CN110781896B - Track garbage identification method, cleaning method, system and resource allocation method - Google Patents

Abstract

The invention discloses a track garbage identification method, cleaning method, system and resource allocation method. The track garbage identification method includes the steps of: acquiring a track video stream; decomposing the video stream frame by frame, and preprocessing each frame of images to Filtering image noise, the preprocessing includes Cohen-Sutherland cropping, inputting the preprocessed image into a trained garbage identification model, image sample feature extraction, multi-scale prediction, and bounding box prediction to identify garbage; garbage classification; remove bounding boxes of the same garbage detected multiple times through non-maximum suppression. The invention simplifies the image through Cohen-Sutherland cropping, only retains the image information between the two sides of the track, removes the influence of environmental factors on the image, further improves the image detection degree, and improves the image detection performance.

Description

A 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 cleaning method, a system and a resource allocation method.

Background technique

At present, the domestic research on modern tram track cleaning technology and equipment is still in its infancy. For modern tram track cleaning, manual cleaning is the main method, but this method is time-consuming, labor-intensive, low in efficiency, and has serious dust pollution. . The utility model patent (CN204875654U) applied by Shiruida Company is basically similar to the utility model patent (CN204875656U) applied by Degaojie Company. And in the stage of cleaning the track when the whole process is turned on, the cleaning car is expensive and cannot achieve the degree of automation. At the same time, due to the inability to accurately determine the location of garbage during manual judgment, the cleaning range and error rate increase, which means that cleaning resources are consumed and higher efficiency cleaning cannot be achieved.

The tram trough-type track garbage cleaning vehicles currently developed on the market do not have the function of automatically detecting garbage, but need to manually watch the screen to obtain the garbage situation. Compared with machine detection, manual processing is very easy because of watching videos for a long time. It causes visual fatigue and misses the information carried in the video image, resulting in garbage not being cleaned up or by mistake. Nowadays, most of the domestic trough rail garbage detection systems are in their infancy. At present, the relatively mature one is the composite track road automatic cleaning vehicle proposed by the Chinese invention patent CN 104047248A. The visual detection technology adopted is based on the color difference between the garbage and the track. However, the limitation of this method is that the color difference gradient between the garbage and the track must be too large, and the detection speed is slow due to the large amount of calculation.

Compared with ordinary object detection, a major difficulty in garbage detection is that the definition of garbage is very wide. For example, in urban road scenes, plastic bags, water bottles, paper scraps, and even sand and gravel are garbage. Although they are all garbage, they are very different in color, texture, geometry, and across different object classes.

The features extracted by traditional image recognition methods are lacking in representativeness and robustness, and the detection accuracy will be greatly reduced under the influence of changing factors such as illumination, occlusion, and scale. Second, because video images are often affected by various natural environments Or the interference of man-made noise, which makes the program unable to obtain the moving target from the video image clearly and clearly.

Therefore, it is necessary to provide a method for accurate cleaning of tram trough-shaped track garbage based on deep learning. By locating and identifying the trough-shaped track garbage, different cleaning tools mounted on the cleaning vehicle are called to improve the modern tram trough. It can improve the cleaning efficiency of trough rails, save cleaning resources and human resources, effectively reduce the hidden danger of modern tram safety accidents, ensure the safe operation of trams, make trough rail cleaning vehicles develop towards a higher degree of automation, and promote the development of trams in the future. The promotion and development of vast cities in my country.

SUMMARY OF THE INVENTION

The purpose of the present invention is to overcome the shortcomings and deficiencies of the prior art, and to provide a method for identifying rail waste, a cleaning method, a system, and a method for resource allocation. The method and system drive the cleaning system to complete the cleaning work by accurately identifying the rail waste and improve the cleaning efficiency.

The object of the present invention is achieved through the following technical solutions:

A method for identifying orbital garbage, comprising the steps of:

Get the track video stream;

The video stream is decomposed frame by frame, and each frame of image is preprocessed to filter image noise. The preprocessing includes Cohen-Sutherland cropping, the preprocessed image is input into the trained garbage recognition model, image sample feature extraction, multi-scale Prediction, bounding box prediction to identify garbage;

Sort the identified garbage through multi-label classification;

The bounding boxes of the same garbage detected multiple times are removed by non-maximum suppression.

A method for cleaning orbital garbage, after the garbage is identified by the above-mentioned method for identifying orbital garbage, the cleaning system is driven to complete the cleaning work.

An orbital trash cleaning system comprising:

The cleaning module is a cleaning vehicle, which is used to obtain control instructions to complete the cleaning work;

The cleaning control module is connected with other modules of the track garbage cleaning system and controls the operation of other modules, including the MFC program command input unit and the physical command input unit.

