CN111967434B

CN111967434B - Machine vision anti-pinch system based on deep learning

Info

Publication number: CN111967434B
Application number: CN202010895766.6A
Authority: CN
Inventors: 丁玲; 宗加飞; 廖海斌; 徐斌
Original assignee: Hubei University of Science and Technology
Current assignee: Hubei University of Science and Technology
Priority date: 2020-08-31
Filing date: 2020-08-31
Publication date: 2023-04-07
Anticipated expiration: 2040-08-31
Also published as: CN111967434A

Abstract

The invention provides a machine vision anti-pinch system based on deep learning, and belongs to the technical field of intelligent control. A convolutional neural network CNN is used for building a learning table closing process into a convolutional neural network model, and the identification accuracy is improved. The system of this paper research makes full use of above-mentioned characteristics, according to the appearance of the closed department barrier of video image analysis study platform that acquires, automatic closing action of correct control to reach and prevent pressing from both sides the purpose, thereby effectively overcome the not enough of traditional anti-pinch mode. The invention has the advantages of high precision, no contact and the like.

Description

Machine vision anti-pinch system based on deep learning

Technical Field

The invention belongs to the technical field of intelligent control, and relates to a machine vision anti-pinch system based on deep learning.

Background

At present, a shielding door is arranged between a waiting area and a rail running area of a subway platform to ensure the safety of passengers when waiting. However, a certain gap is formed between the train and the shield door, and when the train runs, the situations of passenger clamping, article dropping, bag clamping, clothes clamping and the like easily occur in the gap, so that great potential safety hazards and accident risks exist. Aiming at safety detection between a subway train and a shield door, the conventional mode is mainly manual observation, but the mode cannot accurately detect overlong linear platforms and curve platforms. The automatic detection method mainly adopts laser correlation and infrared light curtain for detection, but the laser correlation detection and the infrared light curtain detection belong to point and surface detectors, and can cause higher false alarm rate due to vibration. Meanwhile, for large-gap platforms, huge detection blind areas exist.

In addition, high-grade automobiles in the market are also provided with door and window anti-pinch systems, the systems are controlled by touch, once the control system detects that the glass is stressed in the glass lifting process, the control system stops the glass lifting and instructs the glass to descend, but at the moment, a pinch injury event is generated, only the intensity of the pinch injury is not too large, and psychological injury can be caused to a clamped person.

In summary, the anti-pinch systems existing on the market at present generally have the following problems:

1. has higher misjudgment rate,

2. Has huge dead zone for detection,

3. The safety is poor.

Disclosure of Invention

The invention aims to provide a machine vision anti-pinch system based on deep learning aiming at the problems in the prior art, and the technical problem to be solved by the invention is how to provide an anti-pinch judgment system in a deep learning mode.

The purpose of the invention can be realized by the following technical scheme: a machine vision anti-pinch system based on deep learning is characterized by comprising a Convolutional Neural Network (CNN), a data processing unit and a video image analysis learning platform, wherein the video image analysis learning platform comprises a closing driving unit, a camera, a substrate and a cover plate hinged on the substrate, and the closing driving unit can drive the cover plate to be close to or far away from the substrate around a hinged point;

the first step is as follows: collecting training samples

Setting the maximum opening angle between the substrate and the cover plate as A, taking 1-degree span between opening angles A-1 as an acquisition node, and enabling the convolutional neural network CNN to acquire a plurality of pictures at each acquisition node through a camera and store the pictures in an (A-X) sample, wherein X in the first acquisition node is A-1, and X in the last acquisition node is 0;

the second step: testing of

Selecting any acquisition node, placing a certain barrier corresponding to the acquired picture between a plate and a cover plate, taking a picture of the barrier by a camera, if the data processing unit can match the acquired picture information with the corresponding sample and control the closed driving unit to stop, indicating that the acquisition of the training sample is reliable, otherwise, correcting by adding a picture library;

the third step: and (4) adapting the visual anti-pinch system after deep learning to an application scene.

