CN110781896A

CN110781896A - Track garbage identification method, cleaning method, system and resource allocation method

Info

Publication number: CN110781896A
Application number: CN201910989410.6A
Authority: CN
Inventors: 谢勇君; 黄靖雯; 张弛; 马文韬; 董瀚江; 刘欣; 李进桂; 黄衍铭; 林泽楠; 胡建硕; 黄晓杰; 吴倩童; 严冬松; 武建华; 庄师强
Original assignee: Jinan University
Current assignee: Jinan University
Priority date: 2019-10-17
Filing date: 2019-10-17
Publication date: 2020-02-11
Anticipated expiration: 2039-10-17
Also published as: CN110781896B

Abstract

The invention discloses a track garbage identification method, a cleaning method, a system and a resource allocation method, wherein the track garbage identification method comprises the following steps: acquiring a track video stream; decomposing a video stream frame by frame, preprocessing each frame of image to filter image noise, wherein the preprocessing comprises Cohen-Sutherland clipping, inputting the preprocessed image into a trained garbage recognition model, and performing image sample characteristic extraction, multi-scale prediction and bounding box prediction to recognize garbage; classifying the identified garbage by multi-label classification; and eliminating the bounding box of the same garbage detected for multiple times through non-maximum suppression. The image is simplified through Cohen-Sutherland clipping, only the image information between the tracks on the two sides is reserved, the influence of environmental factors on the image is removed, the image detection degree is further improved, and the image detection performance is improved.

Description

Track garbage identification method, cleaning method, system and resource allocation method

Technical Field

The invention relates to the field of garbage identification and treatment, in particular to a track garbage identification method, a track garbage cleaning system and a track garbage resource allocation method.

Background

At present, domestic research on modern tram track cleaning technology and equipment is in a starting stage, and the track cleaning of the modern tram mainly adopts a manual cleaning mode, but the mode is time-consuming and labor-consuming, low in efficiency and serious in dust pollution. The utility model patent (CN204875654U) that the time Ruidao company applied for and the utility model patent (CN204875656U) that Degao clean company applied for the scheme is similar basically, and domestic tramcar track cleaning cart still is in the manual work and controls the cleaning means through the display screen and accomplishes the action and whole start carries out the stage that cleans to the track, and the cleaning cart price is expensive and can't reach degree of automation. Meanwhile, as the position of the garbage cannot be accurately judged during manual judgment, the cleaning range and the error rate are enlarged, cleaning resources are consumed, and higher-efficiency cleaning cannot be achieved.

The tram cell type track rubbish cleaning cart that develops on the market now does not have the function of automated inspection rubbish, but needs the manual work to watch the screen and obtain the rubbish condition, compares in machine detection, and manual handling is very easy because watch the video for a long time and cause visual fatigue and then lose the information that carries among the video image, leads to rubbish not clear up or the mistake clearance. However, the domestic detection system for the groove rail garbage is mostly in the starting stage, and the existing relatively mature detection system for the combined type automatic cleaning vehicle for the track road surface is provided by the Chinese invention patent CN 104047248A, the visual detection technology adopted by the detection system detects the garbage position according to the color difference between the garbage and the track, however, the method has the limitation condition that the color difference gradient between the garbage and the track is required to be overlarge, and the detection speed is slow due to large calculation amount.

One of the difficulties in detecting garbage compared with detecting general objects is that the definition range of garbage is very wide, and for example, in an urban road scene, plastic bags, water bottles, paper scraps, and even sand and stones are all garbage. Although they all belong to garbage, the differences in color, texture, geometry are very large, spanning different object classes.

The characteristics extracted by the traditional image identification method are deficient in representativeness and robustness, and the detection precision is greatly reduced under the influence of the variation factors such as illumination, shielding and dimension; secondly, the video image is often interfered by various natural environments or man-made noises, so that the program can not obviously and clearly acquire the moving target from the video image.

Therefore, the precise cleaning method for the groove-shaped rail garbage of the tramcar based on deep learning is needed to be provided, different cleaning tools carried on a cleaning vehicle are called through positioning and identifying the groove-shaped rail garbage, the cleaning efficiency of the groove-shaped rail of the modern tramcar is improved, cleaning resources and manpower resources are saved, the potential safety hazard of the modern tramcar is effectively reduced, the running safety of the tramcar is guaranteed, the groove-shaped rail cleaning vehicle is developed to a direction with higher automation degree, and the popularization and development of the tramcar in wide cities in China are promoted.

Disclosure of Invention

The invention aims to overcome the defects in the prior art and provide a track garbage identification method, a track garbage cleaning system and a resource allocation method.