A method for allocating orbital garbage cleaning resources, comprising the following steps:

Divide a track into several segments, and each segment is equally divided into several points

x _ij (i=1,2,...,n,j=1,2,...,m);

When the cleaning vehicle starts from the starting point of the track at a fixed speed, reset the count coefficient count=1, and the timer starts to work;

Define the identification coefficient flag:

During the traveling of the cleaning vehicle, if garbage is identified, then flag=1, and record the position coordinate y _count1 at this time;

Continue to move, if no garbage is identified, then flag=0, record the position coordinate y _count0 at this time, and add 1 to count;

Count the location of garbage in each track of the day;

Continuously record daily data, use several days of data to count the probability mean of each sampling point of each segment of the track, and allocate cleaning resources according to the season and the probability mean.

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

1. The track garbage identification method of the present invention simplifies the image through Cohen-Sutherland cropping, only retains the image information between the two tracks, removes the influence of environmental factors on the image, further improves the image detection speed, and improves the image detection performance.

2. The track garbage cleaning method of the present invention drives the cleaning system to complete the cleaning work after accurately identifying the track garbage, thereby improving the cleaning efficiency.

3. The track garbage cleaning system of the present invention separates the physical control panel from the MFC program instruction input unit to avoid conflicts between control programs.

4. The present invention saves priority cleaning resources through rational allocation of cleaning resources.

Description of drawings

Figure 1 is a schematic diagram of the principle of the Cohen-Sutherland clipping algorithm;

Fig. 2 is the schematic diagram of utilizing Cohen-Sutherland clipping algorithm preprocessing in embodiment 1 of the present invention;

3 is a schematic structural diagram of a multi-scale prediction module in Embodiment 1 of the present invention;

Fig. 4 is the network structure diagram of embodiment 1 YOLOv3 improvement of the present invention;

5 is a schematic diagram of bounding box prediction in Embodiment 1 of the present invention;

6 is a main body diagram of Embodiment 2 of the present invention;

Fig. 7 is the working flow chart of the orbital garbage cleaning system according to the third embodiment of the present invention;

FIG. 8 is a flow chart of a method for allocating resources for cleaning wastes on a track according to Embodiment 4 of the present invention.

Detailed ways

In order to better understand the technical solutions of the present invention, the embodiments provided by the present invention are described in detail below 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

Approximate the collected image as a set of line segments. For a single line segment, if it is completely within the boundary, keep the line segment; if it is completely outside the boundary, discard the line segment; if some line segments are within the boundary and some are outside the boundary, then Part of the line segment within the boundary is preserved.

Generally speaking, in order to enable the computer to quickly determine the relationship between a straight line segment and the window, the following coding method is used:

Extend the sides of the window, dividing the 2D plane into nine regions. As shown in FIG. 1 , taking a single line segment P ₁ P ₂ as an example, the line segment in the middle area is partially reserved, that is, P ₃ P ₄ is reserved.

(2) YOLOv3 feature extraction

YOLO is a brand new method that applies an entire image to a neural network at once. The network divides the image into different regions, then gives the bounding box prediction and probability of each region, and assigns weights to all bounding boxes according to the probability. Finally, set the threshold, and only output the detection results whose score (probability value) exceeds the threshold. YOLOv3 is the latest version of YOLO.

(3) Multi-label classification

Multi-label classification means that the candidate set is a multi-class, not just a binary classification - a question of yes or no, but a question of which of the multiple classes it belongs to. A sample belongs to and only belongs to one of multiple categories, a sample can only belong to one category, and different categories are mutually exclusive. For multi-label classification, the label of a sample is not limited to one category, but can have multiple categories, and different categories are related. For example, a piece of clothing has attributes such as long sleeves and lace. These two attribute tags are not mutually exclusive, but related.

(4) Nonpolar sigmoid activation function

The sigmoid activation function is generally used for two classifications. Assuming that there are 10 classifications in total, it first converts a true value between [0, 1], which are [0.01, 0.05, 0.4, 0.6, 0.3, 0.1, 0.5, 0.4, 0.06, 0.8], and then set a probability threshold, if it is greater than a probability threshold (usually 0.5), it is considered to belong to a certain category, otherwise it does not belong to a certain category. This property makes it suitable for use in multi-label classification. In essence, it is actually a sigmoid classifier for the results of each classification calculation in logits to determine whether the sample belongs to a certain category. It is also assumed that the final neural network model The output is such a 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 will detect the same object multiple times (with slightly different centers and sizes), and in most cases only a small number of pixels with different detections are required, it is necessary to use non-maximum suppression (NMS) as a post-processing algorithm to solve the problem of the same object. The problem of multiple detection of images.

