JPH07306717A - Autonomous-running vehicle - Google Patents

Abstract

PURPOSE:To make a desired autonomous running coping with changes in the conditions of a travel road without preparations such as previous learning. CONSTITUTION:A driving means such as motors 6 and 24 and an engine and a control means 8 which is equipped with a television camera 9 for photographing travel environment and a neural network 17 having real-time learning ability and controls the operation of the driving means is mounted on the vehicle body; and an image photographed by the television monitor 9 is processed by the neural network 17 to control the operation of the driving means.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention is capable of automatic driving by utilizing a neural network, and can be used for industrial, agricultural, hospital,
The present invention relates to self-supporting vehicles that can be used in welfare or leisure facilities and other areas.

[0002]

2. Description of the Related Art Conventionally, as a self-supporting vehicle for transporting luggage in a factory, for example, an aluminum tape serving as a traveling line is attached to a floor surface, and the aluminum tape is detected by a photo sensor attached to the lower portion of the vehicle body. It is known that the vehicle travels while traveling, or the vehicle travels while detecting the distance to or from a specific object in the traveling direction or on the side by ultrasonic waves or electromagnetic waves. In addition, a CCD camera captures an image of aluminum tape or the like, performs image processing, calculates the position of this aluminum tape, and controls the traveling direction of the vehicle body so that the position of the aluminum tape is always a fixed position of the camera. It has been known.

However, such a photo sensor or CC
In a conventional self-driving vehicle using a D camera, it is necessary to dispose a detection target object such as an aluminum tape along the traveling path. Further, a self-sustaining vehicle that travels using ultrasonic waves, electromagnetic waves, or the like needs to have a specific object that can be detected on the front or side of the traveling direction. Therefore, there is a problem that it is difficult to drive an unspecified place based on a certain rule.

In view of the above, Japanese Patent Application No. 172476/1993 proposes a method in which the image output of a television camera mounted on the vehicle is judged based on the data of the traveling path learned by the neural network and the vehicle is properly driven. It has been proposed by the present inventor as an issue. That is, in a self-sustaining vehicle equipped with a driving unit including a motor and an engine, a control unit for the driving unit, and a TV camera, a neural network learned in advance is incorporated in the control unit.
The image taken by the television camera is arithmetically processed by a neural network to control the driving means to control the traveling. According to this, the self-supporting traveling vehicle can travel in an unspecified place without disposing the detection target object on the traveling path.

[0005]

However, in the case of this proposed example, it is necessary to make the neural network beforehand learn the condition of the traveling road and the way of traveling using the data of the traveling road before the autonomous traveling vehicle is driven. There is. For this reason, road data must be prepared for learning before running on the road. It is not possible to swiftly respond to new situations that occur during driving. There is an inconvenience.

[0006]

According to a first aspect of the present invention, there is provided a self-sustaining vehicle including a driving means such as a motor and an engine, a television camera for photographing a traveling environment, and a neural network having a real-time learning ability. A control means for controlling the operation of the means is mounted on the vehicle body, and the image taken by the television camera is processed by the neural network to control the operation of the driving means.

A self-sustaining vehicle according to a second aspect of the present invention includes a driving means including a motor and a motor driving circuit, a CCD camera for photographing a traveling environment, and an image processing means for compressing a video output of the CCD camera and binarizing it. Of the motor drive circuit, comprising a feature quantity calculating means for calculating the feature quantity of the video output based on the output of the image processing means and a hardware neural network with a multilayer perceptron having a real-time learning ability by the error backpropagation learning method. A control means for controlling the operation is mounted on the vehicle body, and the characteristic amount calculated based on the image captured by the CCD camera is input to the hardware neural network to determine the state of the traveling environment. Based on the output, the operation of the motor drive circuit is controlled.

A self-supporting vehicle according to a third aspect of the present invention is the self-supporting vehicle according to the second aspect, wherein the hardware neural network is a pulse density logical operation type hardware neural network.

According to another aspect of the present invention, there is provided a self-sustaining vehicle having a fuzzy controller in the control means of the self-sustaining vehicle according to the second or third aspect, which outputs a control signal to a motor drive circuit using an output of a hardware neural network as an input. It is provided.

