A trolley tracking method based on a mobilenet v2 model and a grayscale difference ratio

Abstract

The application discloses a trolley tracking method based on a mobilenet V2 model and a gray difference ratio, and the trolley tracking method can recognize routes of various color lines as long as the target tracking route and the environmental background color difference are obvious, for example, a white color is used as a bottom plate, and routes of black color, blue color and red color can be well recognized, and when a cross road or a T-shaped road is faced, an identification can be automatically recognized, and it is judged that the next step is left turning, right turning or straight going. The tracking method has strong anti-interference ability and small calculation amount of tracking, can adapt to various scenes and illumination conditions, and has high flexibility.

Description

Technical Field

[0001] This invention relates to a car tracking method based on the Mobilenet v2 model and grayscale difference ratio. Background Technology

[0002] In university laboratories, the handling of electronic components and other small items is common. Existing intelligent transport carts in laboratories use infrared or grayscale sensors for tracking. While these methods are low-cost, they cannot correctly execute transport commands at intersections or T-junctions, are easily affected by external environmental factors (such as lighting), and have poor anti-interference capabilities. Therefore, they have high requirements for the site and their application range is greatly limited. Summary of the Invention

[0003] This invention provides a car tracking method based on the Mobilenet v2 model and grayscale difference ratio to solve the problems existing in the prior art.

[0004] The technical solutions adopted in this invention are as follows:

[0005] A car tracking method based on the Mobilenet v2 model and grayscale difference ratio includes the following steps:

[0006] S1: Place room number signs outside each laboratory, and set up directional signs leading to the corresponding laboratory at each crossroads or T-junctions. At the same time, assign a unique intersection number to each crossroads or T-junction. Copy all the images of the room number signs, directional signs, and intersection numbers to the root directory of the OpenMV camera's SD card. Then, stick black tape on the path leading to each laboratory.

[0007] S2: Input the nameplate of the laboratory room number of the expected destination into the OpenMV camera. The OpenMV camera determines the intersection where the laboratory is located based on the input nameplate. The MT9V034 main drilling machine captures images of the black tape on the ground to control the vehicle's trajectory. When the vehicle reaches a crossroads or T-junction, the OpenMV camera first identifies the intersection number to determine if the laboratory is located at that intersection. If it is, the OpenMV camera identifies the directional sign of the laboratory at that intersection to determine the direction of travel and turn. After turning, the vehicle continues to travel by capturing images of the black tape on the ground using the MT9V034 main drilling machine until the OpenMV camera identifies that the nameplate of the laboratory room number of the destination matches the input nameplate of the expected destination laboratory. Then, it stops and waits for the item to be retrieved. After the item is retrieved, it returns along the original route.

[0008] If the car is not at that intersection, it continues to the next crossroads or T-junction until it finds the intersection where the laboratory is located.

[0009] Furthermore, the specific process of controlling the vehicle's trajectory by acquiring images of the black tape on the ground using the main drilling rig MT9V034 camera includes:

[0010] 1) Acquire the image;

[0011] Black tape was applied to the white ground. After removing the obstructions covering the black tape, the MT9V034 camera captured and collected images with a resolution of 120*188 pixels as the car moved.

[0012] 2) Determine the left and right boundary points of the black tape;

[0013] Select one pixel as point A from all pixels at the vertical center of the image. The row containing point A is the starting row. Select point B in the starting row, which is 4 pixels to the left of point A. Based on the gray values ​​of point A and point B at this time, calculate the difference ratio C for the first time using the formula C = abs(100*(AB) / (A+B)).

[0014] Then, points A and B are simultaneously shifted to the left of the starting line, with each pixel as an offset unit. After each shift, the difference ratio C is calculated again using the formula C = abs(100*(AB) / (A+B)) based on the gray values ​​of points A and B after the shift. This process is repeated until point B traverses all pixels horizontally from the vertical center of the image to the left along the starting line.

[0015] During the horizontal traversal to the left, the sum of the difference ratios C will suddenly increase twice. The points where the difference suddenly increases are the two boundary points of the black tape. At this time, the black tape is located on the left side of the image.

