JP2009262702A - Safe driving support system - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To support safe driving by adaptively learning relevance between driving operation of a driver and an external environment corresponding to various state changes in cycle and speed generated to the driver during driving, and by recognizing deviation from an ordinary internal state of the driver. <P>SOLUTION: A correspondence between risk of travel environment recognized in an environment risk recognition part 2 and a characteristic amount of driving operation data quantized by an operation characteristic amount discretization part 5 is learned in a model learning part 6 to obtain a model parameter according to the cycle set by a schedule control part 9. A state estimation part 7 estimates an internal state of the driver from the relevance between the operation characteristic amount and an environment risk level by use of the already learned model parameter by the estimated cycle set by the schedule control part 9. An alarm-support part 8 compares the internal state of the driver and the environment risk level, evaluates whether a vehicle is in an essentially safe or dangerous state, and performs an alarm, the operation support, and the like. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a safe driving support system that estimates the state of a driver who drives a moving body and provides support for safe driving.

  In recent years, with the goal of enabling safe movement of moving objects such as automobiles, techniques have been developed to ensure the safety by monitoring the state of the driver while driving to detect a decrease in wakefulness and falling asleep. . The estimation of the driver state is mainly performed by measuring the living body state and that estimated from the driving operation data, and various proposals have been made regarding these techniques.

  As a technique related to estimation of a driver state by measuring a biological state, there are techniques disclosed in Patent Documents 1 to 5. The technology of Patent Document 1 estimates the driver's state by measuring eye movement using an electrode attached to the driver's face, a camera that captures the eyeball, and the like. The degree of arousal is estimated based on the opening of the heel.

  Moreover, the technique of patent document 3 and patent document 4 estimates a driver's state by measuring a driver's heartbeat signal, and the technique of patent document 5 detects a brain current signal. .

On the other hand, there are technologies disclosed in Patent Documents 6 and 7 as technologies related to estimation of the driver state based on driving operation data. The technique of Patent Document 6 discriminates driving operation data of a driver under a predetermined condition and reflects it in a change in vehicle control characteristics. The technique of Patent Document 7 is based on fuzzy inference from vehicle state data. Is intended to estimate the intention and psychological state.
JP 2003-230552 A JP 2004-89272 A Japanese Patent No. 3596198 Japanese Patent Laid-Open No. 7-10024 JP-T-2006-524157 Japanese Patent No. 3058966 Japanese Patent No. 3036183

  As mentioned above, in order to realize safe driving of automobiles, it is effective to perform warnings, change control characteristics, assist operations, etc. using technology that estimates the state of the driver. Has been made.

  However, in the technology for performing living body measurement for estimating the driver state as disclosed in References 1 to 5, generally, a measurement amount having a large noise or individual difference must be dealt with. There is a problem that it cannot cope with the environment and the diversity of people who use it.

  That is, Patent Document 1 gives an example in which electrodes are arranged on the face for measurement of eye movement, but it is troublesome that a driver must be equipped with a measurement device that involves physical restraint during actual driving. Not only lacks reality.

  In Patent Document 2, data is measured based on information from a camera that captures the eyes of the driver. However, from the cost of adding additional devices to the vehicle, changes in the driver's attitude, and the external environment such as West Japan There is a possibility that photographing cannot be performed normally due to the influence of light.

  Patent Document 3 describes only electrodes and other devices installed on the seat and the steering, and Patent Document 4 merely describes psychological state detection means, but both provide examples of measuring heartbeats. However, these technologies also measure the electrocardiogram, which is relatively easily disturbed by noise, require additional equipment on the vehicle, and determine the characteristics of the heartbeat that vary from individual to individual. It must be said that there is a problem in view of the fact that it is used in the actual environment of an automobile, such as the point that must be performed.

  Patent Document 5 describes non-invasive measurement by mounting a device such as MRI on the vehicle, but the device described in the document directly or indirectly restrains the driver's posture and other conditions. Similarly, it must be said that the industrial application at the present time is not realistic, including its cost.

  In addition, as a matter common to Patent Documents 1 to 5, we are trying to measure the degree of arousal and fatigue with signs that are considered common to general humans even though there is advance calibration. There are some that do not actively deal with environmental relevance, but rather focus on smoothing out personal differences through preprocessing. Although such a method can be expected to some extent, there is no mechanism to actively handle individual differences and the external environment when estimating the driver state with higher accuracy. There is a certain limit to its performance.

  On the other hand, in Patent Documents 6 and 7, which use operation or vehicle data for estimating the driver state, when the driver's operation satisfies a predetermined condition, the vehicle characteristics in a predetermined direction that seems to correspond to the intention. Is changing. In this method, the vehicle characteristics are changed with the characteristics of the operation that are considered common to general humans, so the state estimation according to each driver's individuality is not made in the true sense, and it is wrong for some people There is a risk of being estimated to the state.

  As described above, the conventional technology not only has problems to be considered with respect to noise and additional costs, but also has insufficient capability due to uniform indicators and conditions for estimating the driver state. It is difficult to broadly estimate various state changes due to the period and speed generated, which is insufficient to realize safe driving of a car.

  The present invention has been made in view of the above circumstances, adaptively learning the relationship between the driving operation of the driver and the external environment in response to various state changes in the period and speed generated in the driver during driving, An object of the present invention is to provide a safe driving support system that can recognize a deviation from the normal internal state of a driver and support safe driving.