The purpose of the invention is realized by the following technical scheme:

a track trash identification method comprises the following steps:

acquiring a track video stream;

decomposing a video stream frame by frame, preprocessing each frame of image to filter image noise, wherein the preprocessing comprises Cohen-Sutherland clipping, inputting the preprocessed image into a trained garbage recognition model, and performing image sample characteristic extraction, multi-scale prediction and bounding box prediction to recognize garbage;

classifying the identified garbage by multi-label classification;

and eliminating the bounding box of the same garbage detected for multiple times through non-maximum suppression.

A rail garbage cleaning method is characterized in that after garbage is identified through the rail garbage identification method, a cleaning system is driven to complete cleaning work.

A track trash cleaning system comprising:

the cleaning module is a cleaning vehicle and is used for acquiring a control command to complete cleaning work;

and the cleaning control module is in data connection with other modules of the rail waste cleaning system and controls the other modules to operate, and comprises an MFC program instruction input unit and a physical instruction input unit.

A rail waste cleaning resource allocation method comprises the following steps:

equally dividing a section of track into a plurality of sections, each section being equally divided into a plurality of sampling points

x _ij(i＝1，2，...，n，j＝1，2，...，m)；

When the cleaning vehicle starts from the track starting point at a fixed speed, resetting the count coefficient to be 1, and starting the timer;

defining the identification coefficient flag:

when the garbage is identified during the traveling of the cleaning vehicle, the flag is 1, and the position coordinate y at the moment is recorded _count1；

Continuing to move, if no garbage is identified, setting flag to be 0, and recording the position coordinate y at the moment _count0Count plus 1;

counting the garbage occurrence positions of each section of track in one day;

continuously recording data every day, counting the probability mean value of each sampling point of each section of track according to data of a plurality of days, and distributing clean resources according to seasons and the probability mean value.

Compared with the prior art, the invention has the following advantages and beneficial effects:

1. the track garbage identification method simplifies the image by Cohen-Sutherland clipping, only retains the image information between the tracks at two sides, removes the influence of environmental factors on the image, further increases the image detection speed and improves the image detection performance.

2. According to the track garbage cleaning method, the cleaning system is driven to complete cleaning work after the track garbage is accurately identified, and the cleaning efficiency is improved.

3. The track garbage cleaning system separates the physical control panel from the MFC program instruction input unit, and avoids conflict among control programs.

4. The invention saves the prior cleaning resources by reasonably allocating the cleaning resources.

Drawings

FIG. 1 is a schematic diagram of the Cohen-Sutherland clipping algorithm;

FIG. 2 is a schematic diagram of the pretreatment by the Cohen-Sutherland clipping algorithm in example 1 of the present invention;

FIG. 3 is a schematic structural diagram of a multi-scale prediction module according to embodiment 1 of the present invention;

FIG. 4 is a network structure diagram of improvement of YOLOv3 in embodiment 1 of the present invention;

FIG. 5 is a schematic diagram of the bounding box prediction in embodiment 1 of the present invention;

FIG. 6 is a main view of embodiment 2 of the present invention;

FIG. 7 is a flowchart of the operation of a rail waste cleaning system according to embodiment 3 of the present invention;

fig. 8 is a flowchart of a method for allocating track garbage cleaning resources according to embodiment 4 of the present invention.

Detailed Description

For better understanding of the technical solutions of the present invention, the following detailed description is provided for the embodiments of the present invention with reference to the accompanying drawings, but the embodiments of the present invention are not limited thereto.

First, the embodiments of the present invention are based on the following principles:

(1) Cohen-Sutherland clipping algorithm

Approximating the acquired image as a set of line segments, and for a single line segment, if the line segment is completely within the boundary, reserving the line segment; if the line segment is completely out of the boundary, discarding the line segment; if a portion of the line segments are within the boundary and a portion of the line segments are outside the boundary, then the portion of the line segments within the boundary is retained.

Generally, to make a computer quickly determine what relationship a straight line segment is to a window, the following encoding method is used:

the sides of the window are extended to divide the two-dimensional plane into nine areas. As shown in fig. 1, by a single line segment P ₁P ₂For example, the line segment portion of the middle regionRetention, i.e. retention P ₃P ₄。

(2) YOLOv3 feature extraction

YOLO is a new method to apply a whole picture to a neural network at once. The network divides the picture into different regions, then gives the frame prediction and probability of each region, and assigns weights to all frames according to the probability. Finally, a threshold is set, and only the detection result that the score (probability value) exceeds the threshold is output, and YOLOv3 is the latest version of YOLO.

(3) Multi-label classification

Multi-label classification means that the candidate set is a multi-classification, not just a two-classification-yes or no problem, but rather a problem of which of the classes it belongs to. One sample belongs to one of a plurality of classifications, one sample belongs to one class, and the different classes are mutually exclusive. For multi-label classification, the label of a sample is not limited to only one class, but can have multiple classes, and there is a relationship between the different classes. Such as a piece of clothing, which has the characteristic categories of long sleeves, lace and other attributes, and the two attribute labels are not mutually exclusive but are related.