[0016] After traversing horizontally from the vertical center of the image to the left between points A and B, the sum of the differences C does not show two sudden increases. In the same way, traverse from the vertical center of the image to the right along the starting row between points A and B and find the points where the sum of the differences C suddenly increases twice. This will show that the black tape is located at the two boundary points on the right side of the image.

[0017] If, after traversing horizontally from the vertical center of the image to the left, the sum of the differences C only shows one sudden increase, and the sum of the differences C also only shows one sudden increase when traversing horizontally to the right, then the black tape is located in the middle of the image.

[0018] 3) Obtain all boundary points of the black tape in the image;

[0019] After obtaining the boundary point of the black tape in the starting row, the midpoint D0 of the black tape in the starting row is obtained based on the coordinates of the boundary point of the starting row.

[0020] Then, the midpoint D0 is moved vertically upwards for the first time, with an offset unit of one pixel, to obtain the midpoint D1 of the black tape in the row of pixels above the starting row. Then, let D1 be point A, and take one pixel each from the left and right of D1 as point B. Similarly, using the formula C = abs(100*(AB) / (A+B)), traverse the pixels to the left and right of the row where D1 is located to obtain the two boundary points of the black tape in the row where D1 is located. Then, based on these boundary points, the actual midpoint of the black tape in the row where D1 is located is obtained again. Then, the obtained midpoint is moved vertically upwards for the second time, with an offset unit of one pixel, and so on, to find all the boundary points of the black tape above the starting row in the image and the midpoint of the black tape in each row of pixels, and obtain the coordinates of the midpoint of the black tape in each row of pixels in the image.

[0021] 4) Determine the direction of the trolley's offset;

[0022] Let the row of pixels in the image containing the midpoint of the black tape be the valid pixel row.

[0023] The column coordinates of the midpoint of each pixel in the image are successively subtracted from the 94 column coordinates of the image. Then, all the subtracted values ​​are added together and divided by the number of effective pixel rows to obtain an average deviation. If the average deviation is greater than 0, the car is moving to the right; if it is less than 0, it is moving to the left.

[0024] When the car approaches an intersection or T-junction, the OpenMV camera identifies the directional signs at the intersection or T-junction to control the car's steering.

[0025] Furthermore, after finding the left or right boundary point of the black tape using the difference ratio and value C, it is necessary to eliminate the deviation between the found left and right boundary points. The method for eliminating the deviation is as follows:

[0026] For the right boundary point of the black tape, first take the point where the C value suddenly increases as the theoretical right boundary point of the black tape. Then, select 5 pixels horizontally to the right from the theoretical right boundary point to form six adjacent comparison points. Subtract the six comparison points in pairs and find the point with the largest absolute difference. This point is the actual right boundary point of the black tape.

[0027] For the left boundary point of the black tape, first take the point where the C value suddenly increases as the theoretical left boundary point of the black tape. Then, starting from the theoretical left boundary point, select 5 pixels horizontally to the left to form 6 adjacent comparison points. Find the actual left boundary point of the black tape by subtracting the gray values ​​in the same way.

[0028] Furthermore, after vertically moving the actual midpoint of the black tape one pixel upwards, when searching for the left or right boundary point of the moved-up black tape using the formula C = abs(100*(AB) / (A+B)), stop traversing other pixels in the next row as soon as the sum of the difference C suddenly increases.

[0029] Further, row 116 of the initial behavior image.

[0030] The present invention has the following beneficial effects:

[0031] The vehicle tracking method of this invention can identify routes of various colors as long as there is a significant color difference between the target route and the background environment. For example, using white as the background, it can effectively identify black, blue, and red routes. Furthermore, when facing intersections or T-junctions, it can automatically recognize signs and determine whether the next step is a left turn, right turn, or straight ahead. The tracking method of this invention has strong anti-interference capabilities and low computational load, adaptable to various scenarios and lighting conditions, and possesses high flexibility. Attached Figure Description

[0032] Figure 1 This is a hardware diagram used in this invention.

[0033] Figure 2 This is a schematic diagram of the laboratory's layout.

[0034] Figure 3 The accuracy rate for each type of laboratory identification.

[0035] Figure 4 The grayscale difference ratio is used to obtain a schematic diagram of the boundary points of the black tape.