  In order to achieve the above object, a safe driving support system according to the present invention is a safe driving support system that estimates the state of a driver driving a mobile body and provides support for safe driving, and includes an external environment of the mobile body. To extract the feature quantity and to recognize the environmental risk included in the external environment based on the feature quantity, and the probabilistic state transition model for the correspondence between the environmental risk and the driving operation of the driver. The model learning unit that learns and constructs the learning model, and based on the learning model, the driver's risk recognition state is determined from the driving operation data including vehicle behavior data as a reflection of the driving operation of the driving driver. Compare the environmental risk with the driver's internal state, and obtain support information related to safe driving of the moving body And a schedule control unit that adaptively varies the estimation period based on the stochastic state transition model by setting data input to the stochastic state transition model in a plurality of different periods. It is characterized by.

  According to the present invention, the relationship between the driving operation of the driver and the external environment can be learned adaptively in response to various state changes in the period and speed generated in the driver during driving. Recognize deviations from internal conditions and support safe driving.

  Embodiments of the present invention will be described below with reference to the drawings. 1 to 16 relate to an embodiment of the present invention, FIG. 1 is a basic configuration diagram of a safe driving support system, FIG. 2 is an explanatory diagram showing a learning phase and an estimation phase of the safe driving support system, and FIG. FIG. 4 is an explanatory diagram showing a stochastic state transition, FIG. 5 is a graph showing a state transition probability after learning, and FIG. 6 is a graph showing an output probability after learning. 7 is an explanatory diagram of state estimation by the maximum likelihood sequence, FIG. 8 is an explanatory diagram of state estimation by a forward algorithm, FIG. 9 is an explanatory diagram showing state estimation by time driven, and FIG. 10 is an explanatory diagram showing state estimation by event driven. 11 is an explanatory diagram showing the occurrence of a transition trigger by event driven, FIG. 12 is a graph showing the state transition probability by time driven and event driven, and FIG. 13 is time driven. 14 is a graph showing an example of internal state estimation by event driven, FIG. 14 is an explanatory diagram of adaptive sampling, FIG. 15 is a graph showing an example of internal state estimation using both time driven and adaptive sampling, and FIG. It is explanatory drawing shown.

  The safe driving support system according to the present invention recognizes a deviation from the normal internal state of the driver by comparing the risk included in the external environment with the risk recognized by the driver when driving a moving body such as an automobile. Thus, it is possible to provide appropriate driving assistance without unnecessary intervention, and it is possible to improve the safety when driving a car while eliminating the “dissonance between system and driver”.

  Hereinafter, in the present embodiment, driving assistance for automobiles will be described. The safe driving support system according to the present embodiment includes an operation of a steering wheel, an accelerator, a brake, etc. including vehicle behavior data as a reflection of the driving operation of the driver as main data for estimating the driver state (internal state of the driver). Data is used, and the “relationship between the surrounding environment of the vehicle and the operation” in the normal driving of the driver is used as a probabilistic reference model as a driver state estimation index without using uniform conditions.

  At that time, the system continuously updates the model by continuously driving the vehicle, improves the state estimation accuracy according to the individual, sets the driver state estimation period optimally, and To increase.

  For example, if importance is attached to the estimation accuracy of the period corresponding to “moment relaxation of moment”, the estimation accuracy for states such as “blurred”, “decreased arousal”, and “accumulation of fatigue” with a longer period deteriorates. There is a possibility that it has a high estimation capability for a state change at a specific frequency during operation, but the estimation accuracy may be relatively deteriorated for state changes at other frequencies. For this reason, in this system, it is possible to estimate various internal states by optimally setting the state estimation cycle, thereby widely estimating state changes with various cycles and speeds generated during driving. be able to.

  Specifically, as shown in FIG. 1, the safe driving support system 1 in the present embodiment includes a single computer system or a plurality of computer systems connected via a network or the like, such as a camera or a radar. Environmental risk recognition unit 2 for recognizing risks (environmental risks) included in the external environment around the vehicle from the external information by the sensing device, environmental risk discretization unit 3 for discretizing the environmental risk, and operation for measuring the driving operation of the driver The measurement unit 4, the operation feature quantity discretization unit 5 that discretizes (quantizes) the feature quantity of the measured driving operation data of the driver, the correspondence between the driving operation (operation feature quantity) and the recognition environment (environmental risk level). Model learning unit 6 that learns using a probabilistic normative model, the driver state from the relationship between the learned model and the current operation data / exterior environment A state estimation unit 7 to be determined, and a warning / support unit 8 that obtains support information for safely driving the vehicle according to the estimated driver state, and performs warnings and operation support, and further includes a driver In order to perform optimal estimation in response to changes in conditions such as arousal level and fatigue level and running conditions, data input to the probabilistic normative model is set to multiple different periods, and the estimation period based on the probabilistic normative model is set. An adaptively variable schedule control unit 9 is provided.

  In this embodiment, an in-vehicle camera is used as a sensing device for the external environment.

  As shown in FIG. 2, the system operation of the safe driving support system 1 includes a “learning phase” in which a model of the internal state of the driver is learned, and an internal state of the driver is estimated based on the obtained model. It is divided into “estimation phase”.