(4) Non-polar sigmoid activation function

The sigmoid activation function is generally used for binary classification, and assuming that there are 10 classes in total, it first converts a true value to between [0,1], respectively [0.01,0.05,0.4,0.6,0.3,0.1,0.5,0.4,0.06,0.8], then sets a probability threshold, and if it is greater than a probability threshold (generally 0.5), it is considered to belong to a certain class, otherwise it does not belong to a certain class. This property makes it suitable for use in multi-label classification, which is essentially a sigmoid classifier is applied to the result of each classification calculation in logits, and the final output of the neural network model is the vector logits [1,2,3,4,5,6,7,8,9,10], which is the final fully-connected output of the neural network.

(4) Non-maximum suppression (NMS)

Since the detector detects the same object many times (slightly different in center and size), and in most cases only a small number of different pixels are needed for detection, the problem of multiple detections on the same image needs to be solved by using non-maximum suppression (NMS) as a post-processing algorithm.

(5) softmax activation function

Wherein z represents a class neuron; j represents the element number in the category; z is a radical of _jA jth element output representing a z category; sigma (z) _jA multivariate output representing z kinds of neurons; z is a radical of _kOutput for the kth element of the z category, K ∈ [1, K ∈]And K represents the total number of symbols in the category.

The softmax function is used for multi-class neural network output, i.e. some z _jGreater than the other z, the component of this mapping is close to 1, the rest is close to 0, and is mainly applied to multi-classification. The reason for the exponentiation is the behavior of the analog max, and the need for a derivative function.

Example 1

A track trash identification method comprises the following steps:

s1, acquiring the track video stream: acquiring a track video stream through a camera mounted on the cleaning vehicle;

s2, decomposing the video stream frame by frame, and preprocessing each frame of image to filter image noise: the preprocessing comprises Cohen-Sutherland clipping, as shown in FIG. 2, dividing the image into three regions by taking the track in the image as the boundary, reserving the middle region, and taking a single line segment P ₁P ₂For example, the line segment part of the middle area is reserved, i.e. P is reserved ₃P ₄。

S3, inputting the preprocessed image into a trained garbage recognition model, and recognizing garbage through image sample feature extraction, multi-scale prediction and bounding box prediction;

s4, classifying the identified garbage through multi-label classification;

and S5, eliminating the bounding box of the same garbage detected for multiple times through non-maximum suppression.

Furthermore, the cameras comprise two groups of high-speed cameras for respectively acquiring the left rail and the right rail, and the two groups of cameras acquire the rail videos and then combine the rail videos into a bus rail video stream.

Further, the preprocessing the image further includes graying and gaussian filtering the image:

(1) graying:

P _gray＝0.11B+0.3R+0.59G

where B represents the pixel value of the blue component in the original color image, R represents the pixel value of the red component, and G represents the pixel value of the green component in the color image. P _grayIs a converted gray scale image.

The image acquired by the camera is an RGB color image, and the final gray value is obtained by comprehensively carrying out weighted average on RGB components of the image; the weight setting refers to the characteristics of human vision in physiology, and human eyes have the highest sensitivity to green and the lowest sensitivity to blue. Therefore, the weight set for blue is the lowest, and the combination of such proportions can more prominently show useful image information such as tramway, sign line, and garbage.

Such images retain essential features, reduce the total amount of information, and reduce the amount of computation.

(2) Gaussian filtering:

the original input image size is 416 x 416, and in order to acquire more transverse features, the image size is changed into 576 x 320 resolution images as network input;

wherein x represents the abscissa of the coordinate point; y represents the ordinate of the coordinate point; sigma ²Represents the variance of the random variable; g (x, y) represents the probability of the coordinate point (x, y) appearing.

The image is influenced by external conditions such as light, haze and the machine, a lot of noise interference often appears in the image, interference information can be well removed by using a denoising technology, and basic information of the image is restored.

Gaussian filtering is a linear smoothing filter that can effectively suppress noise. Gaussian filtering is a process of weighted averaging of the entire image. The value of each pixel point is obtained by weighted averaging of the pixel point and other pixel point values in the neighborhood. That is, each pixel in the image is scanned by using a template (or convolution and mask), the weighted average gray value of the pixels in the neighborhood determined by the template is used to replace the value of the pixel point in the center of the template, and the template coefficient of the gaussian filter is reduced along with the increase of the distance from the center of the template, so that the sharpness of the image after overall filtering is not changed greatly.

Because the size of the garbage in the image decomposed from one video segment is an invariant value, the horizontal feature expression of the garbage is less than that of the vertical feature, and the addition of the horizontal feature expression is beneficial to garbage recognition. In the embodiment of the invention, the length-width ratio of the network model input is changed, more transverse features are extracted by using a rectangular input network, and the size 416 multiplied by 416 of the original network input image is changed into the image with 576 multiplied by 320 resolution as the network input in consideration of reducing the influence of the change of the image resolution on the network as much as possible. On one hand, the transverse features can be more accurately extracted; on the other hand, the two have similar pixel numbers, and the real-time detection is not greatly influenced.

