A loading station chute automatic loading method based on monocular vision recognition

Abstract

The present application relates to a kind of automatic loading method of loading station chute based on monocular vision identification, comprising: identifying car height;Chute swing to initial position;Identify chute mouth distance and height;Chute swing extends to coal pressing position;Chute swing retracts to flat coal position;Judge whether chute mouth is scraped and bumps car;Judge whether whole column loading is finished.The present application uses monocular vision to collect image information, without prior knowledge of camera calibration, without explicit relationship between camera and environment, determines the affine scene structure from the image, completes image recognition, has the function of identifying car height, relative distance between chute mouth and car front and rear side and relative height with car roof plane, thereby replacing people to complete certain work, and machine will not be tired, work efficiency is higher.Through logical sequence, automatic loading process is completed, more manpower can be liberated, and a new solution is provided for realizing "someone supervises, no one operates" automatic loading.

Description

Technical Field

[0001] This invention relates to an automatic loading method for railway loading stations based on monocular vision recognition. It is a method for automatically acquiring signals and loading loading machinery, and an automatic detection and control method for transporting bulk materials at railway loading stations. Background Technology

[0002] Rapid quantitative loading stations play a crucial role in bulk material transportation, and their loading efficiency directly impacts the station's loading capacity. Currently, the level of automation in loading in my country needs improvement. Most chute control relies on manual loading; a few use optical gratings or lidar to acquire motion signals for automated loading. For manual loading, operators observe whether the wagon is in position to open the unloading gate and activate the chute's swing mechanism. If uneven loading occurs, the operator manually adjusts the chute height using the control handle to ensure even loading. Crucially, operators need to carefully observe the wagon's position and height, and loading efficiency varies depending on the operator, resulting in low efficiency. For optical grating and lidar-based loading, the main focus is on detecting the wagon's position, and a cable sensor detects the hydraulic cylinder's position to determine if it has reached the set location. It's evident that current automated loading processes are relatively simple, only detecting the wagon's own position and lacking a solution that integrates the wagon's side opening and chute opening positions. Therefore, it is necessary to develop a new automated loading solution. With the continuous development of computer technology and the integration of the requirements for "automatic loading" and "rapid loading" put forward by railway loading yards, machine vision technology is increasingly being applied. It can replace humans in completing certain tasks, and the machines do not get fatigued, resulting in high work efficiency. Therefore, how to combine machine vision, image processing and recognition technology with automated loading at loading stations to achieve the recognition of the relationship between the car side opening and the chute opening is an urgent problem to be solved. Summary of the Invention

[0003] To overcome the problems of existing technologies, this invention proposes an automated loading method for loading stations using monocular vision recognition. The method acquires image information through monocular vision, identifies planned parameters, and completes the automated loading process according to a logical sequence. Based on this, not only can loading efficiency be improved, but more manpower can be freed up, providing a new solution for achieving automated loading with "manual supervision and unmanned operation."

[0004] The purpose of the present application is achieved in that a kind of automatic loading method of chute of loading station based on monocular vision recognition, the system used by the method includes: the chute of loading station, chute control subsystem, vision subsystem, including two parts: car height identification module and chute mouth and car side relative position identification module, and image recognition subsystem, the steps of the method are as follows:

[0005] Step 1, identify the height of the car: before the car starts loading, identify the height of the car, get the initial swing angle signal of the chute and the height of the coal level;

[0006] Step 2, the chute swings to the initial position: according to the action signal identified in step 1, the chute control system drives the chute to move, so that the chute moves to the initial loading position and waits for the next instruction;

[0007] Step 3, identify the distance and height of the chute mouth: a lightweight single vision Figure Three reconstruction fusion straight line detection algorithm is used to identify the relative distance of the chute mouth and the front and rear sides of the car and the relative height to the car roof plane, to reconstruct the car side in three dimensions and construct the car side opening plane and range; a trajectory tracking algorithm based on kinematic constraints is used to accurately estimate the position and posture of the chute mouth bottom by fusing Kalman filter, and to calculate the relative position relationship between them;

[0008] Step 4, the chute swings out to the coal pressing position: through real-time detection of the distance between the front edge of the chute and the front car side in step 3, the detection data is compared with the distance S between the front edge of the chute mouth and the front car side when the chute is extended, and after the condition S>L / cosβ-V0·L / V is met, the chute control device drives the chute to move, first swings the chute by a set angle, and then makes the chute extension section move to the coal pressing position at a set speed and a set length; wherein: V0 is the speed of the car; L is the extension length of the chute; V is the extension speed of the chute; and β is the swing angle of the chute;

[0009] Step 5, the chute swings back to the coal level: through real-time detection of step 3, when the bottom edge height of the chute mouth is at the coal level height, the image processing device sends a signal to stop the movement of the chute extension section;

[0010] Step 6, judge whether the chute mouth scratches the car side: when the distance between the rear edge of the chute and the rear car side reaches the set instruction, one car loading is completed, and if the lower coal level is lower than the roof, the chute needs to be controlled to retract to the initial position;

[0011] Step 7, judge whether the whole train loading is completed: if the next car is detected, then after step 7 is completed, the next loading cycle is started, and steps 3-7 are repeated; if no next car is detected, the loading is completed.