(5) softmax activation function

Among them, z represents the category neuron; j represents the element number in the category; z _j represents the jth element output of the z category; σ(z) _j represents the multivariate output of the z category neuron; z _k is the kth category of the z category. element output, k ∈ [1, K], where K represents the total number of symbols in the category.

The softmax function is used for multi-classification neural network output, that is, if one z _j is larger than other z, the component of this map is approximated to 1, and the rest is approximated to 0. It is mainly used in multi-classification. The reason for taking the exponent is to simulate the behavior of max, and the second is to need a derivable function.

Example 1

A method for identifying orbital garbage, comprising the steps of:

S1. Get the track video stream: obtain the track video stream through the camera installed on the cleaning vehicle;

S2. Decompose the video stream frame by frame, and preprocess each frame of image to filter image noise: the preprocessing includes Cohen-Sutherland cropping, as shown in Figure 2, the image is divided into three parts with the track in the image as the boundary region, the middle region is reserved, taking a single line segment P ₁ P ₂ as an example, the line segment part of the middle region is reserved, that is, P ₃ P ₄ is reserved.

S3. Input the preprocessed image into the trained garbage identification model, image sample feature extraction, multi-scale prediction, and bounding box prediction to identify garbage;

S4, classify the identified garbage through multi-label classification;

S5. Remove the bounding box of the same garbage detected multiple times through non-maximum suppression.

Further, the camera includes two groups of high-speed cameras for obtaining the left track and the right track respectively, and the two groups of cameras obtain the track video and then synthesize it into a bus track video stream.

Further, the preprocessing of the image also includes grayscale and Gaussian filtering of the image:

(1) Grayscale:

P _gray = 0.11B+0.3R+0.59G

Among them, 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 _gray is the converted grayscale image.

The image obtained by the camera is an RGB color image, and the final gray value is obtained by comprehensively weighting the three RGB components of the image; the weight setting refers to the characteristics of human vision in physiology. The human eye is most sensitive to green and sensitive to blue. lowest. Therefore, the weight of blue is set to be the lowest, and this combination of ratios can highlight useful image information such as tram tracks, marking lines, and garbage.

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

(2) Gaussian filter:

The original input image size is 416×416. In order to obtain more horizontal features, the image size is changed to 576×320 resolution image as the network input;

Among them, x represents the abscissa of the coordinate point; y represents the ordinate of the coordinate point; σ ² represents the variance of the random variable; G(x, y) represents the probability of the coordinate point (x, y).

Affected by external conditions such as light, haze, and the machine itself, images often have a lot of noise interference. By using denoising technology, the interference information can be better removed and the basic information of the image can be 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 is weighted and averaged by itself and other pixels in the neighborhood. That is to say, a template (or convolution, mask) is used to scan each pixel in the image, and the weighted average gray value of the pixels in the neighborhood determined by the template is used to replace the value of the center pixel of the template. The template coefficient decreases as the distance from the template center increases, so the overall image clarity after filtering does not change much.

Since the size of the garbage in the decomposed image in a video is a constant value, the horizontal feature expression of garbage is less than the vertical feature, and increasing the horizontal feature expression is helpful for garbage identification. In the embodiment of the present invention, the aspect ratio of the input of the network model is changed, the rectangular input network is used to extract more lateral features, and the original network input image size is changed to 416×416 in consideration of reducing the influence of the change of image resolution on the network as much as possible. A 576×320 resolution image is used as the network input. On the one hand, the lateral features can be extracted more accurately; on the other hand, the two have a similar number of pixels, which has little impact on the real-time detection.

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

S3.1. Collect orbital garbage image samples;

S3.2. Repeatedly input garbage image samples to the YOLOv3 network structure, and obtain a garbage identification model through image sample feature extraction, multi-scale prediction, bounding box prediction and training process;

The YOLOv3 network structure adopts a full convolution structure, including 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. The YOLOv3 network structure diagram is shown in Figure 4 As shown, the structure uses a series of convolutional layers with 3*3 and 1*1 convolution kernels;

Shortcut connections are set between the residual network layers of the feature extraction module with the same feature map size and the same number of convolution kernels; this Res (residual network residual network) structure can be very good Control the propagation of gradients to avoid situations such as gradient disappearance or explosion that are not conducive to training. The learned features are passed to the classifier/regressor for prediction of bounding box coordinates, class labels, etc.