The system based on machine vision has the advantages of high speed, high precision, non-contact and the like, and a convolutional neural network CNN is used for building a convolutional neural network model by the closing process of a learning table, so that the identification accuracy is improved. The system of this paper research makes full use of above-mentioned characteristics, according to the appearance of the closed department barrier of video image analysis study platform that acquires, automatic closing action of correct control to reach and prevent pressing from both sides the purpose, thereby effectively overcome the not enough of traditional anti-pinch mode. The anti-pinch system can not only protect the safety of people and prevent users from being damaged by clumsy machines, but also protect equipment and prevent the equipment from meeting obstacles and being damaged by the obstacles in the closing process.

This prevent pressing from both sides system has following characteristics:

1. the machine vision technology is developed increasingly, and the system based on the machine vision is faster;

2. the requirement on the application environment is lower;

3. the calculation and judgment speed is higher;

4. the accuracy is high, and the misjudgment rate is low;

5. the detection blind area is smaller;

6. the device has the advantage of non-contact;

7. a convolutional neural network (cnn) training sample is adopted;

8. and (3) segmented training and segmented detection and identification are adopted.

Drawings

Fig. 1 is a flow chart of the operation of the present system (the drive unit is a motor).

FIG. 2 is a block diagram of a convolutional neural network training process.

Detailed Description

The following are specific embodiments of the present invention and are further described with reference to the drawings, but the present invention is not limited to these embodiments.

The maximum opening degree (the angle between a cover plate and a plane when the cover plate is maximally opened) of the learning table is set to be 100 degrees, the cover plate of the learning table is closed from 100 degrees each time, when a sample is collected, the span of 1 degree is taken as a collection unit, namely 1000 pictures are uniformly collected in the process that the cover plate is closed from 100 degrees to 99 degrees and stored in a sample 100-99 folder, and 1000 pictures are uniformly collected in the process that the cover plate is closed from 99 degrees to 98 degrees and stored in a sample 99-98 folder, so that 8230and 8230are analogized.

Sample collection example:

training a network model by using the collected positive sample; the samples in the 100 folders collected in the first step are respectively used for training a convolutional neural network cnn model, the samples in the folder of samples 100-99 are trained into the cnn model and are stored as train _ model _001.H5, the samples in the folder of samples 99-98 are trained into the cnn model and are stored as train _ model _ 002.h5. 8230\ \ 825, and the samples in the folder of samples 1-0 are trained into the cnn model and are stored as train _ model _100.H5.

Testing the model by using the test sample, if the effect is not good, increasing the sample for training, and properly increasing the training times;

the specific test method comprises the following steps:

we start the cover plate closing from 100 degrees, and then test or actually use the cover plate in units of 1 degree, we return a signal to the system at the time of starting the closing of the motor and at the time of completing the closing of 1 degree, so that we can extract a picture from the real-time camera and call the cnn model corresponding to the picture to calculate the picture before receiving the signal of completing the closing of one degree after the cover plate starts the closing of the motor (for example, after the system receives the signal of starting the closing of the motor and before receiving the signal of completing the closing of the first degree, the system calls the first cnn model, namely train _ model _001.H5, to calculate the picture \\8230; \\\ 8230;).

Training until a good enough model is obtained;

mounting the test card on actual hardware for testing; (test methods are as above).

Adjusting until reaching the standard.

The specific system work flow is as follows:

starting a system, sending a closing instruction to a closed motor, returning a signal to the motor and starting to close (returning a signal to the system every time the motor completes one-degree closing), before receiving the signal completing one-degree, extracting a picture from a real-time camera by the system and calling a first model and the extracted picture for calculation, and setting a variable with an initial value of 0 to measure the number of signals received by the system so as to determine to call a plurality of trained cnn models; if the calculation result shows that an obstacle exists, the system sends a pause instruction to the motor, meanwhile, the system gives an alarm, the system continues to extract real-time pictures from the camera and calculates the real-time pictures with the corresponding model, if the calculation result shows that the obstacle exists, the real-time pictures are continuously extracted and calculated, if the calculation result shows that the obstacle does not exist, the alarm is closed, a closing instruction of 8230, is sent to the motor, and after the cover plate is completely closed, the whole system is closed.

The training process of the convolutional neural network comprises the following steps:

1. initializing a weight value by the network;

2. the input data is transmitted forwards through a convolution layer, a down-sampling layer and a full-connection layer to obtain an output value;

3. calculating the error between the output value of the network and the target value;

4. when the error is larger than the expected value, the error is transmitted back to the network, and the errors of the full connection layer, the down sampling layer and the convolution layer are obtained in sequence. The error of each layer can be understood as the total error of the network, and the network can bear the total error; when the error is equal to or less than our expected value, the training is ended.