[0036] Figure 5 The image shows the black tape located on the right side of the image.

[0037] Figure 6 The image shows the black tape positioned in the center.

[0038] Figure 7 This is a schematic diagram showing how to find the boundary points of the tape at the midpoint of each row of pixels in an image using black tape. Detailed Implementation

[0039] The invention will now be further described with reference to the accompanying drawings.

[0040] likeFigure 1 This invention discloses a car tracking method based on the MobileNet v2 model and grayscale difference ratio. The hardware for implementing the tracking method mainly includes a power supply module, an STM32F103RCT6 microcontroller, an OpenMV camera, and a MT9V034 camera.

[0041] The power supply module is used to supply power to the STM32F103RCT6 microcontroller, the OpenMV camera, the MT9V034 drilling camera, and the vehicle's power module.

[0042] The STM32F103RCT6 microcontroller is used to process data from the OpenMV camera and the MT9V034 drilling machine camera. After processing all the data, it controls the power module of the trolley.

[0043] OpenMV cameras are used to determine whether a person has reached their destination at intersections or T-junctions, and then feed back the result (left turn, right turn, go straight, or return along the same route) to the STM32F103RCT6 microcontroller for processing.

[0044] The MT9V034 camera is used to acquire images of the road ahead. After processing the images by grayscale difference ratio, the data is fed back to the STM32F103RCT6 microcontroller for processing.

[0045] The method of the present invention will be described in detail below.

[0046] The tracking method of the present invention includes:

[0047] S1: Place room number signs outside each laboratory, and set up directional signs leading to the corresponding laboratory at each crossroads or T-junctions. At the same time, assign a unique intersection number to each crossroads or T-junction. Copy all the images of the room number signs, directional signs, and intersection numbers to the root directory of the OpenMV camera's SD card. Then, stick black tape on the path leading to each laboratory.

[0048] S2: Input the nameplate of the laboratory room number of the expected destination into the OpenMV camera. The OpenMV camera determines the intersection where the laboratory is located based on the input nameplate. The MT9V034 main drilling machine captures images of the black tape on the ground to control the vehicle's trajectory. When the vehicle reaches a crossroads or T-junction, the OpenMV camera first identifies the intersection number to determine if the laboratory is located at that intersection. If it is, the OpenMV camera identifies the directional sign of the laboratory at that intersection to determine the direction of travel and turn. After turning, the vehicle continues to travel by capturing images of the black tape on the ground using the MT9V034 main drilling machine until the OpenMV camera identifies that the nameplate of the laboratory room number of the destination matches the input nameplate of the expected destination laboratory. Then, it stops and waits for the item to be retrieved. After the item is retrieved, it returns along the original route.

[0049] If the car is not at that intersection, it continues to the next crossroads or T-junction until it finds the intersection where the laboratory is located.

[0050] like Figure 2 The diagram shows the layout of the laboratories. Taking six laboratories as an example, two directional signs are placed at intersection 1: the left end of intersection 1 is the directional sign for laboratory 1, and the right end is the directional sign for laboratory 2; similarly, the left end of intersection 2 is the directional sign for laboratory 3, and the right end is the directional sign for laboratory 4; the left end of intersection 3 is the directional sign for laboratory 5, and the right end is the directional sign for laboratory 6.

[0051] If the intended laboratory is Laboratory 5 (i.e. Figure 2 If the lab is located at the top left corner, the room number of lab 5 will be identified and recorded by the OpenMV camera on the car. After the OpenMV camera identifies the room number of lab 5, it will determine that lab 5 is located at intersection 3.

[0052] During the transport of experimental equipment, the OpenMV camera first identifies intersection signs (i.e., intersection numbers). Upon identifying sign "Intersection 1," based on pre-determined information (i.e., Laboratory 5 is located at Intersection 3), the car proceeds straight through the intersection. Upon identifying sign "Intersection 2," it similarly proceeds straight through the intersection. When it identifies sign "Intersection 3," the OpenMV camera determines that Laboratory 5 is located at that intersection, and the car stops. Then, the OpenMV camera begins to identify the directional signs at the intersections. Based on the location of the directional sign at Intersection 3 containing the words "Laboratory 5," it determines which side of Intersection 3 the destination Laboratory 5 is on, and then controls the car's direction accordingly. Figure 2Turn left from the center. After turning left, if the OpenMV camera recognizes a room number plaque with the words "Laboratory 5" at the entrance of Laboratory 5, the car will stop.