The multi-scale prediction module includes arranging three convolution layers in sequence. As shown in Figure 3, the three convolution layers can be arranged in order from top to bottom, and the middle convolution layer is subdivided into a first part and a second part. The first part of the network is compressed into a shallower (low-level) network (in the arrangement shown in Figure 3, the first part and the second part are arranged up and down, the first part is the upper half of the convolutional layer, and the second part is The lower half of the convolutional layer), truncate the connection between the second part of the network whose weight is less than the set value, obtain a sparsely connected network and retrain it, use the weight to share the weight of the quantized connection, and then use the quantized weight and codebook. The coding method combining predictive coding and run-length coding is used for compression; the essence of the weight sharing quantization is to first obtain the cluster center of the weight of the layer through the clustering method, and then use the cluster center value to represent the original weight value; the predictive coding is The predicted value is obtained according to the statistical characteristics of the data, and then the difference signal between the image pixel and its predicted value is transmitted, so as to reduce the transmission code rate and achieve the purpose of compression; The encoding of the number of pixels. Combining the above two encoding methods can further improve the compression rate;

In convolutional neural networks, more convolutional layers and deeper network structures tend to have better extraction effects on target features. But on the other hand, when the network is deepened, the additional convolutional layers will cause too many network model parameters, thus increasing the computational load of the network. At the same time, in the field of object detection, although the deep network can respond to semantic features, it does not contain much geometric information, which is not conducive to object detection; although the shallow network contains more geometric information, there are not many semantic features of the image. , which is not conducive to the classification of images. Therefore, three scales (13*13, 26*26, 52*52) are used to make predictions, which enhances the accuracy of the improved YOLO algorithm for small target detection.

Further, as shown in Figure 5, the anchor boxes of the bounding box prediction module are obtained by clustering, and four coordinate values (t _x , t _y , t _w , _th ) are predicted for each bounding box, and the Each image is divided into S×S grid cells. For the predicted cell, the bounding box can be predicted by combining the width p _w and height p _h of the bounding box;

b _x =σ(t _x )+c _x

b _y =σ(t _y )+ _cy

When training several coordinate values of the bounding box, the sum of squared error loss (sum of squared distance error loss) is used. The error of this method can be quickly calculated and the detection efficiency is increased. YOLOv3 predicts the score of an object through logistic regression for each bounding box. If the overlap between the predicted bounding box and the actual bounding box value is greater than all other predicted bounding boxes, set its value to 1. If the overlapping part does not reach the preset threshold (the threshold set here is 0.5), then the predicted bounding box will be ignored, that is, there is no loss value.

Further, the specific steps for classifying the identified garbage through multi-label classification are:

A k-dimensional ground truth vector (a1, a2, a3, ...) is mapped to a (b1, b2, b3, ...) through the softmax activation function, where bi is a constant from 0 to 1, and the sum of the output neurons is 1.0, equivalent to the probability value, the multi-classification task is carried out according to the probability of bi; the softmax activation function is mainly used to calculate the cross entropy, convert the logits into a probability distribution and then calculate, and then take the largest probability distribution as the final. Classification results; 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 approximate any nonlinear function arbitrarily, so that the neural network can be applied to many nonlinear models.

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

Initialize, set detection threshold;

Filter bounding boxes based on the objectness score of the target garbage; bounding boxes with scores lower than the set threshold will be ignored;

Sort the candidate boxes according to the class classification probability, and then keep the required bounding boxes according to the following steps:

Mark the maximum probability bounding box that needs to be preserved;

Starting from the maximum probability rectangle, determine whether the overlap IOU of other probability rectangles and the maximum probability bounding box is greater than the set threshold, and discard the rectangle whose overlap exceeds the threshold;

In the bounding box that has never been thrown away, select the bounding box with the maximum probability, mark it as the bounding box to be retained, and then judge the overlap between the bounding box with the maximum probability and other bounding boxes at this time, and throw away the boundary whose overlap exceeds the set threshold. frame;

Repeat the above operation until all remaining bounding boxes are removed and all bounding boxes to be kept are marked.

According to the actual situation around the track, the garbage images are roughly divided into six categories: the first category is leaves, the second category is beverages, the third category is plastic bags, the fourth category is food residues (banana peels), and the fifth category is Food scraps (apple cores), the sixth category is cigarette butts. The collected images are rotated, cropped, and color converted to expand the sample set. The data sample set has a total of 4068 images, involving six categories.

When the number of iterations is less than 1000, it is saved every 100 times, and when it is greater than 1000, it is saved every 5000 times.

