a RAM 1C which is a volatile memory such as a DRAM, and a flash memory 1D which is a non-volatile storage unit. The behavior prediction device 1 realizes various functional units by the CPU 1A expanding and executing the program stored in the ROM 1B in the RAM 1C. The vehicle control device 2 is an electronic control device equipped with a CPU and a storage device, similar to the behavior prediction device 1. By the CPU executing the control program stored in the storage device, it controls various functions such as the steering system, drive system, and braking system, for example. The vehicle control device 2 can control the steering system, drive system, braking system, etc. based on the prediction result of the behavior prediction device 1. When using the prediction result for controlling each system of the host vehicle, by planning the trajectory of the host vehicle so as not to interfere with the predicted trajectories of other vehicles, a safe autonomous driving with less risk of collision can be realized.

[0016] The external sensor 3 can use, for example, LiDAR, a camera, a millimeter-wave radar, etc. The detection result by the external sensor 3 is input to the object detection unit 10 of the behavior prediction unit 1.

[0017] The display 4 is, for example, a display in the meter panel or a head-up display, etc., but is not limited to these forms. The prediction result of the behavior prediction device 1 can be notified to the driver via the display 4. By notifying the driver of a vehicle predicted to have a risk of contact with the host vehicle, the driver can drive while paying particular attention to the state of the notified vehicle.

[0018] The communication device 5 is a communication module that performs wireless communication with the outside of the vehicle 9. The communication device 5 is configured to be able to communicate with a server and, for example, acquires map information around the host vehicle.

[0019] The GPS receiver 6 is a receiver for GPS (Global Positioning System) that measures the position of the vehicle based on radio waves from satellites.

[0020] Hereinafter, the details of each part of the behavior prediction device 1 will be sequentially described. <<Object Detection Unit 10>> The object detection unit 10 is a functional unit that detects moving objects including other vehicles such as vehicles and people, motorcycles, bicycles, pedestrians, stationary objects including parked vehicles and buildings, using external information from external sensors. Further, the object detection unit 10 tracks the moving objects and estimates state quantities such as the position, speed, and direction of the objects.

[0021] <<Peripheral information management unit 11>> The peripheral information management unit 11 is a functional unit that manages peripheral information around the host vehicle. The peripheral information includes road structure, display of signals and road signs, lanes, and the like. The peripheral information can be held in advance as map information, acquired through communication with the outside, or acquired by recognition using external sensors mounted on the vehicle. Acquisition by external sensors is performed, for example, by inputting a camera image and performing region division using a learned machine learning model.

[0022] <<Action prediction unit 12>> The action prediction unit 12 is a functional unit that predicts one or more future actions of an object. In this embodiment, as a future action, a sequence of trajectory points that the object will pass through in the future is predicted. Each point of the trajectory point sequence has position coordinates and the time of a future moment as state quantities, and a trajectory is formed by connecting the points in time series. In order to predict the trajectory point sequence, it is necessary to set the prediction period and the time resolution of the point sequence, respectively. Also, it is necessary to set how many trajectory point sequences are to be predicted. Therefore, the action prediction unit 12 sets the number of predicted trajectory point sequences, the prediction period, and the time resolution as output conditions for prediction and executes the prediction. The action pattern specifying unit 12a, the output condition setting unit 12b, and the prediction processing unit 12c that sequentially execute these processes will be described in detail below.

[0023] In the action pattern identification unit 12a, one or more possible action patterns of the object are identified based on at least one of the state of the object or the surrounding information. The action pattern refers to the type of action assumed for the object, such as turning right or left, going straight, stopping, and parking. Further, the action pattern identification unit 12a can also calculate accompanying information such as the occurrence probability or risk degree of each action pattern. The method for identifying the action pattern and the method for calculating the accompanying information may be conditional branching based on simple rules, may be a machine learning method, or may be a statistical method. Here, it is desirable that the method for identifying the action pattern has a low computational load. This is because if a method with a high computational load is used in the action pattern identification unit 12a, even if the computational load is reduced in the subsequent prediction process, the overall computational load cannot be reduced.