[0012] Further, the step 3 of identifying the distance and height of the chute mouth includes the following sub-steps:

[0013] Sub-step 1, target tracking of the chute opening: The KCF algorithm is used for target tracking of the moving chute opening. First, the target image to be tracked is transformed into the Fourier domain, and a nonlinear classifier is trained. The goal of KCF training is to find a function f(z) = ω. T z, so that sample x i and its regression target y i The squared error between them is minimized; where: ω is the model weight; T is the matrix transpose; z is the cropped image patch;

[0014] Sub-step 2: Optimize target tracking: Use a Kalman filter to further optimize object tracking, and use the Kalman filter for prediction from frame t-1 to frame t; where: t is the t-th frame of the video stream;

[0015] Sub-step 3: Solve for the vertical distance from the bottom edge of the chute to the top plane of the train car: Select four reference points to calculate the auxiliary straight line equation. Select two known points in the side plane of the chute and two points on the edge of the upper surface of the train car. Determine a straight line through the two points to obtain the straight line equation. Considering that the lowest end of the chute will insert into the train car and cause obstruction, the selection of the target tracking center point is particularly important. Calculate the distance between the target tracking center and the upper plane of the train car based on geometric relationships. Finally, correct the calculated distance.

[0016] The advantages and beneficial effects of this invention are as follows: By employing monocular vision to acquire image information, this invention eliminates the need for prior knowledge of camera calibration and a clear understanding of the relationship between the camera and the environment. It determines the affine scene structure from the image, completing image recognition. It possesses the ability to identify the height of the car body, the relative distance between the chute opening and the front and rear sides of the car body, and the relative height to the roof plane. This allows it to replace humans in performing certain tasks, and the machine does not experience fatigue, resulting in high work efficiency. By automating the loading process through a logical sequence, it can free up more manpower, providing a new solution for achieving automated loading with "manual supervision and unmanned operation." Attached Figure Description

[0017] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0018] Figure 1 This is a schematic diagram of the system structure used in the method described in Embodiment 1 of the present invention;

[0019] Figure 2 This is a schematic diagram of the system structure used in the method described in Embodiment 1 of the present invention. Figure 1 View from direction A;

[0020] Figure 3 This is a schematic diagram illustrating the principle of square pile identification using the method described in Embodiment 1 of the present invention;

[0021] Figure 4 is a flow chart of the method of the embodiment one of the present application;

[0022] Figure 5 is the calculation principle of the chute extension position of the chute of the embodiment one of the present application;

[0023] Figure 6 is the calculation principle of the chute opening recognition of the embodiment two of the present application. DETAILED DESCRIPTION

[0024] Embodiment one:

[0025] The embodiment is a loading station chute automatic loading method based on monocular vision recognition. The system used in the method comprises a loading station chute 1, a chute control subsystem, a vision subsystem, which comprises two parts: a car height recognition module 2 and a chute opening and car side relative position recognition module 3, and an image recognition subsystem, as shown in Figure 1 、 2 , 3.

[0026] The steps of the method are as follows, and the flow is as shown in Figure 4 .

[0027] Step 1: Recognize the car height: Before the car starts loading, the car height is recognized to obtain the initial swing angle signal of the chute and the flat coal position height, and the signal database is determined according to the operation experience.

[0028] The height of the car is recognized by monocular vision. First, two square piles 4 with known actual sizes are set in the vision detection scene of the car height recognition module. The sides of the two square piles are not parallel to the side of the car, and have a certain angle (the size of the angle is generally greater than 10 degrees and less than 80 degrees. The car side and the square pile side are observed with the naked eye to have a relatively obvious non-parallelism), as shown in Figure 1 、 2 , 3, which are used to determine two independent parallel lines.

[0029] The specific operation details of the recognition are as follows:

[0030] (1) Eight points need to be clicked on the image to select two independent parallel lines. Each parallel line is composed of two line segments and contains four points. The two planes determined by the two parallel lines cannot be parallel, and are both perpendicular to the common reference surface (the rail surface 5, see Figure 1 、 2 , and Figure 3The plane is indicated by double-dot line). The eight points selected as shown in Fig. 3 are A, B, C, D, E, F, G and H, which form two groups of parallel lines, AB and CD (two vertical thick solid lines) and EF and GH (two nearly horizontal thick solid lines), respectively, wherein AB and CD intersect at a point v1 outside the image and EF and GH intersect at a point v2 outside the image, and then the equation of the straight line l passing through points v1 and v2 can be obtained v .

[0031] (2) The reference (square post) and the target (carriage) must be perpendicular to the common reference plane, and the height values are both relative to the track surface as the reference plane. The top vertex t r and the end point b r of the reference (square post) must be selected in turn, and the top vertex t x and the end point b x of the target (carriage) must be selected in turn.

[0032] (3) The height to be calculated is obtained as follows:

[0033]

[0034] wherein α is a scaling factor.

[0035] Similarly, Z y can be calculated.