The model trained by the improved YOLOv3 algorithm can greatly improve the speed of garbage detection and classification, making the invocation of the required cleaning tools faster and the communication more efficient.

Example 2

A method for cleaning rail garbage, after the garbage is identified by the method for identifying rail garbage according to any one of claims 1-5, a cleaning system is driven to complete the cleaning work.

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

Example 3

An orbital trash cleaning system comprising:

The cleaning module is a cleaning vehicle, which is used to obtain control instructions to complete the cleaning work;

The cleaning control module is connected with other modules of the orbital garbage cleaning system and controls the operation of other modules, including the MFC program command input unit and the physical command input unit;

Specifically, the cleaning control module includes:

(1) PLC programming module

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

1. Communication programming module

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

1) Strictly set relevant communication parameters in the first scan;

2) In the first scan, connect "receive complete interruption" and "send message interruption";

3) Turn on the receiving command RCV and wait for the sending request from the master station;

4) In the receiving completion interrupt program, judge whether the received data is correct, if the request command is judged correctly, organize the corresponding data into the buffer, and call the sending command XMT; if it is not correct, call the receiving command RCV again;

5) In the sending completion interrupt program, call RCV to receive the instruction.

No parity is used in the communication parameter setting program, each character has 8 data bits, the communication baud rate is 9600bps, and the free port communication protocol is transmitted through the port port0.

2. Physical button main programming module

In the design of the main program, the manual mode or the automatic mode is selected by inputting the signals I0.0 and I0.1. At the same time, the two networks use a start-stop circuit and realize mutual self-locking, that is, release the input signal and output the signal. still remain. At the same time, when the reset command is activated by the rising edge switch of the stop switch I2.0, all output signals (except the high-pressure water return signal), the intermediate relay and the input register are reset. In addition, in the automatic mode and the manual mode, the physical buttons are used to trigger the rising edge and the falling edge to realize the switching action of each component.

3. Touch screen virtual key programming module

The touch screen program is based on the physical button main program to change the input area. In order to prevent control conflicts between the physical control panel and the touch screen, it is necessary to strictly distinguish the two. The touch screen key program in the embodiment of the present invention only needs to set the input point of the external input area I of the main program separately.

(2) IPC touch screen programming module

The touch screen program is developed using the Visual Studio platform, and the programming work is completed by establishing an MFC project. MFC (Microsoft Foundation Classes) is the abbreviation of Microsoft Foundation Class Library. It is a C++ class library implemented by Microsoft. It mainly encapsulates most of the windows API functions, which can greatly reduce the development cycle of the program.

The working principle of the touch screen program is roughly as follows. When using the MFC program on the industrial computer for input, the user can press the control button on the touch screen of the industrial computer to control the cleaning tool to start or close. When the button is pressed, the MFC program will use the After the serial control signal of CByteArray class that supports dynamic establishment of an array in bytes, the information sending function under the MSCOMM control is used to send the combined control information to the PLC and execute the operation.

Due to the different programming languages between the upper computer and the lower computer, they cannot be directly identified by communication. In order to ensure the correctness and integrity of data transmission, a Cyclic Redundancy Check (CRC) algorithm is also set in the program. The format of the protocol sends data commands. Cyclic redundancy check is a hash function that generates a short fixed-digit check code based on data such as network packets or computer files. It is mainly used to check errors that may occur after data transmission or storage. Any data code composed of a binary bit string can correspond to a polynomial with a coefficient of '0' or '1' one by one. Using this property, the binary bit string control information sent by MFC can calculate the CRC check code, and It is sent to the PLC together with the control information. After the PLC receives the control information, it calculates the CRC check code again. Only when the CRC check code calculated by the PLC using the ladder diagram is consistent with the check code sent by the upper computer, it is considered that the CRC check code is the same. The control information is valid, then the data is stored in the data buffer and the corresponding operation is started.

The signal sent by the user to the PLC on the interface uses a continuous built-in input contact starting from I10.0. These contacts do not occupy the external input contact, nor can they be input by external physical, so as to ensure the MFC program and physical keys independence. The physical input panel control subsystem is for emergency use only.

After opening the MFC program on the industrial computer, according to the preset program logic, first open the serial port to communicate with the PLC through RS485. According to the Modbus protocol, the communication protocol is customized. At the same time, the RS485 will be connected to the 0 port of the PLC, which can be used for the PLC. Freeport programming to accommodate defined communication protocols.

After the communication channel is set up, the mode selection information is sent first. The user can choose the automatic or manual mode. After selecting the mode, the MFC program will send a series of control information to the PLC. The control information indicates that the automatic mode is activated.