[0053] This invention copies the weight model obtained by training all room number nameplate images and directional nameplate images using EIQ to the root directory of the OpenMV camera's SD card, specifically:

[0054] When building up data from OpenMV cameras, first prepare basic data including various fonts for "Laboratory 1", "Laboratory 2", "Laboratory 3", "Laboratory 4", "Laboratory 5", "Laboratory 6", "Intersection 1", "Intersection 2", and "Intersection 3" to facilitate training of the MobileNet v2 lightweight convolutional neural network. After preparation, write a data augmentation script to augment the above data and save it in the pictureENH folder. "Laboratory 1" refers to the room number plaques and directional plaques, i.e., the identification of the corresponding laboratory.

[0055] The specific enhancement method involves rotating the image 15 degrees horizontally, then adding a brightness variation from dark to light and Gaussian noise and salt-and-pepper noise on top of the rotation. The enhanced data is then converted into an .npy file and labeled with training tags, namely ['lab1','lab2','lab3','lab4','lab5','lab6','lukou1','lukou2','lukou3'].

[0056] To address this, a file named `make_dataset.py` was created to convert images into .npy data. The specific conversion process is as follows:

[0057] First, the `pictureENH` folder is traversed to generate an iteration directory. Then, images are read using the OpenCV library and their formats are checked. If the image format is correct, a resize operation is performed to adjust the pixel size of each image, ensuring a uniform size for subsequent model training. Next, the uniformly sized image data is added sequentially to a cache list, along with a label array. After the iteration process is complete, the data array and label array are stored as `.npy` files. `x.npy` stores the image data, and `y.npy` stores the label data.

[0058] The following describes how to create the EIQ project. In the previous section, we prepared the .npy data (the standard data format of the NumPy library in Python). Then, using the `converters.pyi` open-source file in the EIQ project directory of the NXP open-source project, we converted the .npy data into the DeepView project dataset. After the conversion, we can use the EIQ open-source software to open the newly created DeepView project dataset for model training.

[0059] After training, the model's accuracy was evaluated using EIQ's built-in model evaluation tool. Once the target accuracy was achieved, the model was quantized (i.e., the data precision was reduced to facilitate deployment on platforms with limited computing power) and exported. The original float32 data was quantized using int8 precision, and the quantized weight data was exported in the tflite data storage format. Most lab markers were accurately predicted, but a small number of predictions were inaccurate. Figure 3 The accuracy of each type of laboratory identification was evaluated for this invention. After evaluation, the model was exported and named num.tflite.

[0060] The following provides a detailed explanation of the MT9V034 main drilling air camera's line-following logic.

[0061] 1) Acquire the image;

[0062] Black tape was applied to a white surface (i.e., tracking the delivery cart was performed on a white background covered with black tape). Obstacles covering the black tape were removed. To complete the tracking, a Zongzuanfeng MT9V034 grayscale camera was used to acquire images. The grayscale image acquired by the Zongzuanfeng MT9V034 grayscale camera had a resolution of 120*188 pixels. Figure 4 Only a portion of the pixel grid is drawn in the image;

[0063] 2) Determine the left and right boundary points of the black tape;

[0064] Because the black tape on the delivery cart is relatively thin and long, and the direction of line tracing is from near to far, the first consideration in algorithm design is to identify the path boundary points of the bottom row of the image. Simply put, path boundary points are the edge points between the black tape and the white background. For example... Figure 4 In the acquired grayscale image, it can be observed that there are more obstructions in the distance of the delivery vehicle's field of vision, but the gradient distortion is lower and there are almost no obstructions in the area closer to the delivery vehicle's field of vision, resulting in more usable effective paths. Therefore, after observing the acquired grayscale image, it was confirmed that the direction of line finding is from near to far. Thus, this invention selects row 116 of the image as the starting line.