[0010]

In the self-supporting vehicle according to claim 1, the neural network is driven by inputting the data obtained as a result of numerical processing of the image obtained by photographing the image obtained by the television camera mounted on the vehicle body. Train the road situation. Also, in parallel with this, or
After the learning is completed, the data obtained through the television camera is input to the neural network to determine the condition of the traveling road, and the traveling is controlled by the output of the neural network. As a result, it is possible to run independently even in an unspecified place where the detection target is not placed.
Since this neural network has a real-time learning capability, learning and running control can be performed simultaneously in parallel. Therefore, the learning can be performed while actually traveling on the traveling path without the preliminary preparation of learning in advance. That is, the desired self-sustaining running can be performed only by gaining the running experience in the actual running scene. Furthermore, even if the running environment such as the shape and color of the road, shoulders, walls, etc. changes, it is possible to drive appropriately while adapting to the changed environment. Therefore, for example, in the case of repeatedly making a round trip between the same two points, it becomes possible to make a round trip in a gradually shorter time because the traveling becomes smoother. Alternatively, even when the obstacle frequently occurs, it is possible to quickly avoid the obstacle. Further, since such a neural network is mounted on the vehicle body, the traveling range is not restricted as compared with the one using an external neural network connected by a signal line or the like without being mounted.

In the self-supporting vehicle according to the second aspect, in addition to such an action, the characteristic amount calculating means calculates the characteristic amount based on the image output of the CCD camera and inputs it to the hardware neural network. The load on the neural network is reduced, learning of the neural network is facilitated, and the scale of the neural network is reduced. Also, since the hardware neural network is based on the multilayer perceptron, it operates faster than other types of neural networks such as Boltzmann machines.

In the self-driving vehicle according to the third aspect, since the pulse neural network of the pulse density logical operation type which is lightweight and can learn at high speed is used, the traveling speed can be kept high.

In the self-supporting vehicle according to the fourth aspect, since the fuzzy controller is integrated with the neural network, the control capability is enhanced.

[0014]

DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described with reference to the drawings. First, the schematic configuration of the self-supporting vehicle 1 according to the present embodiment will be described with reference to FIG. The self-supporting vehicle 1 is based on a vehicle body 4 provided with a pair of driving wheels 2 and driven wheels 3 at a lower rear end and a lower front end, and a vehicle body panel 5 is mounted on the vehicle body 4. Has been. The shape of the vehicle body panel 5 is a box shape in the illustrated example, but the shape is not limited to this, and may be another shape, for example, a half-split egg shape, depending on the environment in which the self-supporting traveling vehicle 1 travels.

A drive motor 6 for rotating the drive wheels 2 is mounted on the vehicle body 4. Further, on the vehicle body 4, a power source 7 by a battery, a control device (control means) 8 including a computer and a neural network having a real-time learning ability described later, a television camera, here a CCD camera 9 with illumination and this CC
An image processing device (image processing means) 10 for processing a video output obtained by the D camera 9 is mounted. The CCD camera 9 is for taking a picture of a front situation (running environment), and is mounted on a driving device 11 which can be vertically and horizontally swung and vertically moved, and a sound collecting microphone 12 is provided below the driving device 11.

The control system block diagram of the electric system of the self-sustaining vehicle 1 will be described with reference to FIG. First, CCD
The image processing apparatus 10 that processes an external video signal obtained by shooting with the camera 9 includes, for example, an image processing unit 13 that appropriately performs image processing on the video signal, an image compression unit 14 that appropriately performs compression processing, It is configured by a binarization unit 15 that performs binarization processing.

As a result, the red, green, and blue 3 pixels, which are, for example, 256 × 256 pixels and are captured by the CCD camera 9.
The brightness of each video signal is divided into 64 levels and stored. The figure is moved when necessary, but in the case of the traveling control of the self-supporting vehicle 1, the CCD camera 9 is fixed and judgment is made from the overall position of the figure, so the figure movement is not always necessary. Further, if the information (video signal) of the CCD camera 9 is processed as it is, the memory capacity of the ordinary personal computer forming the control device 8 becomes insufficient. Therefore, image compression (for example, 32 × 32 pixels) may be performed if necessary. Then, the thresholds of the three video signals of red, green, and blue may be selected and used as processed signals.