  In the learning phase, when the front of the vehicle is imaged with an in-vehicle camera and the risk level of the driving environment is recognized by learning, the correspondence between this risk level and the operation feature value data obtained by quantizing the feature value of the driver's drive operation data is obtained. Learn and acquire learned parameters to build a learning model.

  In the estimation phase, the relationship between the operation feature amount and the environmental risk level is acquired using the learned model parameter, and the current internal state of the driver is estimated. The estimated current internal state of the driver is compared with the environmental risk level in the current driving environment to assess whether the vehicle is inherently safe or dangerous.

  The learning phase and the estimation phase do not actually proceed separately from each other, but both proceed simultaneously. As a result, it is possible to realize a function that can adapt to changes in the environment in which the driver is driving and changes in the characteristics of the driver itself, and to provide a system that increases the performance as the vehicle travels. Can do. In addition, the model parameters in the learning phase can all be acquired by online learning while driving. However, it is recognized that the learning speed can be improved by using offline learning in part with a simulator. An improvement in accuracy can be expected.

  Next, the function of each part of the safe driving support system 1 will be described. The environmental risk recognition unit 2 degenerates the state of the environment outside the vehicle into a single scalar value (or vector) by learning the relationship between the feature quantity of the image obtained from the in-vehicle camera and the risk level at that time. To do. The environmental risk discretization unit 3 discretizes the environmental risk level from the environmental risk recognition unit 2 by data division or the like using a threshold value, and transmits it to the model learning unit 6 as an observation state of the probabilistic normative model.

  For the recognition of the risk level from the image feature amount, for example, the technology of the online risk learning system proposed in Japanese Patent Application No. 2007-77625 by the present applicant can be adopted. As described in detail in Japanese Patent Application No. 2007-77625, this technique is based on an image feature amount (edge information, motion information, color information, etc.) of an N-dimensional vector depending on an event such as a sudden accelerator return operation or a brake depression. Is converted into a one-dimensional state, and the risk contained in the environment is learned and recognized from the correlation between this state and the teacher information created from the vehicle information (driver operation information).

  In the present embodiment, an example using the risk level extracted from the image feature amount will be described. However, the risk level is not limited to this, and for example, the risk level is extracted from the inter-vehicle distance or the like. Anyway.

  The operation feature quantity discretization unit 5 uses the data from the sensor or the operation data of the driver such as accelerator, brake, steering, etc. acquired by the operation measurement unit 4 via the in-vehicle network (not shown) as the characteristics of the measurement frequency. In response, the signal is quantized and transmitted to the model learning unit 6 and the state estimation unit 7 as an observation symbol of the probabilistic reference model.

  In other words, the driving operation data has a very large amount of information as it is, and it is difficult to handle the relationship with the risk. For this reason, statistical processing to learn features while preventing loss of information contained in the data by appropriately quantizing in consideration of the distribution (appearance tendency) of the observed data Is possible.

  Quantization of observation data can be performed by data division or data degeneration using thresholds. In this embodiment, driving operation data is quantized using a self-organizing map (SOM). . SOM is a type of neural network that models the visual cortex of the cerebral cortex of living organisms. Each unit arranged in M dimensions has a vector value (usually called the weight of the connection with the input). The winner unit is determined based on the vector distance.

  Then, the reference vector values of the winner unit and the surrounding units are updated so as to approach the input vector. By repeating this, competitive learning is performed so that the entire distribution of the input data can be expressed optimally, the dimensions of the input information are compressed based on this competitive learning, and clustering and visualization are performed according to the characteristics of the data Can do.

  In this embodiment, an example in which input data is degenerated by unsupervised competitive learning using S0M will be described. However, a vector quantization (LVQ; Learning Vector Quantization) model that is supervised competitive learning can also be used. It is.

  Examples of SOM parameters are shown below. The feature values adopted are the steering angular speed, accelerator opening, brake pressure, accelerator opening change, brake pressure change, and vehicle speed. However, absolute values are used for the rudder angular velocity, and the accelerator opening and the brake pressure are normalized to be dimensionless amounts from -100 (maximum brake pressure) to +100 (accelerator fully open). The change amount of the accelerator and the brake is also expressed as the change amount of the normalized value. As a result, the dimension of the input feature quantity is 4, which is compressed to one dimension by SOM and quantized to 256 units.

[SOM parameters]
Number of SOM dimensions 1
256 units
Input feature value ADSTR (steering angular velocity absolute value)
PEDAL (accelerator / brake operation normalized amount)
DPEDAL (PEDAL value change amount)
SPEED (vehicle speed)

  FIG. 3 shows data learned by SOM (however, only three of the four input dimensions), and data obtained by traveling in an urban area (corresponding to 5-hour driving) is learned by SOM. The measurement data is plotted like a cloud, and the one-dimensional SOM network is arranged in a string shape. It can be seen from FIG. 3 that learning is performed according to the density and distribution of measurement data, and that units are appropriately arranged. The measured data is quantized and reduced to one-dimensional data represented by a unit number by being approximated by the nearest SOM unit.

  The unit arrangement of the SOM is determined by learning measured driving data, and it can be expected that the same person will adapt to changes in driving characteristics due to time variations even for the same person.

  Further, the period for discretizing the SOM and the environmental risk level is not a single fixed period, but is set by the schedule control unit 9 according to the state estimation period. The discretization period and the state estimation period will be described later.