[0024] In the output condition setting unit 12b, the output conditions for the prediction are set based on the type of each action pattern, the accompanying information, the computing resources of the computer, and the like.

[0025] In the prediction processing unit 12c, the prediction processing is executed according to the output conditions set by the output condition setting unit 12b. The prediction method may be a machine learning method, a method for searching for an optimal trajectory, or a method based on rule-based conditional branching. Further, in order to cope with the dynamically set output conditions, the prediction processing unit 12c can hold a plurality of prediction methods. For example, in order to perform predictions with different time resolutions, a machine learning model suitable for making a prediction with a low time resolution and a machine learning model suitable for making a prediction with a high time resolution can be held. Also, for the prediction of one trajectory point sequence, a plurality of prediction methods can be combined and used.

[0026] <Operation example of the action prediction device 1> Next, with reference to the flowchart of FIG. 2, an operation example of the action prediction device 1 of this embodiment will be described in detail. Hereinafter, an in-vehicle camera will be used as the external sensor, and the operation of the action prediction device 1 will be described by taking the case where the object to be predicted is another vehicle in the vicinity of the host vehicle as an example.

[0027] First, in procedure P1, the object detection unit 10 detects an object to be predicted (hereinafter referred to as the prediction target object) using a learned machine learning model that utilizes camera images. Also, by tracking the object to be predicted in a time series and applying a Kalman filter, state quantities such as the position, speed, and orientation of the prediction target object are estimated.

[0028] Next, in procedure P2, the surrounding information management unit 11 acquires information on the surroundings of the host vehicle. Specifically, the surrounding information management unit 11 identifies the current position of the host vehicle based on GPS information and acquires map information around the host vehicle from the server. Also, the surrounding information management unit 11 acquires the shape of the road and the signal display using a learned machine learning model that utilizes camera images.

[0029] In procedure P3, the action pattern identification unit 12a identifies the action patterns that the prediction target object can take. Hereinafter, with reference to FIGS. 3A to 3C, an example of the method for identifying action patterns will be described. Although the illustration is omitted, it is assumed that the host vehicle exists at a position where each vehicle in the figure can be observed by an in-vehicle camera (external sensor).

[0030] FIG. 3A is an example of identifying an action pattern based on the position and connection of lanes. In this example, since only a straight-ahead lane exists in front of the prediction target vehicle V1, straight-ahead is identified as the action pattern. Also, since two lanes, left turn and straight-ahead, exist in front of the prediction target vehicle V2, these two are identified as the action patterns. Further, since three lanes, left turn, straight-ahead, and right turn, exist in front of the prediction target vehicle V3, these three are identified as the action patterns.

[0031] FIG. 3B is an example of identifying an action pattern based on the display of the traffic signal in front of the prediction target vehicle V4. In this example, when the traffic signal is red, stop is identified as the action pattern, when it is blue, straight-ahead is identified as the action pattern, and when it is yellow, two action patterns, stop and straight-ahead, are identified respectively.

[0032] FIG. 3C is an example of specifying an action pattern based on the positional relationship among the vehicle V5 to be predicted, an oncoming vehicle, and an obstacle. In the situation shown in the figure, it is assumed that the road is narrowed by the obstacle and only one vehicle can pass through at a time. Therefore, in this example, as the possible action patterns of the vehicle V5 to be predicted, two action patterns are specified: a pattern in which the vehicle V5 to be predicted passes through between the obstacles first, and a pattern in which the oncoming vehicle passes through between the obstacles first.

[0033] In procedure P4, the action pattern specifying unit 12a calculates the occurrence probability of each action pattern. Hereinafter, the method of calculating the occurrence probability of the action pattern of each vehicle in FIGS. 3A to 3C will be described.

[0034] Regarding the vehicle V1 to be predicted in FIG. 3A, since the action pattern is only straight ahead, the action pattern specifying unit 12a calculates the occurrence probability of going straight as 100%.