[0065] Select one pixel from all pixels at the vertical center of the image as point A, i.e. Figure 4 94 columns ( Figure 4 The K-line is a line connecting all 94 columns of pixels. The row where point A is located is the starting row. Point B is selected in the starting row and is 4 pixels to the left of point A. Based on the gray values ​​of point A and point B at this time, the difference ratio C is calculated for the first time using the formula C = abs(100*(AB) / (A+B)).

[0066] Then, points A and B are simultaneously shifted to the left of the starting line, with each pixel as an offset unit. After each shift, the difference ratio C is calculated again using the formula C = abs(100*(AB) / (A+B)) based on the gray values ​​of points A and B after the shift. This process is repeated until point B traverses all pixels horizontally from the vertical center of the image to the left along the starting line.

[0067] During the horizontal traversal to the left, the sum of the differences C will experience two sudden increases. These two points of sudden increase are the two boundary points of the black tape. At this time, the black tape is located on the left side of the image, as shown below. Figure 4 .

[0068] After traversing horizontally from the vertical center of the image to the left between points A and B, the sum of the differences C did not show two sudden increases. Therefore, in the same manner, traverse from the vertical center of the image to the right along the starting row between points A and B, finding the points where the sum of the differences C suddenly increases. This reveals the two boundary points on the right side of the image where the black tape is located. Figure 5 .

[0069] If, after traversing horizontally from the vertical center of the image to the left, the sum of the differences C between points A and B only experiences one sudden increase, and the sum of the differences C also only experiences one sudden increase when traversing horizontally to the right, then the black tape is located in the middle of the image. Figure 6 .

[0070] Taking the pixels on the left side of the image where the black tape is located as an example, the grayscale values ​​of rows 116 to 112 and columns 64 to 80 are shown in the table below.

[0071]

[0072]

[0073] If we take column 80 as point A, then column 75 is point B. Then we can calculate C using abs(100*(AB) / (A+B)).

[0074] Then, taking column 79 as point A, and column 74 as point B, we calculate point C using abs(100*(AB) / (A+B)). This process continues. Finally, the calculation shows that column 78 is a boundary point of the black tape.

[0075] 3) Obtain all boundary points of the black tape in the image;

[0076] This invention requires determining the valid portion of a path in an image to complete tracking for each frame of the camera image. Specifically:

[0077] like Figure 7 After obtaining the boundary point of the black tape in the starting row (row 116), the midpoint D0 of the black tape in the starting row is obtained based on the coordinates of the boundary point. Then, the midpoint D0 is moved vertically upwards for the first time (row 115) with an offset unit of one pixel, and the midpoint D1 of the tape width in the row of pixels above the starting row is obtained. Then, let D1 be point A, and take one pixel each from the left and right of D1 as point B. Similarly, according to the formula C = abs(100*(AB) / (A+B) Iterate through the left and right pixels of the row containing D1 to find the two boundary points of the black tape in the row containing D1. Then, based on these boundary points, find the actual midpoint of the black tape in the row containing D1. Then, move the obtained midpoint vertically upwards again by one pixel as an offset unit. By doing so, find all the boundary points of the black tape above the starting row and the midpoint of the black tape in each pixel row in the image, and obtain the coordinates of the midpoint of the black tape in each pixel row in the image.

[0078] To save computation time: After vertically moving the actual midpoint of the black tape one pixel upwards, when using the formula C = abs(100*(AB) / (A+B)) to find the left or right boundary point of the moved-up black tape, if the sum of the differences C suddenly increases, stop traversing and move to other pixels in the next row. Figure 7 In the process, after finding point A on the left side of point A in D1, we stop searching to the left and the same applies to the right side.

[0079] 4) Determine the direction of the trolley's offset;

[0080] Let the row of pixels with the midpoint of the black tape in the image be the effective pixel row. Take the column coordinates of the midpoint of each pixel row of the black tape in the image and subtract them from the 94 column coordinates of the image. Then add all the subtraction values ​​together and divide by the number of effective pixel rows to get an average deviation. If the average deviation is greater than 0, the car is moving to the right; if it is less than 0, it is moving to the left.

[0081] When the car approaches an intersection or T-junction, the OpenMV camera identifies the directional signs at the intersection or T-junction to control the car's steering.