By the image processing device 10 as described above,
A feature amount calculation unit (feature amount calculation means) 16 is provided in the input unit of the control device 8 that receives a binarized signal. The feature amount calculated by the feature amount calculation unit 16 is, for example, a binarized image itself or a moment invariant amount (inertia moment, position of the center of gravity, screen 2, 4, or 4) of the image obtained by calculation by a computer. It is the moment of inertia of each or the position of the center of gravity when divided into 9 or more). A neural network 17 having a real-time (real-time) learning ability, which is a main part of the control device 8, is provided with the feature amount calculated by the feature amount calculating portion 16 and the output signal from the image processing device 10 as inputs. . That is, the function required for the neural network 17 is not required to slow down the traveling speed or stop the traveling for learning, and has a sufficiently high-speed learning function, which can be mounted on the vehicle body 4 of the self-supporting vehicle 1. It is small and can operate standalone.

As a neural network satisfying such requirements, for example, the applicant of the present invention has disclosed Japanese Unexamined Patent Publication No. 4-549.
Japanese Patent Laid-Open No. 4-111185 and Japanese Patent Laid-Open No. 5-2
The neural network already proposed by 10650 gazette and others may be used. That is, as one of the digital types, it is a pulse density type neural network with a self-learning function in which a signal is expressed by a pulse train, and JP-A-4-549 discloses the basics of the forward process. No. 111185 gazette shows the basics of the learning process. More specifically, VL
A pulse density logical operation type hardware neural network consisting of SI neurons.
A multilayer perceptron having one or a plurality of intermediate layers and an output layer and having a learning function by an error backpropagation method (backpropagation method). Such a multi-layer perceptron has a feature that it can be operated at a higher speed than other types of neural networks with learning ability such as Boltzmann machine. Particularly, according to the pulse density logical operation type hardware neural network, since it is lightweight and can perform learning at high speed, it becomes possible to control while maintaining the traveling speed of the self-supporting vehicle 1 at high speed.

Here, although learning of the neural network 17 by the error back propagation method is well known, it will be briefly described. First, when an input signal is given to the neural network 17, the neural network 17 sequentially processes the signal from the input layer through the intermediate layer to the output layer and outputs the output signal. This is called the "forward process" of the neural network. Next, the neural network 17 generates an error signal from this output signal and the given teacher signal, and conversely, while propagating this error signal from the output layer through the intermediate layer to the input layer, the synapse part is generated. Change the coupling coefficient of. This is called the "backward process" of the neural network.
One learning process is configured by a combination of one forward process and one backward process.
When the value of the coupling coefficient changes to an appropriate value by repeatedly executing this learning process, an output signal equal to the teacher signal is obtained from the neural network 17, and the learning of the neural network 17 ends. . It is also possible to execute only the forward process without the backward process.

After causing the neural network 17 to perform the learning process on the input signal n times,
A discriminating and processing unit 18 is provided which receives the output signal obtained by the forward process of the neural network 17 for the subsequent m input signals. That is, while the autonomous vehicle 1 is traveling, the learning process n times and the forward process m times are repeated. here,
n and m are 0 or an arbitrary integer, and may be changed each time the learning process and the forward process are repeated, or may be maintained at the same value. If n is set to a sufficiently large number to complete learning, the self-supporting vehicle 1 travels while surely learning the traveling environment in each scene of the traveling process. If n is not a large enough number to complete the learning, the learning is not completed during one iteration of the learning process and the forward process, but the learning is completed within several iterations.

By the way, the learning process is not separated from the execution process of the output signal of the forward process, for example, the forward process is executed m times after the learning process of n times. The output signal of the neural network 17 at the time of execution may be output to the identification and processing unit 18. That is, the learning of the neural network 17 and the control of the autonomous traveling vehicle 1 by the neural network 17 may be simultaneously executed in parallel. In this case, the self-sustaining vehicle initially travels improperly, such as hitting a wall, but learning progresses while the vehicle is traveling, gradually changing to an appropriate traveling, and finally the given traveling environment. You will be able to drive properly inside.

