DE4211556A1 - Adaptive driver monitoring and warning system for motor vehicle - uses neural nets, image acquisition and processing to determine approach to safety limits in order to generate warnings - Google Patents

Abstract

The method for driver identification, condition monitoring and adaptive warning system employs a number of sensors for data acquisition, including image sensors (3) and processing, where a situation is classified (4), input data is passed to two neural networks (1,2) the outputs of which are processed to time values (7), time value tolerances (5) and driver identification. The spread of actual driver performance figures is used to generate drive warnings for approach of safety limits for a traffic situation (8) and also for the driver condition (9). ADVANTAGE - Automatic support and warning systems to assist driver.

Description

Invention:

Procedure for driver-adaptive, situation-specific modeling of driver behavior in a real driving environment.

Technical field:

Anthropotechnics, mechanical driver support, automatic Vehicle guidance.

State of the art:

Previous modeling approaches were based primarily on models with conventional control technology transmission elements or statistical evaluation. First approaches with neural networks for driver-adaptive drivers Modeling has so far been of a very basic nature (e.g. Kraiss, Küttelwesch), without a description of the application dung in real driving environment. It was the number of situations descriptive parameters and therefore not the trained rules realistic and the network structure for use in real driving environment too easy.

Literature:

• - Johannsen, Boller, Donges, Stein: People in the control loop Oldenbourg, Munich (1977)
• - Donges: A two-level control model of the human steering behavior in motor vehicles Journal of Road Safety 24 (1978)
• - VDI reports 948: The human-machine system in traffic Pp. 211-225 VDI publishing house (1992)
• - Kraiss, Küttelwesch: Identification and Application of Neuronal Operator models in a car driving situation IJCNN 1991

Tasks:

Tasks that involve knowing how the driver is normal behaves, can be solved:

• - Adaptation of a support system for vehicle guidance to the respective driver
  → Compare secondary claim 2: driver-adaptive warning
• - Determine whether the driver is no longer driving normally (e.g. due to fatigue or exposure to alcohol)
  → Compare secondary claim 3: driver status detection
• - Determine which person from a given driver group actually drives the vehicle.
  → Compare secondary claim 4: driver identification.
• - Adaptation of a system for automatic vehicle guidance to the respective driver
  → Compare secondary claim 5: driver-adaptive, automatic vehicle guidance

Commercial usability:

The invention can be used commercially in the motor vehicle industry, where Be Efforts are in place to install the driver accordingly Support sensors and on-board electronics when driving the vehicle Zen.

Beneficial effects:

Accident prevention by timely warning of the driver, driver safety equipment, driver relief.

Execution:

The execution is based on the example of the situation class "lane keeping" explains:

The components and of drawing 1 represent the trained, parallel operated information processing structures (e.g. neural networks) and can be implemented as software modules in a computer. Both components can be trained online ie while driving or offline ie after driving. Both components have the network topology common to all situation classes according to drawing 2 and exist z. B. according to drawing 3 in the situation class "lane keeping" from a given number of input units for the input of information about the state of vehicle motion and the course of the road, a number of hidden units minimized on the basis of the previously defined input units and the output unit, which is the steering angle δ _N (nT) or its scatter indicates.

The required input values describing the situation for the input units (represented by a vector in drawing 1) are provided by conventional sensors, systems for machine perception (e.g. machine vision) and communication as well as a situation classifier (component) of the network group for the respective situation class, that is, in the case of the selected example, networks 1 and 2 are supplied to the situation class "keeping track". (see component of drawing 1). The situation classifier, if necessary, classifies the traffic situations (head 92). Its result is also (not shown in drawing 1) the components and made available.

In the drawings 2 and 3, the hidden units form a sum of the weighted inputs; its output is a non-linear function this sum. The output unit works in the same way as that Hidden unit. In the example of the driving task "keeping track" there is only one output unit that measures the steering angle (component) or the scatter (component). The network pair, consisting from the components and, is with a training record trained, the one of a previous driving section at normal Driving style is taken from the driver. This network pair describes the normal driver behavior. Another pair of networks can be used during who trains on the current driving behavior of the driver the.