Using the nature of the CRC check code, the binary bit string control information sent by the MFC can calculate the CRC check code, and send it to the PLC together with the control information. After the PLC receives the control information, it calculates the CRC check code again. Only when the PLC When the CRC check code calculated by the ladder diagram is the same as the check code sent by the industrial computer, the control information is considered valid, and then the data is stored in the data buffer area, and the corresponding operation is started.

When the PLC receives the control information and performs a CRC check to determine that the information is valid, it stores the data string in the data storage area. When the PLC reads the data, the switch value is used to directly read the data bits that already exist in the data storage area, and According to whether the data bit is set or reset, the corresponding built-in input contact is set or reset, and then it is judged by the ladder logic, and then the output contact is refreshed. In addition, the mode selection information will open the intermediate relay, if there is no mode selection information, that is, when no mode is selected, all operations on the cleaning device will be considered invalid.

After the MFC program has sent the information related to the manual and automatic modes, if the cleaning vehicle is working in the automatic mode, the MFC program will automatically turn on the camera, and hand over the images collected by the camera to the trough rail garbage accurate detection and positioning algorithm. Image processing, after the accurate detection and positioning algorithm of the trough rail garbage, the location and size of the garbage will be identified, and a control signal will be sent to the PLC immediately after the garbage is identified, and the cleaning devices such as high-pressure water guns will be automatically activated for cleaning. When the monitoring interface cannot detect garbage, it will send a control signal to the PLC again to turn off cleaning devices such as high-pressure water guns. If the cleaning truck works in manual mode, the user can turn on the camera by himself. Similarly, the precise detection and positioning algorithm of the trough rail garbage will also identify the location and size of the garbage. The user presses the corresponding control button to control the cleaning device to clean. If the third rail needs to be cleaned, the user can also turn on the rotation of the front sweeping disc, the front sweeping disc for dust reduction, and the third rail suction port for dust reduction to clean the third rail.

When the PLC receives the control information such as mode information sent by the MFC program, it will make a logical judgment through the internally loaded ladder diagram to make the corresponding external output contact on or off. The PLC has a strong load capacity. It can directly drive the general solenoid valve and AC contactor. When the external output contact is turned on, the electric diaphragm pump, solenoid valve, electromagnetic clutch and other controllers are energized. When the relay is pulled in, the solenoid valve is energized. When pulling in, the solenoid valve loses power, and then drives the hydraulic actuator and the electric actuator to make the corresponding cleaning device work.

Further, the bottom of the front of the cleaning vehicle of the cleaning module is provided with a dust suction port, and a clean water tank and a sewage tank are installed in the compartment of the cleaning vehicle;

The orbital trash cleaning system also includes:

The wind pressure monitoring module, in this embodiment, selects the wind pressure sensor, which is installed at the suction port to measure the pressure at the suction port. Only when the pressure value is large enough, can the dust, leaves, and sediment on the grooved rail be removed. Clean up, excessive wind pressure will cause unnecessary power loss; waterway monitoring module is used to detect the water level of the clean water tank and the sewage tank respectively.

Further, the waterway monitoring module is a liquid level transmitter, which is installed at the bottom of the side of the clean water tank or the sewage tank for easy maintenance and replacement, and the opening position is on the same plane as the water outlet of the water pipe.

Further, the rail garbage cleaning system also includes a temperature monitoring module, specifically a temperature sensor, and a DS18B20 digital temperature sensor can be selected, which is installed at the air inlet grille at the front of the air-circulating cleaning vehicle; the cleaning vehicle is equipped with a sprayer, which is installed Near the engine water tank of the cleaning vehicle, when the temperature detected by the temperature sensor is higher than the set threshold, indicating that the working environment temperature of the cleaning vehicle is too high, the sprayer will be turned on through the cleaning control module.

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

The rail garbage cleaning system also includes a hydraulic module, specifically an oil pressure sensor, which is used to measure the oil pressure and is installed on the engine main oil passage on the side of the engine. When the oil pressure is lower than a certain value, the oil pressure warning light is turned on.

Specifically, it includes hydraulic pumps, oil cylinders, and hydraulic cycloid motors. The entire hydraulic module is an open-type single-pump multi-actuator hydraulic system. The cleaning control module controls and reverses the mechanism, including completing the lifting and lowering of the sweeping disc. , the rotation of the sweeping disc, the lifting and lowering of the front-guided cleaning mechanism, the lifting of the rear-guided cleaning mechanism, the tipping of the garbage bin, the opening and closing of the garbage bin door, and the overturning of the grooved rail cleaning device. In this system, the oil pressure sensor can be used to measure the oil pressure, and the oil pressure sensor can be installed on the main oil passage of the engine on the side of the engine to monitor the oil supply pressure of the engine lubrication system, judge whether the lubrication system works normally, and prevent the engine from being caused by lack of oil. damage.