[0082] After finding the left or right boundary point of the black tape using the difference ratio and value C, it is necessary to perform deviation elimination on the found left and right boundary points respectively. The deviation elimination method is as follows:

[0083] During the process of continuously shifting points A and B left (right) and calculating C, it was found that the value of C suddenly increases near the boundary point. Based on this characteristic, since the difference calculation is performed every 4 pixels, the currently found pixel is not accurate enough. Therefore, the grayscale values ​​of the next 5 pixels are subtracted, and the point with the largest absolute difference is found. This point is the actual boundary point of the black tape. Details are as follows:

[0084] For the right boundary point of the black tape, first take the point where the C value suddenly increases as the theoretical right boundary point of the black tape. Then, select 5 pixels horizontally to the right from the theoretical right boundary point to form six adjacent comparison points. Subtract the six comparison points in pairs and find the point with the largest absolute difference. This point is the actual right boundary point of the black tape.

[0085] For the left boundary point of the black tape, first take the point where the C value suddenly increases as the theoretical left boundary point of the black tape. Then, starting from the theoretical left boundary point, select 5 pixels horizontally to the left to form 6 adjacent comparison points. Find the actual left boundary point of the black tape by subtracting the gray values ​​in the same way.

[0086] Taking the black tape located on the left side of the image as an example, when the black tape is on the left side of the image, during the process of shifting points A and B from the vertical center of the image to the left and calculating the difference ratio C, if the value of C suddenly increases, the point where the value suddenly increases is the theoretical right boundary point of the black tape. Then, starting from the theoretical right boundary point, select 5 pixels horizontally to the right, thus forming 6 adjacent comparison points. Subtract each pair of adjacent comparison points to find the point with the largest absolute difference, which is the actual right boundary point of the black tape on the left side of the image. Continue to find the theoretical left boundary point on the same horizontal line as the actual right boundary point by calculating the difference ratio C. Then, starting from the theoretical left boundary point, select 5 pixels horizontally to the left and find the actual left boundary point of the black tape by subtracting the gray values ​​in the same way.

[0087] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several improvements without departing from the principle of the present invention, and these improvements should also be considered within the scope of protection of the present invention.