[0035] Regarding the vehicle V2 to be predicted and the vehicle V3 to be predicted in FIG. 3A, the method of calculating the probability differs depending on whether the external sensor can recognize the blinking of the turn signal of each vehicle to be predicted. First, consider the case where the blinking of the left turn signal is recognized. In this case, the occurrence probability of a left turn is assumed to match the recognition correct rate of the blinking of the left turn signal, and the occurrence probability of each action pattern is set by equally distributing the remaining probability among the remaining action patterns. The case where the blinking of the right turn signal is recognized is calculated in the same way.

[0036] When the blinking of the turn signal cannot be recognized, the occurrence probability of each action pattern is calculated based on the direction of the vehicle. Therefore, the action pattern specifying unit 12a previously holds a table showing the relationship between the direction of the vehicle at the time of entering the intersection and the occurrence probability of each action pattern such as a left turn, going straight, and a right turn. This table is created by previously collecting data of vehicles entering the intersection. Thereby, the action pattern specifying unit 12a can obtain the occurrence probability of each action pattern even when the blinking of the turn signal cannot be recognized, based on the direction of the vehicle to be predicted estimated by the object detection unit 10 and the table.

[0037] Regarding the target vehicle V4 in FIG. 3B, when the action pattern specifying unit 12a recognizes the red light of the traffic signal, it sets the stop probability to 100%, and when it recognizes the green light of the traffic signal, it sets the straight-ahead probability to 100%. On the other hand, when the action pattern specifying unit 12a recognizes the yellow light, it calculates the probability based on the distance from the vehicle to the traffic signal. Therefore, the action pattern specifying unit 12a previously holds a table showing the relationship between the distance from the vehicle to the traffic signal and the occurrence probability of each action pattern. This table is created by collecting data of vehicles entering the intersection in advance. Thereby, the action pattern specifying unit 12a can calculate the distance from the vehicle to the traffic signal using the vehicle position information estimated by the object detection unit 10 and the position information of the traffic signal acquired by the surrounding information management unit 11, and use it in combination with the table to obtain the occurrence probability of each action pattern.

[0038] Regarding the target vehicle V5 in FIG. 3C, the action pattern specifying unit 12a calculates the occurrence probability of each action pattern based on the distance between the vehicle and the straight line defined by connecting two obstacles in a straight line. Therefore, the action pattern specifying unit 12a previously holds a table showing the relationship between the distance to the obstacle and the occurrence probability of each action pattern. Thereby, the action pattern specifying unit 12a can calculate the distance using the vehicle position information estimated by the object detection unit 10 and the position information of the obstacle, and use it in combination with the table to obtain the occurrence probability of each action pattern.

[0039] Hereinafter, taking the target vehicle V3 in FIG. 3A as an example, an example will be described in which the prediction output conditions corresponding to each action pattern are set in procedures P5, P6, and P7, and the trajectories T1, T2, and T3 as shown in FIG. 4 are predicted in procedure P8. It is assumed that in procedure P4, the occurrence probabilities of each action pattern, i.e., 80% for left turn, 17% for straight ahead, and 3% for right turn, have been calculated.

[0040] In step P5, the output condition setting unit 12b sets the number of predicted trajectories for each action pattern. Therefore, the output condition setting unit 12b pre - holds a table showing the correspondence between the occurrence probability and the number of predicted trajectories. For example, if the occurrence probability of an action pattern is less than 10%, zero trajectories are predicted; if it is 10% or more and less than 50%, one trajectory is predicted; and if it is 50% or more, two trajectories are predicted. Considering it together with the occurrence probability calculated in step P4, in this example, the output conditions for the number of predicted trajectories are set such that there are two predicted trajectories for a left turn, one predicted trajectory for going straight, and zero predicted trajectories for a right turn.