In the identification and processing section 18, the output signal of the neural network 17 is identified, the identification result is compared with the reference data 19 to perform classification, and the output is provided in the output stage of the controller 8. Fuzzy controller 2
It is given to 0.

On the other hand, a voice discriminating circuit 21 to which a voice signal from the sound collecting microphone 12 mounted on a part of the vehicle body 4 is inputted is provided. The voice identification circuit 21 converts a voice signal into a predetermined command signal and inputs it into a memory (not shown) in the identification and processing section 18. As a result, the autonomous vehicle 1 stops, turns right,
It is supposed to be able to turn left, accelerate, and decelerate.

The fuzzy controller 20 to which the output signal from the identification and processing unit 18 is input has a ROM programmed in advance, and the processed signal (binarized signal, image signal, etc.) of the neural network 17 is inputted. (Including) is calculated, an output corresponding to each input is calculated, and a control output is given to the drive motor drive circuit 22 and the steering motor drive circuit 23. The drive motor drive circuit 22 drives the drive motor 6, and the steering motor drive circuit 23 drives the steering motor 24 mounted to steer the driven wheels 3. These drive circuits The driving means is composed of 22, 23 and the motors 6, 24 (of course,
If it is an engine car, the driving means is composed of the engine). These motors 6 and 24 are provided with rotary encoders 25 and 26, respectively, and compare with the reference rotation angles 27 and 28 that are preset and stored, the rotation state (rotation speed, rotation speed) and the driven state of the drive wheel 2. The steering status of the wheels 3 can be fed back to the fuzzy controller 20. By fusing the neural network 17 with the fuzzy controller 20 in this manner, controllability is enhanced. An example of such a fuzzy controller 20 is, for example, from June 3 to 5, 1992, in Istanbul, TURKIYE, "" Fuzzy Control of Mo
bil Vehicle ”IFACWORKSHOP ON AUTOMATIC CONTROL FOR
QUALITY AND PRODUCTIVITY, ACQP'92, Preprints, Vo
As for “l.2”, those for which detailed examples have been announced may be used.

Therefore, according to the self-supporting vehicle 1 of the present embodiment, the condition of the traveling road photographed by the CCD camera 9 can be recognized and judged by the neural network 17, so that an unspecified place where the object to be detected is not arranged in advance. Even then, it becomes possible to run independently. Further, since the neural network 17 has a real-time learning capability, learning and running control can be performed in parallel at the same time, or learning is performed while actually running on a running path without preparation for learning in advance. You can also let it. That is, the desired self-sustaining traveling can be performed only by placing the self-sustaining traveling vehicle 1 in the actual traveling scene and gaining traveling experience. Furthermore, even if the traveling environment (for example, the shape or color of the road, shoulders, walls, etc.) changes, it is possible to drive appropriately while adapting to the changed environment. Further, since such a neural network 17 is mounted on the vehicle body 4, there is no restriction on the traveling range of the self-supporting vehicle 1 as compared with a case where an external neural network connected by a signal line or the like is used without being mounted. Will be things.

In this embodiment, the control device 8 has the neural network 17 and the fuzzy controller 20 integrated, but instead of the fuzzy controller 20, a normal PID controller is used to send the control signal to the motor drive circuit 22. , 23 may be given. In this case, the processed signal by the neural network 17 is
According to a preset program, it changes into a motor drive signal, this is given to the motor drive circuits 22 and 23, and the corresponding motors 6 and 24 are controlled, and the traveling of the self-supporting vehicle 1 is controlled.

Further, the control device 8 of the present embodiment is provided with an abnormality / fault identifying circuit 29 which receives the feature quantity calculated by the feature quantity calculating section 16 as an input, and the identification output thereof is the identification and processing. It is configured to be provided to the section 18. The abnormality / fault identifying circuit 29 has a built-in neural network (not shown) different from the neural network 17. That is, it is a hierarchical structure that has an input layer, an intermediate layer, and an output layer, and has already been learned. It has the function of diagnosing abnormalities and failures of specific objects. Here, when the destination is reached and transferred to another carrier, the shape of a specific object such as the carrier is also identified in advance by a neural network, and when the self-supporting vehicle 1 carries a package or the like. It is also possible to allow unmanned luggage transfer.