Further, the garbage recognition model is trained by the following steps:

s3.1, collecting a track garbage image sample;

s3.2, repeatedly inputting the garbage image samples to a YOLOv3 network structure, and obtaining a garbage recognition model through image sample feature extraction, multi-scale prediction, bounding box prediction and training processes;

the YOLOv3 network structure adopts a full convolution structure, and comprises a feature extraction module for feature extraction, a multi-scale prediction module for multi-scale prediction and a bounding box prediction module for bounding box prediction, wherein a YOLOv3 network structure diagram is shown in fig. 4, and the structure uses a series of convolution layers of convolution kernels of 3 × 3 and 1 × 1;

shortcut links (shortcut connections) are arranged between residual network layers of the feature extraction module, which have the same feature diagram size and the same number of convolution kernels; the Res (residual network) structure can well control the propagation of the gradient and avoid situations which are unfavorable for training, such as disappearance of the gradient or explosion. The learned features are passed to a classifier/regressor for prediction of coordinates, class labels, etc. of the bounding box.

The multi-scale prediction module comprises three convolutional layers which are sequentially arranged, as shown in fig. 3, the three convolutional layers can be sequentially arranged from top to bottom, the middle convolutional layer is subdivided into a first part and a second part, the network compressing the first part is compressed into a shallower (lower-level) network (when the arrangement mode is shown in fig. 3, the first part and the second part are arranged up and down, the first part is the upper half part of the convolutional layer, the second part is the lower half part of the convolutional layer), the connection with the weight between the networks of the second part smaller than a set value is cut off, the sparsely connected network is obtained and retrained, the weight of quantization connection is shared by using the weight, and the quantized weight and a codebook are compressed by adopting a coding mode combining predictive coding and run-length coding; the essence of the weight sharing quantification is that a clustering center of the layer weight is obtained by a clustering method, and then the original weight value is represented by a clustering center value; the predictive coding obtains a predicted value according to the statistical characteristics of data, and then transmits a difference signal between an image pixel and the predicted value, so that the transmitted code rate is reduced, and the purpose of compression is achieved; the run-length coding is a coding mode that stores one pixel value and the number of pixels having the same color. The compression rate can be further improved by combining the two encoding modes;

in the convolutional neural network, more convolutional layers and deeper network structures tend to have better extraction effect on target features. On the other hand, when the network is deepened, the extra convolutional layer may cause the network model parameters to be too large, thereby increasing the computation amount of the network. Meanwhile, in the field of object detection, although the deep network can respond to semantic features, the deep network contains few geometric information and is not beneficial to object detection; although the shallow network contains more geometric information, the semantic features of the image are not more, and the classification of the image is not facilitated, so that the prediction is performed by adopting a mode of fusing three scales (13 × 13, 26 × 26 and 52 × 52), and the accuracy of the improved YOLO algorithm on the detection of the small target is enhanced.

Further, as shown in fig. 5, the anchor boxes of the bounding box prediction module are obtained by a clustering method, and four coordinate values (t) are predicted for each bounding box _x,t _y,t _w,t _h) Dividing each image into S multiplied by S grid cells, and combining the width p of a bounding box for the predicted cells _wAnd high p _hA bounding box can be predicted;

b _x＝σ(t _x)+c _x

b _y＝σ(t _y)+c _y

the sum of squared error loss is adopted when training several coordinate values of the bounding box, and the error of the method can be calculated quickly, so that the detection efficiency is increased. YOLOv3 predicts the score of an object by logistic regression for each bounding box, and sets the value to 1 if the overlap between the predicted bounding box and the real border value is larger than that of all other predicted bounding boxes. If the overlap does not reach the preset threshold (the threshold set here is 0.5), the predicted bounding box is ignored, i.e. no loss value is displayed.

Further, the specific steps of classifying the identified garbage through multi-label classification are as follows:

mapping a truth value vector (a1, a2, a3, …) of a k dimension into a (b1, b2, b3, …) through a softmax activation function, wherein bi is a constant of 0-1, the sum of output neurons is 1.0 and is equivalent to a probability value, and performing multi-classification tasks according to the probability magnitude of bi; the softmax activation function is mainly used for calculating cross entropy, calculating after converting logits into probability distribution, and then taking the maximum in the probability distribution as a final classification result; the input of softmax is the image feature value and the output is the probability that the garbage belongs to each category.

The activation function introduces nonlinear factors to the neurons, so that the neural network can arbitrarily approximate any nonlinear function, and the neural network can be applied to a plurality of nonlinear models.