  The model learning unit 6 acquires the relationship between the operation feature amount and the risk level by learning, and enables probabilistic calculation for estimating the internal state of the driver. As shown in FIG. 4 (a), human behavior changes according to mental states such as relief, tension, anxiety, impatience, and anger, and transitions thereof, and is not necessarily deterministic but expressed as probabilistic behavior. can do. Similarly, as shown in FIG. 4 (b), the driver's driving behavior includes a stochastic operation output for scenes such as following a preceding vehicle, passing, parking, lane change, and merging. It appears.

  Therefore, the model learning unit 6 models the internal state of the driver using a hidden Markov model (HMM), which is a kind of probabilistic state transition model, as a reference model for expressing the probabilistic behavior of human behavior. Turn into. The HMM is a model that assumes that a target internal state (state) transitions by a probabilistic conditional branch and that a signal is output to the outside with a different probability depending on the transitioned state.

  In the model using the HMM, in the task of estimating the risk level that the driver is aware of, the risk level that is currently conscious corresponds to the state of the HMM, as shown in FIG. Observed driving operation data corresponds to a signal output to the outside. FIG. 4C illustrates an example in which the risk levels are set to 5 levels and numbers 1 to 5 are assigned to the respective states. The number 1 recognizes that the risk of the external environment is the lowest. The state, number 5, indicates a state where the driver recognizes that the risk of the external environment is the highest. In addition, the risk level that is actually handled is 10 levels of 0, 1,.

  In this way, the driving operation data is discretized, the appearance tendency of the data is obtained, and the internal state of the driver is approximated as a probabilistic model, so that it is deterministically derived from the external environment like the operation data in actual driving. It is possible to appropriately handle information that cannot be handled. However, in order to estimate from which state the operation data observed at the time of estimation is output, it is necessary to perform two probability calculations: a state transition probability and an operation output probability.

  For this reason, the model learning unit 6 first learns the state transition probability by a statistical method based on the scalar value transmitted from the environmental risk recognition unit 2 via the environmental risk discretization unit 3, and then operates the operation feature. Based on the data from the quantity discretization unit 5, the observation probability distribution (operation output probability) of the operation feature amount for each state is learned. For example, the steering angle, the accelerator opening, the brake pressure, and the speed, yaw rate, acceleration, and the like that reflect the operation are discretized as appropriate, and the number of observations for each discrete value is counted to statistically calculate the probability.

  Hereinafter, calculation processing of the state transition probability and the operation output probability using the HMM will be described.

[Specification example of HMM]
Drive frequency variable Number of states 10
Output signal SOM unit number (0 to 255)
Network structure All-state interconnection type

[Calculation of state transition probability]
In general, when driving a car, it is unlikely that there will be a driver who always causes an accident even though it is unintentional every time the car is driven. In other words, it can be regarded macroscopically that the driver is appropriately performing a driving operation corresponding to the risk level of the driving environment. Based on these assumptions, when data is collected and statistically processed over a relatively long time range, it is assumed that the transition of the internal state of the driver in the HMM depends on the transition of the risk level of the driving environment. Can do.

Therefore, the transition probability is calculated according to the following procedures (1-1) to (1-5), and the transition probability of the driver internal state is calculated by obtaining the transition probability of the risk level at the time of learning.
(1-1) Define drive frequency of state transition model (1-2) Define number of states in internal state (for example, 10 states)
(1-3) Discrete the risk level by the number of states (1-4) Count the number of transitions between each state (1-5) Calculate the statistical transition probability between each state

  FIG. 5 shows a graph of the state transition probability of the HMM obtained by learning from the measurement data in the urban area. The larger the state number, the higher the risk level. This graph shows the learning result by the HMM at a predetermined driving frequency. From the learning result, the self-transition (transition staying in the same numbered state) and the transition probability for one state above and below are large, the high risk state Then, it can be seen that the transition probability that the risk level suddenly decreases is relatively large. This is generally equivalent to the fact that moderate and moderate risks rise and fall continuously, and high risk factors disappear rapidly with the passing of the vehicle. It can be judged that there is nothing.

[Calculation of output signal probability]
The SOM described above is used for learning the output probability of the operation data feature in the state transitioned to a certain state. The measured operation data is dimensionally compressed and quantized into SOM unit numbers, and the statistical output signal probability is calculated by counting the number of times each unit number is observed. Since the output probability of each unit obtained for each state is different, it can be modeled that the operation tendency changes depending on the risk level, and the output probability for each state as shown in FIG. 6 can be obtained.

  When the above state transition probability and output probability are acquired by learning, the state estimation unit 7 next estimates the internal state of the driver using both pieces of information. This internal state can be estimated by a technique using maximum likelihood sequence estimation or a technique using a forward algorithm, and is appropriately selected according to system conditions and the like. In the maximum likelihood sequence estimation method, the driver's internal state is output as a discrete number of state numbers, and in the forward-looking algorithm, the driver's internal state is an expected value that expresses the state number precisely as a continuous number. Is output.