Claims

1. A car tracking method based on the MobileNet v2 model and grayscale difference ratio, characterized in that: Includes the following steps: S1: Place room number signs marked with the laboratory room number outside each laboratory, set up directional signs leading to the corresponding laboratory at each crossroads or T-junctions, and assign a unique intersection number to each crossroads or T-junctions. Copy all the images of the room number signs, directional signs, and intersection numbers to the root directory of the OpenMV camera's SD card, and then stick black tape on the path leading to each laboratory. S2: Input the nameplate of the laboratory room number of the expected destination into the OpenMV camera. The OpenMV camera determines the intersection where the laboratory is located based on the input nameplate. The MT9V034 main drilling machine captures images of the black tape on the ground to control the vehicle's trajectory. When the vehicle reaches a crossroads or T-junction, the OpenMV camera first identifies the intersection number to determine if the laboratory is located at that intersection. If it is, the OpenMV camera identifies the directional sign of the laboratory at that intersection to determine the direction of travel and turn. After turning, the vehicle continues to travel by capturing images of the black tape on the ground using the MT9V034 main drilling machine until the OpenMV camera identifies that the nameplate of the laboratory room number of the destination matches the input nameplate of the expected destination laboratory. Then, it stops and waits for the item to be retrieved. After the item is retrieved, it returns along the original route. If the car is not at that intersection, it continues to the next crossroads or T-junction until it finds the intersection where the laboratory is located. The specific process of controlling the car's trajectory by acquiring images of black tape on the ground using the MT9V034 main drilling air camera includes: 1) Acquire the image; Black tape was applied to the white ground. After removing the obstructions covering the black tape, the MT9V034 camera captured and collected images with a resolution of 120*188 pixels as the car moved. 2) Determine the left and right boundary points of the black tape; Select one pixel as point A from all pixels at the vertical center of the image. The row containing point A is the starting row. Select point B in the starting row, which is 4 pixels to the left of point A. Based on the gray values ​​of point A and point B at this time, calculate the difference ratio C for the first time using the formula C=abs(100*(AB) / (A+B)). Then, points A and B are simultaneously shifted to the left of the starting line, with each pixel as an offset unit. After each shift, the difference ratio C is calculated again using the formula C=abs(100*(AB) / (A+B)) based on the gray values ​​of points A and B after the shift. This process is repeated until point B traverses all pixels horizontally from the vertical center of the image to the left along the starting line. During the horizontal traversal to the left, the sum of the difference ratios C will suddenly increase twice. The points where the difference suddenly increases are the two boundary points of the black tape. At this time, the black tape is located on the left side of the image. After traversing horizontally from the vertical center of the image to the left between points A and B, the sum of the differences C does not show two sudden increases. In the same way, traverse from the vertical center of the image to the right along the starting row between points A and B and find the points where the sum of the differences C suddenly increases twice. This will show that the black tape is located at the two boundary points on the right side of the image. If, after traversing horizontally from the vertical center of the image to the left, the sum of the differences C only shows one sudden increase, and the sum of the differences C also only shows one sudden increase when traversing horizontally to the right, then the black tape is located in the middle of the image. 3) Obtain all boundary points of the black tape in the image; After obtaining the boundary point of the black tape in the starting row, the midpoint D0 of the black tape in the starting row is obtained based on the coordinates of the boundary point of the starting row. Then, the midpoint D0 is moved vertically upwards for the first time, with an offset unit of one pixel, to obtain the midpoint D1 of the black tape in the row of pixels above the starting row. Then, let D1 be point A, and take one pixel each from the left and right of D1 as point B. Similarly, using the formula C=abs(100*(AB) / (A+B)), traverse the left and right pixels of the row containing D1 to obtain the two boundary points of the black tape in the row containing D1. Then, based on these boundary points, obtain the actual midpoint of the black tape in the row containing D1. Then, the obtained midpoint is moved vertically upwards for the second time, with an offset unit of one pixel, and so on, to find all the boundary points of the black tape above the starting row and the midpoint of the black tape in each row of pixels in the image, and obtain the coordinates of the midpoint of the black tape in each row of pixels in the image. 4) Determine the direction of the trolley's offset; Let the row of pixels in the image containing the midpoint of the black tape be the valid pixel row. The column coordinates of the midpoint of each pixel in the image are successively subtracted from the 94 column coordinates of the image. Then, all the subtracted values ​​are added together and divided by the number of effective pixel rows to obtain an average deviation. If the average deviation is greater than 0, the car is moving to the right; if it is less than 0, it is moving to the left. When the car approaches an intersection or T-junction, the OpenMV camera identifies the directional signs at the intersection or T-junction to control the car's steering.

2. The car tracking method based on the MobileNet v2 model and grayscale difference ratio as described in claim 1, characterized in that: After finding the left or right boundary point of the black tape using the difference ratio and value C, it is necessary to perform deviation elimination on the found left and right boundary points respectively. The deviation elimination method is as follows: For the right boundary point of the black tape, first take the point where the C value suddenly increases as the theoretical right boundary point of the black tape. Then, select 5 pixels horizontally to the right from the theoretical right boundary point to form six adjacent comparison points. Subtract the six comparison points in pairs and find the point with the largest absolute difference. This point is the actual right boundary point of the black tape. For the left boundary point of the black tape, first take the point where the C value suddenly increases as the theoretical left boundary point of the black tape. Then, starting from the theoretical left boundary point, select 5 pixels horizontally to the left to form 6 adjacent comparison points. Find the actual left boundary point of the black tape by subtracting the gray values ​​in the same way.

3. The car tracking method based on the MobileNet v2 model and grayscale difference ratio as described in claim 1, characterized in that: After moving the actual midpoint of the black tape vertically upwards by one pixel, when searching for the left or right boundary point of the moved-up black tape using the formula C=abs(100*(AB) / (A+B)), stop traversing other pixels in the next row as soon as the sum of the differences C suddenly increases.

4. The car tracking method based on the MobileNet v2 model and grayscale difference ratio as described in claim 1, characterized in that: The 116th row of the initial behavior image.

Images