Further, since the detector detects the same object (the center and the size are slightly different) for multiple times, and in most cases, only a small amount of different pixels are needed for detection, the step of removing the bounding box of the same garbage detected for multiple times is as follows:

initializing and setting a detection threshold value;

filtering the bounding box according to the object score of the target garbage; bounding boxes with scores below a set threshold are ignored;

sorting according to the class classification probability of the candidate boxes, and then reserving the required bounding boxes according to the following steps:

marking a maximum probability bounding box needing to be reserved;

starting from the maximum probability rectangular frame, respectively judging whether the overlapping degree IOU of other probability rectangular frames and the maximum probability boundary frame is greater than a set threshold value, and throwing away the rectangular frames with the overlapping degree exceeding the threshold value;

selecting a maximum probability boundary box from the boundary boxes which are not thrown away, marking the boundary box as a boundary box to be reserved, then judging the overlapping degree of the maximum probability boundary box and other boundary boxes at the moment, and throwing away the boundary box with the overlapping degree exceeding a set threshold value;

and repeating the operation until all the remaining bounding boxes are removed and all the bounding boxes to be reserved are marked.

According to the practical situation of the periphery of the track, the garbage images are roughly divided into six categories: the first is leaves, the second is beverages, the third is plastic bags, the fourth is food residues (banana peel), the fifth is food residues (apple pits), and the sixth is cigarette ends. The collected images are rotated, cut and color-converted to expand a sample set, and the total 4068 pictures in the data sample set relate to six categories.

When the number of iterations is less than 1000, the data is stored once every 100 times, and when the number of iterations is more than 1000, the data is stored once every 5000 times.

The model trained by the improved YOLOv3 algorithm can greatly improve the speed of garbage detection and classification, so that the required cleaning tool can be called more quickly and the communication is more efficient.

Example 2

A rail trash cleaning method, wherein after the trash is identified by the rail trash identification method according to any one of claims 1 to 5, a cleaning system is driven to complete cleaning work.

Further, the driving instruction is obtained through an operator physical instruction or an instruction library call.

Example 3

A track trash cleaning system comprising:

the cleaning control module is in data connection with other modules of the rail waste cleaning system and controls the other modules to operate, and comprises an MFC program instruction input unit and a physical instruction input unit;

specifically, the cleaning control module includes:

(1) PLC programming module

In the embodiment of the invention, the PLC ladder diagram program of the lower computer of the cleaning vehicle is mainly divided into three parts to be written, namely a communication program, a main program (a physical key program) and a touch screen virtual key program.

1. Communication program design module

In the embodiment of the invention, an industrial touch screen industrial personal computer is used as an upper computer, and a PLC is used as a lower computer, so that the PLC is a slave station, and when the PLC is used as the slave station, a programming method and steps of a communication program are as follows:

1) in the first scanning, strictly setting related communication parameters;

2) in the first scan, connect "receive done interrupt" with "send message interrupt";

3) starting a receiving instruction RCV and waiting for a sending request of a master station;

4) in the receiving completion interrupt program, judging whether the received data is correct, if the request instruction is correctly judged, organizing the corresponding data into a buffer area, and calling a sending instruction XMT; if not, re-calling the receiving instruction RCV;

5) and calling the RCV to receive the command in the process of sending the finished interrupt program.

The communication parameter setting program adopts no parity check, each character has 8 data bits, the communication baud rate is 9600bps, and the communication protocol of the free port is transmitted through the port 0.

2. Physical key main program design module

In the design of the main program, a manual mode or an automatic mode is selected through input signals I0.0 and I0.1, and meanwhile, the two networks adopt a start-stop circuit and realize mutual self-locking, namely the input signals are released, and the output signals are still kept. And simultaneously, resetting all output signals (except the high-pressure water return signal), the intermediate relay and the input register when a reset instruction is activated by using a rising edge switch of the stop switch I2.0. In addition, the switch action of each part is realized by triggering the rising edge and the falling edge through a physical key in the automatic mode and the manual mode.

3. Touch screen virtual key programming module

The touch screen program is formed by changing an input area on the basis of a main program of physical keys. In order to prevent control conflict between the physical control panel and the touch screen, the physical control panel and the touch screen need to be strictly distinguished, and the touch screen key program in the embodiment of the invention only needs to additionally set the input points of the external input I area of the main program.

(2) Industrial control computer touch screen program design module

The touch screen program is developed by adopting a visual studio platform, and programming work is completed by establishing an MFC project. MFC (Microsoft Foundation classes) is a short name for Microsoft basic class library, is a C + + class library realized by Microsoft corporation, mainly encapsulates most windows API functions, and can greatly reduce the development period of programs.

The touch screen program has the following working principle, when the MFC program on the industrial personal computer is used for inputting, a user can press a control button on the touch screen of the industrial personal computer according to needs to control the cleaning tool to be started or closed, and when the button is pressed, the MFC program sends the combined control information to the PLC and executes operation after the CByteArray series control signals supporting dynamic establishment of arrays by taking bytes as units are utilized and then the information sending function under the MSCOMM control is used.