5. And updating the weight according to the obtained error. And then proceeds to the second step.

The traditional anti-pinch system is mainly built based on a contact touch panel and an infrared light curtain, the contact touch panel and the infrared light curtain cannot realize pre-judgment, accidents are frequently caused, and the safety is poor; the latter infrared transmission line is susceptible to interference and erroneous determination. The traditional automatic detection method mainly adopts laser correlation and infrared light curtain for detection, but the laser correlation detection and the infrared light curtain detection belong to point and surface detectors, and can cause higher false alarm rate due to vibration. Meanwhile, the large-gap working environment has huge detection blind areas, and the detection accuracy cannot be guaranteed.

In recent years, machine vision technology is developed more and more mature, and a system based on machine vision has the advantages of high speed, high precision, non-contact and the like. The device does not suffer misjudgment due to the influence of a transmission line during working, has extremely simple requirements on working environment, and can automatically and stably run without being disturbed by external factors only by pre-storing all possible conditions in samples no matter what environment the device works in. A

Particularly, a convolutional neural network (cnn) is used for building a network model to train a sample, so that the system is high in accuracy and high in fault tolerance, and the dependence of the traditional machine vision system on light is reduced. The system closing process is subjected to segmented training and segmented detection, and the accuracy of system identification is improved.

The specific embodiments described herein are merely illustrative of the spirit of the invention. Various modifications or additions may be made to the described embodiments or alternatives may be employed by those skilled in the art without departing from the spirit or ambit of the invention as defined in the appended claims.

Claims

1. A machine vision anti-pinch system based on deep learning is characterized by comprising a Convolutional Neural Network (CNN), a data processing unit and a video image analysis learning platform, wherein the video image analysis learning platform comprises a closing driving unit, a camera, a substrate and a cover plate hinged on the substrate, and the closing driving unit can drive the cover plate to be close to or far away from the substrate around a hinged point;

the first step is as follows: collecting training samples

the second step is that: testing

Selecting any acquisition node, placing a certain barrier corresponding to the acquired picture between the plate and the cover plate, taking pictures of the barrier by the camera, if the data processing unit can match the acquired picture information with the corresponding sample and control the closed driving unit to stop, indicating that the acquisition of the training sample is reliable, otherwise, correcting by adding a picture library;

the third step: the vision anti-pinch system after deep learning is adapted to an application scene;

the working process of the visual anti-pinch system is as follows:

starting a visual anti-pinch system, sending a closing instruction to a closing driving unit, returning a signal to the closing driving unit and starting to close, returning a signal to the visual anti-pinch system every time the closing driving unit completes one-degree closing, extracting a picture from a real-time camera and calling a first model and the extracted picture for calculation before the visual anti-pinch system receives the signal completing one-degree, setting a variable with an initial value of 0 to measure the number of signals received by the visual anti-pinch system, and determining to call a plurality of trained convolutional neural network CNN models; if the calculated result shows that an obstacle exists, the visual anti-pinch system sends a pause instruction to the closing driving unit, meanwhile, the visual anti-pinch system gives an alarm, the visual anti-pinch system continuously extracts a real-time picture from the camera and calculates the real-time picture with a corresponding Convolutional Neural Network (CNN) model, if the calculated result shows that the obstacle exists, the real-time picture is continuously extracted and calculated, if the calculated result shows that the obstacle does not exist, the alarm is closed, a closing instruction is sent to the closing driving unit, the steps are repeated, and after the cover plate is completely closed, the visual anti-pinch system is closed;

the training process of the convolutional neural network CNN model is as follows:

1) Initializing the weight value by the network;

2) The input data is transmitted forward through the convolution layer, the down sampling layer and the full connection layer to obtain an output value;

3) Solving the error between the output value of the network and the target value;

4) When the error is larger than the expected value, the error is transmitted back to the network, and the errors of the full connection layer, the down sampling layer and the convolution layer are sequentially obtained; the error of each layer can be understood as the total error of the network, and the network can bear the total error; when the error is equal to or less than our expected value, the training is finished;

5) Updating the weight according to the obtained error; then proceed to step 2).