[0041] In step P6, the output condition setting unit 12b sets the time resolution of each predicted trajectory. The output condition setting unit 12b pre - holds a table showing the correspondence between the type of action pattern and the time resolution. For example, when the action pattern is going straight, the dot sequence is predicted at 1.0 - second intervals, and when it is a left turn or a right turn, the dot sequence is predicted at 0.3 - second intervals. Thus, for a left turn, the interval is set to 0.3 seconds, and for going straight, the interval is set to 1.0 seconds.

[0042] In step P7, the output condition setting unit 12b sets the prediction length of each predicted trajectory. Therefore, the output condition setting unit 12b pre - holds a table showing the correspondence between the occurrence probability and the prediction length. For example, if the occurrence probability of an action pattern is less than 50%, a dot sequence for 2 seconds is predicted, and if it is 50% or more, a dot sequence for 3 seconds is predicted. Accordingly, in this example, using it together with the occurrence probability calculated in step P4, the prediction length of the left - turn predicted trajectory is set to 3 seconds, and the prediction length of the straight - ahead predicted trajectory is set to 2 seconds.

[0043] In step P8, the prediction processing unit 12c predicts the trajectory based on the output conditions set above. Therefore, the prediction processing unit 12c has a prediction model capable of changing the time resolution. The input of this prediction model is the state quantity of the vehicle to be predicted, the road shape, and the time resolution, and the output of the prediction model is the candidate for the position of the vehicle to be predicted one time step ahead and the movement probability corresponding to each position. Here, the resolution of one time step is determined by the time resolution of the input. The prediction model is an encoder-decoder model using a known LSTM (Long Short-Term Memory). After encoding the input information with the encoder, the decoder obtains the position coordinates.

[0044] A method for predicting two trajectories corresponding to the turning-left behavior pattern using such a prediction model will be described. First, a method for obtaining the first trajectory will be described. Using the state quantity of the object to be predicted obtained in step P1, the road shape obtained in step P2, and the time resolution determined in step P6 as inputs, candidates for the position of the object to be predicted one step ahead and the movement probability are obtained. Next, using the position with the highest movement probability as the input to the prediction model, the position of the object to be predicted in the next step is obtained. This operation is repeated to obtain a sequence of points for 3 seconds, which is the desired length, as the first predicted trajectory.

[0045] Next, a method for obtaining the second predicted trajectory will be described. Using the current position of the object to be predicted obtained in step P1 as the input, candidates for the position of the object to be predicted one step ahead and the movement probability are obtained. Here, using the position with the second highest movement probability as the input to the prediction model, the position of the object to be predicted in the next step is obtained. Thereafter, the same operation is repeated using the position with the highest movement probability as the input to the prediction model. Then, a sequence of points for 3 seconds, which is the desired length, is obtained as the second predicted trajectory.

[0046] Through these operations, two trajectories T1 and T2 under the desired output conditions can be obtained. In the case of going straight, with the time resolution of the input set to 1.0 second, a trajectory T3 for 2 seconds is obtained in the same way as when obtaining the first trajectory in the case of a left turn.

[0047] As described above, the behavior prediction device 1 of the present embodiment identifies one or more possible behavior patterns of the object to be predicted, sets prediction output conditions for each behavior pattern, and outputs a prediction for each behavior pattern. As a result, it is possible to properly use a prediction method with a low computational load and a prediction method with a high computational load. The computational load can be reduced as compared with the case where prediction is always performed using a method with a high computational load.

[0048] <Operation and Effect of the Above Operation Example> Through the above operation example, one or more possible behavior patterns of the object to be predicted are identified. As a result, subsequent processing according to the behavior pattern can be executed.

[0049] Through the above operation example, the occurrence probability is predicted for each behavior pattern. As a result, it becomes possible to set prediction output conditions according to the occurrence probability of each behavior pattern, and it becomes possible to set output conditions according to the importance of the prediction.

[0050] Through the above operation example, the number of prediction outputs corresponding to each behavior pattern is set. As a result, the computational load can be reduced by reducing the number of predictions corresponding to behavior patterns with low importance or not performing predictions.