Further, the rail garbage cleaning system includes an auxiliary engine, which is used to drive the hydraulic pump and the fan, and can retract the working device and lift the garbage box to meet the power requirements of the fan and the hydraulic pump. If the water temperature of the auxiliary engine is too high, it will be May cause the auxiliary engine to stop working;

The rail garbage cleaning system also includes an auxiliary engine monitoring module, specifically a temperature sensor, and a DS18B20 digital temperature sensor can be selected, which is placed in the water tank of the auxiliary engine to measure the water temperature of the auxiliary engine water tank, and each pin of the temperature sensor is used. Heat shrink tubing is spaced to prevent short circuits, and the interior is sealed to prevent water and moisture.

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

The positioning module can choose an ultrasonic sensor. During the field work of the cleaning vehicle, the ultrasonic sensor emits high-frequency ultrasonic waves of 25-40 kHz through the crystal oscillator, and then the frequency of the reflected wave is detected by the control module. If there is an object moving in the area, The frequency of the reflected wave will fluctuate slightly, that is, the Doppler effect, so as to measure the distance between the cleaning vehicle and the obstacle and transmit it to the industrial computer system in real time. When it is detected that the distance from the obstacle is less than the predetermined safety distance, the corresponding alarm device is automatically activated to prompt the staff to pay attention to driving safety or make the cleaning vehicle take emergency braking.

Installing a positioning module (such as GPS) on the cleaning vehicle is conducive to understanding the real-time location of the cleaning vehicle, specifically recording the route of each work of the cleaning vehicle, and facilitating the remote control center to understand the status of the cleaning vehicle. In the event of an accident, the precise location of the cleaning vehicle can be determined as soon as possible.

Based on the rail garbage cleaning system including the above modules, the industrial computer receives data and processes it as follows:

According to the actual situation and the comparison of various data transmission methods, GPRS (General Packet Radio Service) can be selected as the method of data transmission, which is responsible for sending data information to the industrial computer module.

When using continuous real-time data transmission, it is necessary to design the transmission data format. In the process of data transmission, the TCP/IP protocol (Internet Protocol Suite) is used to achieve the purpose of high data transmission frequency and complete and reliable transmission to the industrial computer.

The industrial computer needs to quickly and accurately identify different types of data transmitted from the lower computer. Identify the real-time data of the air pressure sensor, liquid level transmitter, oil pressure sensor, temperature sensor, and ultrasonic sensor, respectively, and transfer them to the database to check and judge the data. If the parameters are abnormal, an alarm message will be displayed on the interface. The staff dealt with it in a timely manner.

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

The design of the condition monitoring interface on the industrial computer can be further designed based on the MFC project of the above-mentioned trough rail garbage precise 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. In the main dialog box that has been designed in the above MFC project, create a new sub-thread window, and establish a basic status monitoring interface framework in this sub-window, add the required controls, and then add the corresponding control response function to call The API interface functions related to database operation import the data stored in the database, so as to display the dynamic change diagram of each parameter curve through this interface, and use the expert evaluation system in the system evaluation method to evaluate these parameters and their change curves and knowledge base. The rules in the comparison analysis. According to the warning line preset by each parameter in the normal working state of the cleaning vehicle, the system gives information about the operation safety prompts and warnings of the cleaning vehicle to ensure the safe operation of the cleaning vehicle.

Integrate two major applications that integrate multiple functions into one control panel, reduce the operation steps in the application control process, save the user's operation time during the cleaning device control process, and make the control process more convenient; at the same time, the various modules of the cleaning device are divided into labor Clear, collaborative operation makes the cleaning process more efficient and intelligent.

Example 4

A method for allocating orbital garbage cleaning resources, comprising the following steps:

Divide a track into several segments, and each segment is equally divided into several points

x _ij (i=1,2,...,n,j=1,2,...,m);

When the cleaning vehicle starts from the starting point of the track at a fixed speed, reset the count coefficient count=1, and the timer starts to work;

Define the identification coefficient flag:

During the traveling of the cleaning vehicle, if garbage is identified, then flag=1, and record the position coordinate y _count1 at this time;

Continue to travel, if no garbage is identified, flag=0, record the position coordinate y _count0 at this time, and add 1 to count.

Count the location of garbage in each track of the day;

Continuously record daily data, use several days of data to count the probability mean of each sampling point of each segment of the track, and allocate cleaning resources according to the season and the probability mean.