[Internal state estimation by maximum likelihood sequence estimation]
Maximum likelihood sequence estimation is a method that estimates the most likely state from which the currently observed operation data is output (maximum likelihood estimation), and the maximum likelihood when calculating the transition sequence. A probable sequence corresponds to specifying a sequence with the highest probability of occurrence. Here, the transition sequence of the internal state of the driver is calculated from the time series data of the measured operation data using the Viterbi algorithm, which is one of the maximum likelihood sequence estimation methods for the HMM, and obtained. It is estimated that the internal state of the driver is transitioning along the state transition series.

  The Viterbi algorithm is a method of calculating the maximum probability existing in each state at each time step sequentially from the beginning of the time series observation data based on the state transition probability and the output signal probability. This algorithm has a feature that the calculation amount is small because a calculation method similar to dynamic programming is used.

  In the Viterbi algorithm, the state having the maximum probability at the final step is determined as the estimated internal state at that step as a result of calculation up to the end of the observation data. Next, an operation called backtracking goes back one step at a time from there, and the state having the maximum state transition probability is determined. Finally, the maximum likelihood sequence is estimated by specifying the state having the highest probability of being present in all time steps. In the first step of probability calculation, the prior probability obtained from the state transition probability of the HMM is used as the existence probability in each state.

  Specifically, the maximum likelihood sequence is estimated by performing sequential calculation according to the following steps (2-1) to (2-4). Where π: prior probability, δ: state existence probability, φ: backtrack, a: state transition probability, b: output signal probability, P: estimated probability, q: estimated state series, o: zero vector, Subscript 1 indicates an initial value.

(2-1) Variables are initialized for each state i = 1,.
δ ₁ (i) = π ₁ b ₁ (o ₁ )
φ ₁ (i) = 0
(2-2) The following recursive calculation is executed for each time t = 1,..., T-1 and each state j = 1,.
δ _{t + 1} (j) = max _i [δ _t (i) a _ij ] b _j (o _{t + 1} )
φ _{t + 1} (j) = argmax _i [δ _t (i) a _ij ]
(2-3) End of recursive calculation P = max _i δ _T (i)
q _T = argmax _i δ _T (i)
(2-4) Restoration of optimal state transition sequence by backtracking (the following is executed for T = T-1,..., 1)
q _T = φ _{t + 1} (q _{t + 1} )

  In the above calculation, the Viterbi algorithm has a characteristic of using backtracking, which means that after the event is confirmed in time, the state is determined in the past direction, and it can be applied to online real-time state estimation as it is. There are difficulties. Therefore, when performing online state estimation, all time steps are treated as final steps in the Viterbi algorithm, and a state having the maximum existence probability in each step is output as an estimated state.

  For example, as shown in FIG. 7A, in the online state estimation, the state estimation at time t = 0 is performed using the prior probability determined from the learning result so far, and the prior probability at time t = 0 × The state having the maximum output probability is output, and in the subsequent time steps, the state having the maximum transition probability × output probability is output as a transition sequence. That is, in each time step, each state has a high transition probability as shown by a bold line in FIG. 7A. Of these probabilities, the state having the maximum probability is the transition state in the online state estimation. And As a result, in the online state estimation shown in FIG. 7A, the states 1, 1, 3, 2, and 4 having the maximum existence probability are output as the state transition series.

  In contrast to the online state estimation, in the offline state estimation, when the track is tracked back from time t4, the state 4 having the maximum existence probability at time t4 is shown at the time t3 one step before, as indicated by the bold line in FIG. Step 3 is a transition from Step 3, Step 3 at time t3 is a transition from State 2 at time t2, State 2 at Time t2 is a transition from State 1 at Time t1, and State 2 at Time t2 is State 1 at Time t1 It turns out that it is a transition from. Therefore, in the offline state estimation, as shown in FIG. 7B, among the sequences obtained by the online state estimation, the states at times t2 and t3 are changed by the trackback, and the state of 1 → 1 → 2 → 3 → 4 is changed. A likelihood sequence is output as an estimated state.

[Internal state estimation by forward algorithm]
When the above Viterbi algorithm is strongly intended to be applied online or to improve estimation accuracy, the forward state algorithm is used to estimate the internal state. The forward algorithm is a method of tracing forward on a trellis between an event and each state, and performing forward calculation according to the following steps (3-1) and (3-2), thereby providing a forward probability (state existence probability). α is calculated. In the following equations, π: prior probability, a: state transition probability, b: output signal probability, o: zero vector, and subscript 1 of each variable indicates an initial value.

(3-1) The forward probability is initialized for each state i = 1,.
α ₁ (i) = π ₁ b ₁ (o ₁ )
(3-2) A forward probability is recursively calculated for each time t = 1,..., T−1 and each state j = 1,. Note that Σ in the recursive calculation is the sum for j = 1 to N.
α _{t + 1} (j) = [Σα _t (i) a _ij ] b _j (o _{t + 1} )

  In the state estimation by this forward algorithm, as shown in FIG. 8, all transition probabilities are calculated at each step of time t = 0, 1, 2, 3, 4,. Therefore, the internal state can be estimated with high definition. In the first step of probability calculation, the prior probability obtained from the state transition probability of the HMM is used as the existence probability in each state.