[0051] Through the above operation example, the time resolution corresponding to each prediction result is set. As a result, by reducing the time resolution according to the behavior pattern, the number of operations of the prediction model can be reduced, and the computational load can be reduced.

[0052] Through the above operation example, the prediction period corresponding to each prediction result is set. As a result, by shortening the prediction period of predictions with low importance, the number of operations of the prediction model can be reduced, and the computational load can be reduced.

[0053] <Effect of Reducing Computational Load> Estimate the computational load reduced by this embodiment. For simplicity here, assume that the predicted score and the computational load are in a proportional relationship, and compare the case of predicting for 3 seconds at 0.3-second intervals, two for each of the three action patterns of left turn, straight ahead, and right turn, with a total of 60 points predicted, with the prediction by this embodiment. In the former case, six trajectories are predicted, and each trajectory has 10 prediction points, so a total of 60 points are predicted. In contrast, in the latter case, a total of three trajectories, two for left turn and one for straight ahead, are predicted. Each trajectory corresponding to a left turn has 10 prediction points, and the trajectory corresponding to straight ahead has 2 prediction points, so a total of 22 points are predicted. Therefore, by the operation of this embodiment described above, the computational load can be suppressed to about one-third.

[0054] <When considering the degree of danger instead of the occurrence probability when determining the output conditions> In the above embodiments, the output conditions in the subsequent stage were set using the occurrence probability for each action pattern, but the degree of danger can also be used instead of the occurrence probability. Here, assume that the degree of danger of the action pattern is obtained in three levels: low, medium, and high. Fig. 5 shows the prediction results when straight ahead and right turn are specified as the action patterns. As shown in the figure, since an oncoming vehicle is approaching from the front of the predicted vehicle V6, there is a possibility of a right-straight accident if the predicted vehicle V6 makes a right turn. Therefore, when there is an oncoming vehicle in the oncoming lane at a T-junction, the degree of danger of a right turn is determined to be high. On the other hand, when the predicted vehicle V6 goes straight, the possibility of a collision accident with the oncoming vehicle is low, so the degree of danger of going straight is determined to be low. Then, compared with the trajectory T5 of the action pattern with a high degree of danger, the time resolution of the trajectory T4 of the action pattern with a low degree of danger is set low and the prediction is made.

[0055] From the above, the degree of danger is predicted for each action pattern. This enables setting of the output conditions of the prediction according to the degree of danger of each action pattern, and enables setting of the output conditions according to the importance of the prediction.

[0056] <When combining multiple prediction methods for trajectory prediction> In the embodiment described so far, the prediction using a single prediction model has been executed. However, when predicting one trajectory, multiple prediction methods can be combined for prediction.

[0057] For example, consider the trajectory prediction when the target vehicle V7 in FIG. 6 makes a right turn. As the output conditions here, the number of predictions is set to 1, and the prediction period is set to 5 seconds. Regarding the time resolution, the first 3 seconds are set to 0.3 seconds (section A1 in the figure), and the remaining 2 seconds are set to 1.0 seconds (section A2 in the figure). In this case, the trajectory prediction for the first 3 seconds is performed using the prediction model. For the remaining 2 seconds of trajectory prediction, it is assumed that the vehicle moves linearly, and linear prediction is performed. As a result, compared with the case of using the prediction model for all 5 seconds of trajectory prediction, the number of prediction points decreases, and the computational load can be reduced. Also, since linear prediction has a lower computational load compared to prediction using a machine learning model, the computational load for one-point prediction can also be reduced.