统计用水量Q＝P_ij*Z_s，Z_S表示季节S时的用水量。Further, input the data of the collection points in the past week and the seasonal amount S (S=1, 2, 3, and 4 represent spring, summer, autumn, and winter, respectively) before the configuration, and the statistics of the garbage occurrence positions of each track in a day. , calculate the mean value of garbage probability

Statistical water consumption Q=P _ij *Z _s , Z _S represents the water consumption in season S.

The above-mentioned embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited by the above-mentioned embodiments, and any other changes, modifications, substitutions, combinations, The simplification should be equivalent replacement manners, which are all included in the protection scope of the present invention.

Claims (9)

1. a method for identifying orbital garbage, comprising the steps of: Get the track video stream; The video stream is decomposed frame by frame, and each frame of image is preprocessed to filter image noise. The preprocessing includes Cohen-Sutherland cropping, the preprocessed image is input into the trained garbage recognition model, image sample feature extraction, multi-scale Prediction, bounding box prediction to identify garbage; The spam identification model is trained by the following steps: Collect orbital garbage image samples; Repeatedly input garbage image samples into the YOLOv3 network structure, and obtain the garbage identification model through image sample feature extraction, multi-scale prediction, bounding box prediction and training process; The YOLOv3 network structure adopts a full convolution structure, including 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 set between the residual network layers of the feature extraction module with the same feature map size and the same number of convolution kernels; The multi-scale prediction module includes sequentially arranging three convolutional layers, subdividing the intermediate convolutional layer into a first part and a second part, compressing the network of the first part, and truncating the connection between the second part of the network whose weight is less than the set value. , obtain a sparsely connected network and retrain it, use weights to share the weights of quantized connections, and then compress the quantized weights and codebooks using a combination of predictive coding and run-length coding; Sort the identified garbage through multi-label classification; The bounding boxes of the same garbage detected multiple times are removed by non-maximum suppression. 2 . The method for identifying orbital garbage according to claim 1 , wherein the Cohen-Sutherland cropping divides the image into three areas with the orbit in the image as the boundary, and retains the middle area. 3 . 3 . The method for identifying orbital garbage according to claim 1 , wherein the preprocessing of the image further comprises grayscale and Gaussian filtering of the image; during Gaussian filtering, the horizontal size of the image is increased. 4 . 4. A method for cleaning rail garbage, characterized in that, after the garbage is identified by the method for identifying rail garbage according to any one of claims 1-3, the cleaning system is driven to complete the cleaning work. 5 . The method for cleaning rail debris according to claim 4 , wherein the driving instruction is obtained through an operator physical instruction or an instruction library call. 6 . 6. An orbital garbage cleaning system, characterized in that the system uses the orbital garbage identification method described in any one of claims 1-3 to identify garbage, comprising: The cleaning module is a cleaning vehicle, which is used to obtain control instructions to complete the cleaning work; The cleaning control module is connected with other modules of the track garbage cleaning system and controls the operation of other modules, including the MFC program command input unit and the physical command input unit. 7. The rail trash cleaning system according to claim 6, wherein: The bottom of the front of the cleaning vehicle of the cleaning module is provided with a dust suction port, and a clean water tank and a sewage tank are arranged in the compartment of the cleaning vehicle; The orbital trash cleaning system also includes: The wind pressure monitoring module is located at the suction port to measure the pressure at the suction port; The waterway monitoring module is used to detect the water level of the clean water tank and the sewage tank respectively. 8. The rail garbage cleaning system according to claim 7, wherein the waterway monitoring module is a liquid level transmitter, installed on the bottom of the side of the clean water tank or the sewage tank, and the opening position and the water outlet of the water pipe are on the same plane . 9. A method for disposing of orbital garbage cleaning resources, characterized in that, identifying garbage using the orbital garbage identification method described in any one of claims 1-3, comprising the steps of: Divide a track into several segments, and each segment is equally divided into several sampling points x _ij (i=1, 2,...,n, j=1, 2,...,m);

When the cleaning vehicle starts from the starting point of the track at a fixed speed, reset the count coefficient count=1, and the timer starts to work; Define the identification coefficient flag:

During the traveling of the cleaning vehicle, if garbage is identified, then flag=1, and record the position coordinate y _count1 at this time; Continue to move, if no garbage is identified, then flag=0, record the position coordinate y _count0 at this time, and add 1 to count; Count the location of garbage in each track of the day; Continuously record daily data, use several days of data to count the probability mean of each sampling point of each segment of the track, and allocate cleaning resources according to the season and the probability mean.

Images