From the state existence probability α for each state obtained by the sequential calculation, an expected value μ is calculated as shown below. The expected value μ is output as a driver internal state in which the state number is precisely expressed by a continuous numerical value instead of a discrete value, and becomes a risk level that the driver is aware of. Note that x is the value of the random variable (here, the state number), and Σ is the sum of i = 1 to N.
μ = Σx _i α _i

[Estimation period]
The transition period of the above HMM is not fixed to one period, but is managed as a period according to a plurality of different time periods and changes in observation symbols in the schedule control unit 9, and various internal states can be estimated. It is said. Specifically, the SOM and HMM are hierarchized to drive each layer at different time periods (time-driven), and the observation symbol in the previous stage of the model (SOM output) is used as an event to drive the HMM according to the state of this event ( (Event driven), there is a method of driving the HMM in accordance with an adaptive sampling period, which is appropriately selected and combined depending on system conditions and the like. Next, each method will be described.

<Time driven>
The feature of time-driven is that it is assumed that the state transitions at a constant period. For this reason, in the time driven, as shown in FIG. 9, the environmental risk discretization unit 3, the operation feature quantity discretization unit 5, the model learning unit 6, and the state estimation unit 7 are hierarchized, and each layer is divided into respective time periods. In parallel.

  FIG. 9 shows an example in which each layer operates in parallel in time periods of 10 s, 3 s, 300 ms, and 100 ms, and the observation state and operation data obtained by discretizing the environmental risk level in each period are discretized by the SOM ( The quantized observation symbols are input to the model learning unit 6 and the model parameters updated in the learning step (learning phase) are used in the estimation step (estimation phase). In the estimation step, as shown in FIG. 9, the internal state of the driver is estimated from the feature quantity of the operation data discretized at the corresponding hierarchy (cycle), and the estimated state is output in parallel from each layer.

  In this case, SOMs may be provided independently for each layer, or the same SOM may be shared by a plurality of layers, and only the sampling periods may be different periods. FIG. 9 shows an example in which common data is input to each layer as the operation data. However, the operation data input by each layer may be different. Further, different operation data may be input to a plurality of layers having the same drive cycle.

  By this hierarchization, it becomes possible to widely estimate state changes with various periods and speeds generated in the driver during driving. For example, in the case of momentary relaxation, learning / estimation is performed with a hierarchy of about 100 ms to 1 s, and the driver gradually becomes unable to cope with the external environment due to a decrease in alertness or accumulation of fatigue. Performs learning / estimation based on a layer of 10 s or a longer period.

  By such hierarchization, it is possible to learn a state transition model at a period corresponding to various state changes, and it is possible to estimate various driver internal states. In addition, it is possible to estimate the internal state in more detail by comprehensively determining the estimated outputs of the plurality of layers obtained later.

<Event driven>
A feature of event-driven is that a driver response to an external environmental risk appears as a driving operation. In other words, it is assumed that there is no state transition that does not appear in the driving operation, and state estimation is performed according to the occurrence of an event. That is, as shown in FIG. 10, the discretization result (observed symbol from the SOM) output in the previous stage is used to monitor changes in driving operation and the internal state of the driver, and a transition trigger is triggered by the occurrence of a vent by the observed symbol. generate.

  For example, as shown in FIG. 11, after the observation symbol a is output as a transition trigger at the change point to the observation symbol a, the observation symbol changes from a to b, and when the observation symbol changes from c to d, d to e. When it changes to, the observation symbol at that time is output to the subsequent stage as a transition trigger. The observation symbol changes when the SOM output changes with a predetermined unit difference (for example, a unit difference of 1 or more).

  Also, when the same discretization result continues for a certain time or more, it is regarded as one event occurrence, and as shown in FIG. 11, when the state of the observation symbol c continues for a predetermined time or more, the observation symbol at that time is output. And re-trigger.

  With this event-driven approach, if there is a change in the driving operation of the driver, the state transition model can be transitioned according to the timing of the change, which is more flexible than when the transition period is fixed. It becomes possible to build an estimator. That is, when the change in the internal state of the driver is slow, the change in the driving operation is considered to be the same, and conversely, when the internal state is frequently changing, the driving operation is frequently changed. Conceivable. Therefore, by making the transition period of the state transition model variable by event-driven, it is possible to adaptively execute the estimation calculation with respect to the change in the internal state of the driver.

  FIG. 12 is a graph showing the state transition probability by time driven and event driven. FIG. 12A is a time driven with a period of 300 ms, FIG. 12B is an event driven when the unit difference of SOM output is 1 or more, FIG. 12C shows the state transition probability due to event driven when the unit difference of the SOM output is 5 or more.

  From FIG. 12, it can be seen that self-transition is relatively decreased in event-driven as compared to time-driven, and in particular, self-transition on the high risk side is decreased. This generally indicates that the high risk is avoided by the driving operation of the driver, and the low risk is self-transitioned by the driving operation.

  FIG. 13 shows an example of internal state estimation by event driving. In FIG. 13, ED is event-driven, TD is time-driven with the transition period fixed at a fixed time, R indicates external environmental risk, H1 and H2 locations in the period from time T2 to T3, and period T3- In the portion of H3 during the period of T4, the hunting of the estimation result seen during TD is suppressed in the ED. This is presumably because the transition period of the internal state estimation is adaptively changed.

  Note that this event driven can be applied to a plurality of hierarchies, and event driven and time driven can be combined depending on the hierarchies.

<Adaptive sampling>
Adaptive sampling is used to prevent significant risks and significant sampling omissions that are problematic in the long-period hierarchy when discriminating environmental risk levels and operation data. By combining, the effectiveness can be further improved.