The outputs of the com components and are passed to the components and.

Using physical approaches, the component determines a measure of time, the situation-specific mean time reserve T _{res, m} and its scatter Δ T _{res, m} · T _{res, m} indicate how much time is available on average in the respective situation to prevent accidents.

Component identifies the driver by comparing the Output of component and with the actual Fahrak tions, e.g. B. when "keeping track" with the actually measured steering angle δ.

The component is the current driving action - at "lane keeping" the steering angle δ - supplied. Using physical approaches, this determines the actual time reserve T _res from the given measured variables.

The outputs of the components and become the components and fed.

Component compares the current time reserve T _{res, is} with the situation-specific mean time reserve T _{res, m} and its spread Δ T _{res, m} and controls a delivery of goods to the driver if T _{res is} outside a predetermined range of scatter by T _{res, m} .

The component determines whether the driver behavior differs from the normal driver behavior either by observing the relative frequency of the deviations of the actual time reserve T _res, from a set barrier or by comparison with networks trained on the current driver behavior. Component 9 controls a goods issue if such a deviation (longer time) is found.

The warning output is of an acoustic, visual or haptic nature.

Literature:

(Head 92) Head, Onken: DAISY, a Khowledgeable Monitoring and Maintenance Aid for the Driver on German Motorways IFAC, The Hague (1992).

Claims (6)

Generic term:
Adaptive, situation-specific driver modeling of high quality in real driving environments characterized by

• - Through a multiple corresponding to the considered classes of traffic situations of parallel working, specially trained information processing structures, which are called neural networks in the form of representation used here.
• - through the adaptivity of driver modeling on the basis of off-line training and on-line training of the neural networks.
• - Through a realistic image of the information available to the driver while driving
• a) conventional sensors
• b) Image sensors with image processing
• c) Communication with other vehicles or communication partners external to the vehicle.
• - Through situation-specific driver modeling, ie by taking into account all conceivable situation characteristics of the considered classes of traffic situations.
• - By high quality due to the simultaneous application of the features described in the preceding NEN markings.
• - By applicability of the method to the generation of driver-adaptive warnings, driver status detection, driver identification and automatic vehicle control.

1. Main claim
Driver-adaptive modeling of the driver with several neural networks, at least two networks being used for each situation characteristic of a respective situation class and of these

• a network (in drawing 1) generates a situation-specific mean value of the driver action,
• - A second network (drawing 1) provides a measure of the situation-specific dispersion of the driver action.

Additional claims: 2. Driver adaptive warning (in drawing 1):
This process is based on:

• - Use of the driver-adaptive driver model (main claim)
• - Determination of a hazard level via physical and physiological parameters.
• - Comparison of the current hazard measure with the average hazard that can be derived from the adaptive driver model.
• - Warning of the driver when a hazard threshold value is exceeded.

3. Driver condition detection (in drawing 1):
This process is based on:

• - Use of the driver-adaptive driver model (main claim)
• - A statistical determination of the frequency of the deviations from the normal range of driver actions, which is determined from the adaptive driver model and an evaluation of these frequencies or comparison of two trained driver-adaptive driver models (main claim), one pre-trained for the normal behavior of the driver and a second permanently trained while driving for the current driving behavior of the driver.
• - a warning to the driver when a frequency threshold value is exceeded.

4. Driver identification (in drawing 1)
This procedure is based on

• - Use of driver-adaptive driver models (main claim)
• - Identification of the driver by comparing the driver-adaptive driver models of the drivers in question with the current driver actions.

5. Driver adaptive, automatic vehicle guidance
This process is based on:

• - Use of the driver-adaptive driver model (main claim) for automatic guidance of the vehicle instead of previously used control algorithms.