Because the programming languages of the upper computer and the lower computer are different, the upper computer and the lower computer cannot be directly identified by communication, in order to ensure the correctness and the integrity of data transmission, a cyclic redundancy check algorithm (CRC) is also arranged in a program, and the upper computer sends a data instruction according to the format of a communication protocol. Cyclic redundancy check is a hash function that generates a short fixed bit check code from data, such as network packets or computer files, and is mainly used to check errors that may occur after data transmission or storage. Any data code formed by binary bit strings can be in one-to-one correspondence with a polynomial with a coefficient of '0' or '1', by utilizing the property, the binary bit string control information sent by the MFC can calculate a CRC check code and send the CRC check code to the PLC along with the control information, after the PLC receives the control information, the CRC check code is calculated again, only when the CRC check code calculated by the PLC by utilizing a ladder diagram is consistent with the check code sent by the upper computer, the control information is considered to be valid, then the data is stored in a data cache region, and corresponding operation is started.

The signals sent by the user to the PLC on the interface use a continuous section of internal input contacts starting from I10.0, which do not occupy external input contacts, nor can they be input by external physical input, to ensure the independence of the MFC program from the physical keys. The physical input panel control subsystem is used only in emergency situations.

After an MFC program is opened on an industrial personal computer, according to preset program logic, a serial port is opened firstly, serial port communication is carried out with the PLC through RS485, according to a Modbus protocol, a communication protocol is set up by self, meanwhile, the RS485 is connected to a 0 port of the PLC, and a free port can be adopted for programming the PLC so as to adapt to the defined communication protocol.

After the communication channel is built, mode selection information is sent firstly, a user can select an automatic mode or a manual mode, and after the mode is selected, the MFC program sends a series of control information to the PLC, wherein the control information indicates that the automatic mode is started.

And the PLC receives the control information and then calculates the CRC code again, only when the CRC code calculated by the PLC by using a ladder diagram is the same as the check code sent by the industrial personal computer, the control information is considered to be valid, and then the data is stored in a data cache region and corresponding operation is started.

When the PLC receives the control information and carries out CRC check to determine that the information is valid, the data string is stored in the data register area, the PLC adopts the switching value when reading the data, directly reads the data bit of the existing data register area, determines that the corresponding built-in input contact is set or reset according to the setting or resetting of the data bit, then carries out logic judgment by a ladder diagram, and refreshes the output contact.

After the MFC program sends information related to the manual mode and the automatic mode, if the cleaning vehicle works in the automatic mode, the MFC program automatically opens the camera, and sends an image acquired by the camera to a groove-shaped rail garbage accurate detection and positioning algorithm for image processing, after the groove-shaped rail garbage accurate detection and positioning algorithm is processed, the position and the size of the garbage are identified, a control signal is immediately sent to the PLC after the garbage is identified, cleaning devices such as a high-pressure water gun are automatically started for cleaning, and the control signal is sent to the PLC again until the garbage is not detected on a monitoring interface, and the cleaning devices such as the high-pressure water gun are closed. If the cleaning cart work is in manual mode, the user can open the camera by oneself, and is the same, and the accurate detection of cell type rail rubbish also can sign rubbish place position and size with positioning algorithm, and the user presses corresponding control button, controls cleaning device and cleans, if needs clean third rail, the user also can open before sweep set rotation, preceding sweep set dust fall, third rail suction inlet dust fall clean the third rail.

When the PLC receives control information such as mode information sent by an MFC program, logic judgment is carried out through an internally loaded ladder diagram, and a corresponding external output contact is enabled to be connected or disconnected.

Furthermore, a dust suction port is formed in the bottom of the head of the cleaning trolley of the cleaning module, and a clean water tank and a sewage tank are arranged in the carriage of the cleaning trolley;

the track refuse cleaning system further comprises:

the wind pressure monitoring module is arranged at the dust suction port and used for measuring the pressure at the dust suction port, dirt such as dust, fallen leaves and silt on the groove-shaped rail can be cleaned only when the pressure value is large enough, and unnecessary electric energy loss is caused by overlarge wind pressure; and the waterway monitoring module is used for respectively detecting the water levels of the clean water tank and the sewage tank.

Furthermore, the water path monitoring module is a liquid level transmitter and is arranged at the bottom of the side surface of the clean water tank or the sewage tank so as to be convenient for maintenance and replacement, and the position of the opening is on the same plane with the water outlet of the water pipe.