Diagnosis of the shape of such a specific object, a situation such as an abnormality or a failure is performed by a preprogrammed command or a command input through the sound collecting microphone 12. In this case, when diagnosing the shape, abnormality or failure of the object, the CCD camera 9 needs to face the object to be diagnosed. Therefore, a driving device provided for raising and lowering and turning the CCD camera 9 11 is configured to track an object according to pre-programmed instructions. Even when tracking such an object, the shape of the object, for example, using a neural network,
It is also possible to learn features such as patterns and colors, and use the learned neural network to judge the degree of abnormality or failure of the object. Alternatively, for the tracking of the object, the normal PID control may be used, or a combination of the neural network and the fuzzy control (for example, Japanese Patent Laid-Open No. 5-233 proposed by the present inventor).
No. 062 (see “moving pair identification tracking system”).

As described above, similar to the above-mentioned Japanese Patent Application No. 5-172476, the means for identifying the shape, abnormality and failure of the specific object mounted on the self-supporting vehicle 1 is constructed. That is, when the self-supporting traveling vehicle 1 reaches the destination and transfers the load mounted on the self-supporting traveling vehicle 1 to another transfer device, it is known which transfer device should be transferred, or
This is because it is necessary to know whether or not there is a failure in the transport device, and the image input of the specific object on the traveling path is converted into a numerical process, and the neural network learned in advance is used to input the specific object. It is an identification means for judging the shape, abnormality and failure. This makes it possible to select an object,
It is possible to identify a failure or an abnormal portion at high speed, and it is possible to deal with a connection work with a specific object or an unexpected accident.

Next, a specific example in which actual traveling is assumed using the self-supporting traveling vehicle 1 configured as described above will be described with reference to FIG. When the self-supporting vehicle 1 travels on the road 31, for example, FIG. 3A to FIG. 3A show the state of image output captured by the CCD camera 9 mounted in the front center of the vehicle body 4 facing forward. As shown in (g). In the figure, 32 indicates a shoulder line. Road 31 in the case of FIG.
Is a straight line, and in the case of FIG. 7B, it indicates that there is an intersection in front. When the self-supporting vehicle 1 is driven on such a straight road and on a road having an intersection and the image captured by the CCD camera 9 is shown in FIG.
The neural network 17 is made to learn so that the output that is determined to be 1 is linear, and the output that is determined to be an intersection ahead in the case of FIG. For example, in the former case (straight line), the first neuron in the output layer of the neural network 17 is given a teacher signal with a value of 1.0, and the remaining neurons in the output layer are given a teacher signal with a value of 0.0. In the latter case (intersection), the second neuron in the output layer of the neural network 17 is given a teacher signal with a value of 1.0, and the remaining neurons in the output layer are given a teacher signal with a value of 0.0. . As a result, in the former case, learning is performed so that the first neuron in the output layer of the neural network 17 outputs the value 1.0 and the remaining neurons in the output layer output the value 0.0. Trains the second neuron in the output layer of the neural network 17 to output a value of 1.0 and the remaining neurons in the output layer to output a value of 0.0. Then, for example, when it is determined that there is an intersection, a program input in advance or a command signal from the sound collecting microphone 12 causes
After determining the traveling direction, the vehicle goes straight on or decelerates and turns left or right.

Here, if straight driving is to be performed, the steering angle of the driven wheel 3 is calculated so that the center position of the CCD camera 9 is at a predetermined position on the road 31, and this is used as a steering drive mechanism. , Here, it outputs to the steering motor drive circuit 23. At this time, the control used may be fuzzy control by the fuzzy controller 20 or normal PID control.

In the case of right turn or left turn,
The image output by the CCD camera 9 is shown in FIGS.
The state of the road shoulder line 32 gradually changes. Then, when the right turn traveling is about to end as shown in (e) of the figure, or when the left turn traveling is about to be completed as shown in (f) of the figure, this is recognized and straight traveling is performed. Start preparing. Even in this case,
The steering angle of the driven wheel 3 and the drive wheel 2 are set so that the center position of the CCD camera 9 is at a predetermined position on the road 31 (this predetermined position may be changed depending on the learned image).
A right turn or a left turn is smoothly completed by controlling the rotation speed of the.