  That is, as shown in FIG. 14 (a), in the case of a hierarchy with a short cycle, since the sampling interval is short, it is possible to sufficiently capture the risk level and the temporal change of the driving operation. As shown in b), in a hierarchy with a long cycle, since the sampling interval becomes long, the spike-like risk and the driving operation may not be able to sufficiently capture the temporal change. Therefore, as shown in FIG. 14 (c), when a remarkable value is shown between the previous sampling and the current sampling, an adaptive sampling process in which the current sampling value is used as the representative value is added. As a result, the apparent sampling period is changed.

  By adopting this adaptive sampling, even when discretization is performed for an estimator having a long state transition period, it is not necessary to overlook necessary information. It is possible to extend the period up to 2T without overlooking operations with a large absolute value, spike-like risks, and the like. As representative values of adaptive sampling, for example, a maximum value in the case of risk, a maximum absolute value in the case of driving operation, and the like may be considered, and a minimum value, an average value, a mode value, a median value, and the like may be used.

  FIG. 15 shows an example in which hierarchization and adaptive sampling are used together. The layers are the four layers (0th to 3rd layers) shown in FIG. 9 described above. The 0th layer has a transition period of 100 ms, the first layer is 300 ms, the second layer is 3 s, and the third layer is 10 s. The change of the internal state estimation result in the period of time 0-T1'-T2'-T3'-T4 'is shown. The 100 ms and 300 ms layers give important results in evaluating whether the internal state of the driver is responding to short-term changes in the external environmental risk R, and the driver's “look away” and “relaxation”. The output suitable for estimating the state corresponding to is shown. On the other hand, the 3s and 10s layers give important results in evaluating whether the internal state of the driver follows the long-term changes in the external environmental risk R, such as “disappointing” and “misunderstanding”. The output is suitable for evaluating the internal state of the driver that is close to the degree of fatigue and arousal.

[Output of the driver's conscious risk level]
The state number (or expected value) calculated from the probability of existence in each state as an internal state of the driver by the state estimation unit 7 is filtered by a primary low-pass filter (for example, a cutoff frequency of 0.3 Hz), and then alarmed. Output to the support unit 8. This output value corresponds to the risk level that the driver is aware of and can be regarded as the risk level that the driver deals with. By comparing the output value with the environmental risk level value, the vehicle is inherently safe. Or it becomes possible to evaluate whether it is in a dangerous state.

  That is, if the state estimated by the state estimating unit 7 is the same as the state to which the scalar value currently obtained from the environmental risk recognizing unit 2 belongs as a result of the comparison in the alarm / support unit 8, the driver normalizes the environment. It is determined that a normal operation is performed after recognition. On the other hand, when the state estimated by the state estimation unit 7 is different from the state to which the scalar value currently obtained from the environmental risk recognition unit 2 belongs, it is determined that the driver state is inconsistent with the level required by the environment. Is done.

  For example, if the risk level of the estimated state is lower than the risk level obtained from the environmental risk recognition unit 2, the driver's risk recognition level is low, and there is a high possibility that safety will be impaired. Alert / support from Department 8. Conversely, if the estimated risk level of the state is higher than the risk level obtained from the environmental risk recognition unit 2, it is estimated that the driver is overly tensioned, and safety may be similarly impaired. Alarm is given as high probability.

  FIG. 16 shows a part of data measured in the state estimation running test. From S1 street (one side lane) to the S2 street (one lane) through the narrow path (no center line) indicated by the symbols A, B, C, and further from the position indicated by symbol D, S3 street ( Changes in the environmental risk level and the internal state risk level when traveling up to one lane on one side are shown together with changes in the operation feature amount.

  In the test results shown in FIG. 16, the driver is driving with a normal arousal level and a sense of tension based on the relationship between the operation feature quantity and the risk level, and corresponds to a high risk level situation such as an intersection or a narrow road. Can be confirmed from the estimated internal state, that is, the risk level perceived by the driver is high. In particular, the points A, B, C, and D shown in the graph are characteristic high-risk events, mainly intersections and narrow streets. When the other risk level is high, it is mainly a portion where a pedestrian, a bicycle, an oncoming vehicle, or the like is encountered, and it can be seen that the same estimation result is obtained also in this portion.

  As described above, the safe driving support system according to the present embodiment uses the data such as the accelerator opening, the brake pressure, and the steering angle that are mounted on the vehicle as inputs, and the correspondence relationship with the risk level of the driving environment is stochastic. The characteristics of the driving operation of the driver are learned by approximation with the state transition model, and the correspondence between the risk level of the driving environment and the internal state of the driver is evaluated.

  This system does not judge whether alarms, operation assistance, or other interventions are necessary based only on the risk level of the external environment, but rather uses the evaluation index as to whether the driver correctly understands the external environment. This eliminates the discomfort and annoyance caused by unnecessary interventions that occur when judging whether or not a safety system intervention is necessary based only on the risk of the driving environment, and provides appropriate support when necessary. It is possible to construct an advanced safety system adapted to the driver.

  Moreover, in this system, since the driving operation data is used as the main data for state estimation, it is possible to divert the sensors for observing the operation of the steering wheel and accelerator / brake already installed for various vehicle control, There is no need for additional cost as in the case of a measuring device. Since these sensors are relatively robust against noise, it is possible to stably estimate the state, and further, the normal driving operation of the driver is not hindered and the posture is not restrained. Unnecessary interference that adversely affects the estimation can be eliminated.