[0058] <When setting an upper limit on the number of prediction points> In addition to the present embodiment described above, in the present invention, an upper limit on the number of prediction points can be set to set the output conditions of the prediction. Here, it is assumed that if the total number of prediction points is up to 10, calculations can be performed sufficiently fast. For simplicity, hereinafter, the prediction period is fixed at 3.0 seconds, and the number of predicted trajectories is 1 for each action pattern. Consider the case when, as in the target vehicle V8 in FIG. 7, the left-turn occurrence probability is calculated as 50%, the straight-ahead occurrence probability is 30%, and the right-turn occurrence probability is 20%. In this case, the occurrence probability is multiplied by 10 points, and the prediction points are allocated as 5 points for a left turn, 3 points for straight ahead, and 2 points for a right turn. By dividing each point by the prediction period, the respective time resolutions are set to 0.6 seconds for a left turn, 1.0 seconds for straight ahead, and 1.5 seconds for a right turn. Finally, using each time resolution as the input to the prediction model, trajectories T7, T8, and T9 are predicted for each action pattern. As a result, within a range where the computational load does not exceed the allowable value, the computational resources can be appropriately allocated to predict the trajectory.

[0059] <Effects of this embodiment> According to the action prediction device of this embodiment described above, by dynamically changing the prediction output conditions according to the possible action patterns of the object to be predicted, it is possible to reduce the amount of calculation. Note that the present invention is not limited to the above embodiment, and various modifications understandable by those skilled in the art can be made within the scope of the present invention.

Explanation of Signs

[0060] 1 Action prediction device 10 Object detection unit 11 Peripheral information management unit 12 Action prediction unit 12a Action pattern identification unit 12b Output condition setting unit 12c Prediction processing unit V1 to V8 Vehicles to be predicted T1 to T9 Prediction trajectories

Claims

1. An object detection unit that detects a prediction target based on external information from an external sensor, and an action prediction unit that predicts the action of the prediction target, wherein the action prediction unit has an action pattern specifying unit that specifies an action pattern that the prediction target can execute based on at least one of the state of the prediction target or the surrounding environment of the prediction target, an output condition setting unit that sets an output condition for each of the specified action patterns, and a prediction processing unit that outputs a prediction result for each of the action patterns based on the set output condition, and is characterized by having the above components.

2. In the action prediction device according to Claim 1, the action pattern specifying unit is characterized by specifying the action pattern using at least one of a road structure, traffic rules, signs, and a signal indication on the path of the prediction target.

3. In the action prediction device according to Claim 1, the action pattern specifying unit is characterized by predicting a generation probability for each of the action patterns.

4. In the action prediction device according to Claim 1, the action pattern specifying unit is characterized by predicting a risk level for each of the action patterns.

5. In the action prediction device according to Claim 1, the output condition setting unit is characterized by setting, as the output condition, the number of generated prediction trajectories including prediction points obtained by predicting the action of the prediction target in time series for each of the action patterns.

6. In the action prediction device according to Claim 1, the output condition setting unit is characterized by setting, as the output condition, the time interval between prediction points of a prediction trajectory including prediction points obtained by predicting the action of the prediction target in time series for each of the action patterns.

7. In the action prediction device according to Claim 1, the output condition setting unit is characterized by setting, as the output condition, the prediction period of a prediction trajectory including prediction points obtained by predicting the action of the prediction target in time series for each of the action patterns.

8. In the action prediction device according to Claim 1, the prediction processing unit is characterized by outputting a prediction result by combining a plurality of prediction methods.

9. the external sensor, and a vehicle control device that controls a steering system, a drive system, and a braking system The behavior prediction device according to claim 1, A vehicle equipped with The vehicle control device controls the steering system, the drive system, and the braking system based on the prediction result of the behavior prediction device. A vehicle characterized by this.

10. A behavior prediction method executed by a behavior prediction device equipped with a CPU, An object detection step of detecting a prediction object based on external information from an external sensor; A behavior pattern specifying step of specifying a behavior pattern that the prediction object can execute based on the state of the prediction object or the surrounding environment of the prediction object; An output condition setting step of setting an output condition for each of the specified behavior patterns; A prediction processing step of outputting a prediction result for each of the behavior patterns based on the set output conditions; By having, A behavior prediction method characterized by predicting the behavior of the prediction object.

Images