That is, while the pattern obtained by the image processing is as shown in FIG. 7A, the vehicle travels straight ahead, and when the pattern as shown in FIG. Therefore, the speed is reduced to prepare for a right turn or a left turn, and when the pattern shown in (c) or (d) of the figure is obtained, the right turn or the left turn is continued, and the figure (e).
Alternatively, at the stage where the pattern as shown in (f) is obtained, the right turn or the left turn is completed, so preparation for straight running is made. Since each pattern can be learned, it is possible to drive at an intersection similar to such a pattern. It is understood that learning is applicable not only to such an intersection but also to a three-way road, a five-way road, or an obstacle on the road.

Further, when the vehicle travels on a curved road, the screen as shown in FIG. 9 (g) is displayed, but by learning the built-in neural network 17 so that it is judged to be a curved road screen, the screen shown in FIG. In the case of the screen as shown in FIG. 8), it is determined that the screen is a curved traveling screen, and the rotational speed of the drive wheel 2 is set so that the center position of the CCD camera 9 comes to a predetermined position P (for example, the center) of the screen of the road 31. And the direction of the driven wheel 3 may be calculated using fuzzy control or PID control, and control signals may be output to the motor drive circuits 22 and 23.

Thereafter, the same procedure will be repeated.
In this case, the object to be learned by the neural network 17 is the shape of the entire road 31 and its characteristic amount visible in front of the CCD camera 9. The shape of the white line or the yellow line on the road 31 and others (for example, traffic signal signs, road 31 Color, C
The width of the road 31 that can be seen by the CD camera 9).

A control example when the self-supporting vehicle 1 of this embodiment is applied to a vacuum cleaner or a mowing machine will be described with reference to FIG. 4 (a). That is, this is an example of application to a device that travels in a certain specific area. First, the target object 3 consisting of walls, etc.
3 is photographed by the CCD camera 9 from the position where the self-driving vehicle 1 is currently stopped, and while controlling the image output data thereof, control is performed so as to reach the target 33 by straight traveling, and when the target 33 is reached. A U-turn operation is performed, and a straight running is performed toward a target object learned in advance. At this time, instead of controlling the autonomous vehicle 1 using the actual output signal of the neural network 17,
The self-supporting vehicle 1 is controlled by giving a desired output signal of the neural network 17 from the outside, and at the same time, the desired output signal is given to the neural network 17 as a teacher signal. By repeating such straight running and U-turn operation many times, the neural network 17 can be made to learn this operation. When the learning is completed, the output signal of the actual neural network 17 is used to switch to control the self-sustaining vehicle 1, so that the self-sustaining vehicle 1 travels independently as indicated by the arrow locus in FIG. become.

In this case, the desired output signal of the neural network 17 is generated by a computer mounted on the self-supporting vehicle 1 and given to the neural network 17, so that the neural network 17 can always be learned. For example, even in an environment in which the shape and color of the target object 33 such as a wall changes daily, the self-sustaining vehicle 1 can adapt to the change and maintain desired travel.

Further, as shown in FIG.
When 4 is a golf ball or the like, the neural network 17 learns the image of the target object 34, so that the autonomous traveling vehicle 1 travels so as to approach the target object 34.

Further, another application example will be described. For example,
Another vehicle is run in front of the self-supporting vehicle 1, this vehicle is photographed by the CCD camera 9, and the image output data and the distance from the previous vehicle when the previous vehicle tries to turn right or left are displayed. By letting the neural network 17 learn an image such as a size and controlling the traveling by using the neural network 17 learned in this way, the self-supporting vehicle 1 always keeps a certain distance from the vehicle in front of the vehicle. It becomes possible.