  In addition, in this system, since the correspondence relationship between the environment outside the vehicle and the driving operation is a normative model, there is no need to determine conditions and templates in advance, and by updating the model through learning, individual differences for each driver and each driver Thus, the state estimation accuracy can be improved by repeated traveling.

  In this case, the learning model that becomes the norm is used as a probabilistic state transition model, and the operation data with variations is statistically processed. Relevance can be learned. In addition, it is possible to estimate which risk level the estimated internal state corresponds to by matching the state transition of the probabilistic state transition model with the transition of the risk level recognized from the external environment at the time of learning.

  In other words, when the risk level corresponding to the estimated internal state is lower than the recognized external environment risk level, the driver is not in a safe driving state even if the external environment risk level is high.・ Operation support can be performed. On the other hand, when the risk level corresponding to the estimated internal state is the same or higher than the recognized external environment risk level, the driver is in a state that can handle the environment even if the external environment risk level is high. Since it can be estimated, useless warnings and unnecessary operation support can be suppressed.

  In addition, when estimating the internal state of the driver, the internal state is estimated at a period such as hierarchization at different time periods, the occurrence of an event, or adaptive sampling, rather than fixing the estimation period based on the state transition model. Thus, it becomes possible to widely estimate various state changes in the period and speed generated in the driver during driving, and it becomes possible to construct an advanced system.

  That is, in the estimation by hierarchization, various state changes can be learned by the state transition model of the corresponding hierarchy, and various driver internal states can be estimated. In addition, it is possible to estimate the internal state in more detail by comprehensively determining the estimated outputs of the plurality of layers obtained later.

  In addition, when the change in the internal state of the driver is slow, the change in the driving operation is also considered to be the same, and conversely, if the internal state changes frequently, a frequent change occurs in the driving operation. As a result, if there is a change in the driving operation of the driver due to event driven, the state transition model can be transitioned according to the timing of the change, and the internal state of the driver can be improved compared to when the transition period is fixed. It is possible to execute the estimation calculation adaptively to the change of the state.

  In addition, adaptive sampling enables learning without missing sampling characteristic input values to be observed when discriminating the risk level and operation data input to the layer of the probabilistic state transition model having a long drive cycle.・ Estimation can be performed.

Basic configuration of safe driving support system Explanatory drawing showing the learning phase and estimation phase of the safe driving support system Explanatory diagram showing the network of operation features and self-organizing maps Explanatory diagram showing stochastic state transition Graph showing state transition probability after learning Graph showing output probability after learning Illustration of state estimation by maximum likelihood sequence Illustration of state estimation by forward algorithm Explanatory diagram showing state estimation by time-driven Explanatory diagram showing state estimation by event driven Explanatory diagram showing the occurrence of transition trigger by event driven Graph showing state transition probability by time-driven and event-driven Graph showing examples of internal state estimation by time-driven and event-driven Illustration of adaptive sampling A graph showing an example of internal state estimation using both time-driven and adaptive sampling Explanatory diagram showing estimation test results

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Safe driving assistance system 2 External environment recognition part 3 Environmental risk discretization part 5 Operation feature discretization part 6 Model learning part 7 State estimation part 8 Warning / support part

Claims (9)

A safe driving support system that estimates the state of a driver driving a moving body and provides support for safe driving,
An environmental risk recognition unit that senses an external environment of the mobile body to extract a feature amount, and recognizes an environmental risk included in the external environment based on the feature amount;
A model learning unit that learns the correspondence between the environmental risk and the driving operation of the driver using a probabilistic state transition model, and constructs a learning model;
Based on the learning model, a state estimation unit that estimates a driver's risk recognition state as a driver's internal state from driving operation data including vehicle behavior data as a reflection of the driving operation of the driver while traveling,
A driving support unit that compares the environmental risk with the internal state of the driver and obtains support information related to safe driving of the mobile body,
Furthermore, a safe driving support system comprising a schedule control unit that sets data input to the stochastic state transition model in a plurality of different periods and adaptively varies an estimation period based on the stochastic state transition model . The schedule control unit
2. The safe driving support system according to claim 1, wherein the stochastic state transition model is driven in a hierarchical manner.   The safe driving support system according to claim 2, wherein the hierarchization is performed according to a type of the driving operation data. The schedule control unit
The safe driving support system according to any one of claims 1 to 3, wherein a transition trigger is output in response to an event occurrence on the model input side, and the stochastic state transition model is driven by the transition trigger.   5. The safe driving support system according to claim 4, wherein when the input data to the stochastic state transition model changes, the transition trigger is output as an event occurrence.   5. The safe driving support system according to claim 4, wherein when the state in which the input data to the stochastic state transition model does not change continues for a certain time or more, the transition trigger is output as an event occurrence. The schedule control unit
The safe driving support system according to any one of claims 1 to 6, wherein a data input period to the probabilistic state transition model is changed by a representative value between samplings.   8. The safe driving support system according to claim 7, wherein a maximum value or a minimum value between samplings is used as the representative value.   8. The safe driving support system according to claim 7, wherein an average value, a mode value, or a median value between samplings is used as the representative value.