Furthermore, the rail garbage cleaning system also comprises a temperature monitoring module, specifically a temperature sensor, which can be a DS18B20 digital temperature sensor and is arranged at an air inlet grid of a cleaning vehicle head with air circulation; the cleaning vehicle is provided with the sprayer and is arranged near an engine water tank of the cleaning vehicle, and when the temperature detected by the temperature sensor is higher than a set threshold value, which indicates that the working environment temperature of the cleaning vehicle is too high, the sprayer is started to spray by the cleaning control module.

Further, the track garbage cleaning system is driven by an engine, and the engine is controlled by a cleaning control module

The rail garbage cleaning system further comprises a hydraulic module, specifically an engine oil pressure sensor, which is used for measuring the pressure of engine oil, is arranged on an engine main oil gallery on the side surface of the engine, and lights an engine oil pressure warning lamp when the pressure of the engine oil is lower than a certain specified value.

The cleaning device comprises a hydraulic pump, an oil cylinder and a hydraulic cycloid motor, wherein the whole hydraulic module is an open single-pump multi-execution-element hydraulic system, and the cleaning control module controls and reverses the mechanism and comprises 7 parts of actions of lifting and descending of a sweeping disc, rotation of the sweeping disc, lifting of a front guide cleaning mechanism, lifting of a rear guide cleaning mechanism, tipping of a dustbin, opening and closing of a dustbin door and overturning of a groove-shaped rail cleaning device. The system can adopt the engine oil pressure sensor to measure the engine oil pressure, and the engine oil pressure sensor is arranged on the main oil duct of the engine on the side surface of the engine and used for monitoring the oil supply pressure of the engine lubricating system, judging whether the lubricating system works normally or not and preventing the engine from being damaged due to oil shortage.

Further, the track garbage cleaning system comprises a secondary engine, a working device and a lifting garbage can, wherein the secondary engine is used for driving the hydraulic pump and the fan, and can retract and lift the working device and the lifting garbage can so as to meet the power requirements of the fan and the hydraulic pump, and the secondary engine can possibly stop working if the water temperature of the secondary engine is too high;

the track garbage cleaning system further comprises an auxiliary engine monitoring module, specifically is a temperature sensor, can select a DS18B20 digital temperature sensor, is placed in a water tank of the auxiliary engine to measure the water temperature of the auxiliary engine water tank, each pin of the temperature sensor is separated by a heat shrink tube to prevent short circuit, and the inside of the temperature sensor is sealed with water and moisture proofing.

Further, the track garbage cleaning system comprises an external environment monitoring module, specifically comprises a distance monitoring module for detecting the distance between the cleaning vehicle and the obstacle, and a positioning module for positioning;

the cleaning vehicle is characterized in that an ultrasonic sensor can be selected as the positioning module, in the field working process of the cleaning vehicle, 25-40 kHz high-frequency ultrasonic waves are emitted outwards through a crystal oscillator through the ultrasonic sensor, then the frequency of the reflected waves is detected through the control module, if an object moves in an area, the frequency of the reflected waves slightly fluctuates, namely Doppler effect, so that the distance between the cleaning vehicle and an obstacle is measured and transmitted to the industrial personal computer system in real time. When the distance between the cleaning vehicle and the obstacle is detected to be less than the preset safety distance, the corresponding alarm device is automatically started to prompt a worker to pay attention to driving safety or enable the cleaning vehicle to adopt emergency braking.

The cleaning vehicle is provided with the positioning module (such as a GPS), so that the real-time position of the cleaning vehicle can be known, the working route of the cleaning vehicle at each time can be recorded, and the remote control center can know the state of the cleaning vehicle conveniently. If an accident happens, the accurate position of the cleaning vehicle can be determined at the first time.

Based on the track rubbish cleaning system comprising the modules, the industrial personal computer receives data and processes the data in the following mode:

according to the actual situation and the comparison of various data transmission modes, GPRS (General packet radio Service) can be selected as a data transmission method, and the GPRS is responsible for sending data information to the industrial personal computer module.

When continuous real-time data transmission is used, the data format of the transmission needs to be designed. In the data transmission process, a TCP/IP Protocol (Internet Protocol Suite Internet Protocol) is adopted to achieve the purpose of high data transmission frequency requirement and complete and reliable transmission to the industrial personal computer.

The industrial personal computer needs to quickly and accurately identify different types of data transmitted from the lower computer. The real-time data of the wind pressure sensor, the liquid level transmitter, the engine oil pressure sensor, the temperature sensor and the ultrasonic sensor are respectively identified and transferred to the database, the data are checked and judged, and alarm information is displayed on an interface to remind a worker to process in time if the parameters are abnormal.

The parameters acquired by the system can be stored in the cleaning control module. After the data enters the database, the data is judged according to the standard of normal data and stored in a corresponding data Table for subsequent processing and application.