[0041]

According to the self-sustaining vehicle of the first aspect of the invention, the self-driving vehicle is provided with a driving means such as a motor and an engine, a television camera for photographing the traveling environment, and a neural network having a real-time learning ability. Since the control means for controlling the operation of the means is mounted on the vehicle body and the image taken by the television camera is processed by the neural network to control the operation of the driving means, the image is taken by the television camera. Since the condition of the road can be recognized and judged by a neural network, it can be driven independently even in an unspecified place where the object to be detected is not arranged in advance, and this neural network has a real-time learning capability. Therefore, learning and running control can be performed in parallel at the same time, and there is no preparation to learn in advance, In addition, you can learn while traveling on the road, and even if the traveling environment such as the shape and color of the road, shoulders, walls, etc. changes, you can appropriately drive while adapting to the changed environment Also, since such a neural network is mounted on the vehicle body, the traveling range is not restricted as compared with the one using an external neural network connected by a signal line without mounting the neural network. be able to.

According to the self-supporting vehicle of the second aspect of the present invention, the driving means including the motor and the motor drive circuit, the CCD camera for photographing the traveling environment, and the image output of the CCD camera are compressed and binarized. The image processing means, the feature amount calculating means for calculating the feature amount of the video output based on the output of the image processing means, and the hardware neural network by the multilayer perceptron having the real-time learning ability by the error backpropagation learning method are provided. The control means for controlling the operation of the drive circuit is mounted on the vehicle body,
Since the characteristic amount calculated based on the image captured by the CD camera is input to the hardware neural network to determine the state of the traveling environment and the output based on the determination result controls the operation of the drive circuit. In addition to the effect of the self-supporting vehicle of the invention according to claim 1, CC
Since the characteristic amount is calculated by the characteristic amount calculation means based on the image output of the D camera and is input to the hardware neural network, the load on the neural network can be reduced and the learning of the neural network can be facilitated. In addition, the scale of the neural network can be reduced, and since the hardware neural network is based on the multilayer perceptron, it can be operated at a higher speed than other types of neural networks such as the Boltzmann machine.

According to the autonomous vehicle of the invention described in claim 3, since the hardware neural network in the autonomous vehicle of claim 2 is a pulse density logical operation type hardware neural network, it is lightweight. By taking advantage of the features that enable high-speed learning, it is possible to control while maintaining a high running speed.

According to the self-sustaining vehicle of the fourth aspect of the invention, the control means of the self-sustaining vehicle of the second or third aspect of the invention uses the output of the hardware neural network as an input to the motor drive circuit as a control signal. Since the fuzzy controller for outputting is provided, the control capability can be further enhanced.

[Brief description of drawings]

FIG. 1 is a block diagram of a control system configuration showing an embodiment of the present invention.

FIG. 2 is a schematic configuration diagram showing an autonomous traveling vehicle.

FIG. 3 is an operation explanatory diagram showing each example of a specific video pattern assuming an actual traveling example.

FIG. 4 is an operation explanatory diagram of an application example of the self-sustaining vehicle.

[Explanation of symbols]

6 motor 8 control means 9 CCD camera (television camera) 10 image processing means 16 feature amount calculation means 17 neural network with real-time learning ability 20 fuzzy controller 22, 23 motor drive circuit 24 motor

Claims (4)

[Claims] 1. A vehicle body equipped with driving means such as a motor and an engine, a television camera for photographing a traveling environment, and a control means provided with a neural network having a real-time learning ability to control the operation of the driving means. The self-supporting vehicle is characterized in that the image taken by the television camera is processed by the neural network to control the operation of the driving means. 2. A driving means including a motor and a motor driving circuit, a CCD camera for photographing a traveling environment, and the CCD.
Image processing means for compressing and binarizing the video output of the camera, feature quantity calculating means for calculating the feature quantity of the video output based on the output of the image processing means, and real-time learning ability by the error back propagation learning method. A control means for controlling the operation of the motor drive circuit, which comprises a hardware neural network having a multi-layered perceptron, is mounted on the vehicle body, and the feature amount calculated based on the image taken by the CCD camera is used as the hardware neural network. A self-sustaining vehicle, characterized in that a state of a traveling environment is input to a network and the operation of the motor drive circuit is controlled by an output based on the result of the determination. 3. The self-supporting vehicle according to claim 2, wherein the hardware neural network is a pulse density logical operation type hardware neural network. 4. The self-sustaining vehicle according to claim 2 or 3, wherein the control means is provided with a fuzzy controller which outputs a control signal to a motor drive circuit using an output of a hardware neural network as an input.