The design of a state monitoring interface on the industrial personal computer can be further designed based on the MFC engineering of the groove-shaped rail garbage accurate cleaning control system. The state monitoring interface is designed to display various data including wind pressure, hydraulic pressure, oil pressure, temperature and GPS position measured by the sensor. Creating a new sub-thread window in the main dialog box designed by the MFC engineering, establishing a basic state monitoring interface frame in the sub-window, adding a required control, adding a corresponding control response function, calling an API interface function related to database operation, introducing data stored in a database, displaying a dynamic change diagram of each parameter curve through the interface, and comparing and analyzing the parameters and the change curves thereof with rules in a knowledge base by adopting an expert evaluation system in a system evaluation method. The system gives information such as safe prompt and warning about the operation of the cleaning vehicle according to warning lines preset by each parameter under the normal working state of the cleaning vehicle so as to ensure the safe operation of the cleaning vehicle.

Two large applications integrating multiple functions are integrated into one control panel, so that the operation steps in the application control process are reduced, the operation time of a user in the control process of the cleaning device is saved, and the control process is more convenient and faster; meanwhile, each module of the cleaning device is definite in division and operates cooperatively, so that the cleaning process is more efficient and intelligent.

Example 4

A rail waste cleaning resource allocation method comprises the following steps:

x _ij(i＝1，2，...，n，j＝1，2，...，m)；

defining the identification coefficient flag:

Continuing to move, if no garbage is identified, setting flag to be 0, and recording the position coordinate y at the moment _count0Count plus 1.

Counting the garbage occurrence positions of each section of track in one day;

Further, when data of collection points of the past week and a seasonal amount S (S1, 2,3,4 respectively indicate spring, summer, fall, and winter) are input before the configuration, and the counted positions of occurrence of garbage in each track during the day are calculated, the garbage probability average value is calculated Statistical water consumption Q ═ P _ij*Z _s，Z _SRepresenting the amount of water used during season S.

The above embodiments are preferred embodiments of the present invention, but the present invention is not limited to the above embodiments, and any other changes, modifications, substitutions, combinations, and simplifications which do not depart from the spirit and principle of the present invention should be construed as equivalents thereof, and all such changes, modifications, substitutions, combinations, and simplifications are intended to be included in the scope of the present invention.

Claims

1. A rail refuse recognition method is characterized by comprising the following steps:

acquiring a track video stream;

classifying the identified garbage by multi-label classification;

2. The track trash recognition method of claim 1, wherein the Cohen-Sutherland cropping divides the image into three regions with the track in the image as a boundary and reserves a middle region.

3. The rail trash recognition method of claim 1, wherein the pre-processing of the image further comprises graying, gaussian filtering of the image; when gaussian filtering, the image size lateral size is increased.

4. The rail waste recognition method according to claim 1, wherein the waste recognition model is trained by the following steps:

collecting a rail rubbish image sample;

repeatedly inputting the garbage image samples to a YOLOv3 network structure, and obtaining a garbage recognition model through image sample feature extraction, multi-scale prediction, bounding box prediction and training processes;

the YOLOv3 network structure adopts a full convolution structure and comprises a feature extraction module for feature extraction, a multi-scale prediction module for multi-scale prediction and a bounding box prediction module for bounding box prediction;

a shortcut link is arranged between residual error network layers of the feature extraction module, which have the same feature map size and the same number of convolution kernels;

the multi-scale prediction module comprises three convolutional layers which are sequentially arranged, the middle convolutional layer is subdivided into a first part and a second part, a network of the first part is compressed, connection with the weight between the networks of the second part smaller than a set value is cut off, a sparsely connected network is obtained and retrained, the weight of quantization connection is shared by using the weight, and then the quantized weight and a codebook are compressed by adopting a coding mode of combining predictive coding and run coding.

5. A rail debris cleaning method, characterized in that after debris is identified by the rail debris identification method according to any one of claims 1 to 4, a cleaning system is driven to complete the cleaning work.

6. The rail trash cleaning method of claim 5, wherein the driving instructions are obtained by an operator physical instruction or an instruction library call.

7. A track trash cleaning system, comprising:

8. The rail trash cleaning system of claim 7, wherein:

a dust suction port is formed in the bottom of the head of the cleaning trolley of the cleaning module, and a clean water tank and a sewage tank are arranged in the carriage of the cleaning trolley;

the track refuse cleaning system further comprises:

the wind pressure monitoring module is arranged at the dust suction port and used for measuring the pressure at the dust suction port;

and the waterway monitoring module is used for respectively detecting the water levels of the clean water tank and the sewage tank.

9. The rail trash cleaning system of claim 8, wherein the waterway monitoring module is a liquid level transmitter and is installed at the bottom of the side surface of the clean water tank or the sewage tank, and the opening position is on the same plane as the water outlet of the water pipe.

10. A rail waste cleaning resource allocation method is characterized by comprising the following steps:

x _ij(i＝1，2，...，n，j＝1，2，...，m)；

defining the identification coefficient flag:

counting the garbage occurrence positions of each section of track in one day;