[0036] The height of the car carriage Z is:

[0037]

[0038] The height of the coal is obtained according to the height of the car carriage and the corresponding data in the experience database.

[0039] Step 2: The chute swings to the initial position: according to the action signal obtained in step 1, the chute control system drives the chute to move to the initial loading position, and then waits for the next instruction.

[0040] The initial position should not interfere with the car head, so as to facilitate the next stage of action of the chute after the unloading starts. In the implementation process, the height of the car carriage obtained in step 1 is used to obtain the initial swing angle signal of the chute and the height experience value of the coal.

[0041] Step 3: Identify the distance and height of the chute opening: a lightweight single-view Figure ThreeThe relative distance between the chute mouth and the front and rear sides of the car body and the relative height from the bottom of the chute mouth to the roof plane are identified by a three-dimensional reconstruction of the car body and a linear detection algorithm.

[0042] After the chute reaches the initial position, the car body moves at a constant speed, and the image is collected in real time by vision. The distance between the front and rear edges of the chute mouth and the front and rear sides of the car body and the height of the bottom edge of the chute mouth relative to the roof plane are identified in real time by algorithm design. The specific implementation method is as follows: (1) Because the semantic information of the train car body is relatively rich in the picture, and the train always moves forward, a lightweight single-view tracking algorithm is used to track the front and rear edges of the chute mouth in real time; (2) The color of the chute mouth is single, the color change range is small, and the trajectory is relatively fixed. A trajectory tracking algorithm based on kinematic constraints is used to accurately estimate the position and attitude of the bottom of the chute mouth by fusing Kalman filtering; (3) According to the information of the car body mouth plane and range and the position and attitude information of the chute mouth, the relative position relationship between the two is calculated, so as to realize the control optimization of the subsequent chute mouth. Figure Three

[0042] After the chute reaches the initial position, the car body moves at a constant speed, and the image is collected in real time by vision. The distance between the front and rear edges of the chute mouth and the front and rear sides of the car body and the height of the bottom edge of the chute mouth relative to the roof plane are identified in real time by algorithm design. The specific implementation method is as follows: (1) Because the semantic information of the train car body is relatively rich in the picture, and the train always moves forward, a lightweight single-view tracking algorithm is used to track the front and rear edges of the chute mouth in real time; (2) The color of the chute mouth is single, the color change range is small, and the trajectory is relatively fixed. A trajectory tracking algorithm based on kinematic constraints is used to accurately estimate the position and attitude of the bottom of the chute mouth by fusing Kalman filtering; (3) According to the information of the car body mouth plane and range and the position and attitude information of the chute mouth, the relative position relationship between the two is calculated, so as to realize the control optimization of the subsequent chute mouth.

[0043] Step 4: The chute swings out to the coal pressing position: The distance between the front edge of the chute and the front side of the car body is detected in real time by step 3, and the detection data is compared with the distance S between the front edge of the chute mouth and the front side of the car body at the extension position. When the condition S>L / cosβ-V0·L / V is met, the chute control device drives the chute to move, first makes the chute swing by a set angle, and then makes the chute extension segment move to the coal pressing position at a set speed and a set length.

[0044] In order to avoid large material dust pollution in the unloading process, the chute needs to be extended into the car body for a period of time when loading the front end of each car body. The ideal extension position can be calculated according to the nearest position without interfering with the car body, provided that the car speed V0, the chute extension length L, the chute extension speed V and the chute swing angle β are known, wherein the principle of calculating the chute extension position is shown in Figure 5 The distance between the front edge of the chute and the front side of the car body is detected in real time by step 3, and the detection data is compared with a certain set instruction (the distance S between the front edge of the chute mouth and the front side of the car body at the extension position). When the condition S>L / cosβ-V0·L / V is met, the chute control system drives the chute to move, first makes the chute swing by a set angle, and then makes the chute extension segment move to the coal pressing position at a set speed and a set length. The set angle, the set speed and the set length are obtained according to the operation experience.

[0045] Step 5: The chute swings back to the coal pressing position: Through real-time detection of step 3, when the height of the bottom edge of the chute mouth is located at the coal pressing position, the image processing device sends a signal to stop the movement of the chute extension segment.

[0046] After step 4 is completed, the material continuously piles up at the front end of the car, and then the chute is first returned to the initial angle and then retracted to the coal level, and the coal level height is obtained according to the identification result of step 1. Through real-time detection of step 3, when the bottom of the chute opening is at the coal level height, the image processing system sends a signal to stop the movement of the retractable section of the chute.

[0047] Step 6, judge whether the chute opening is scraped against the car side: when the distance between the rear edge of the chute and the rear car side reaches the set command, the car loading is completed, and if the lower coal level is lower than the car roof, the chute needs to be retracted to the initial position.

[0048] With the uniform movement of the car, the vision system collects images in real time. When the distance between the rear edge of the chute and the rear car side reaches the set command through real-time detection of step 3, the car loading is completed, and if the lower coal level is lower than the car roof, the chute needs to be retracted to the initial position and then stopped.

[0049] Step 7, judge whether the whole train loading is completed: if there is still a next car detected, after step 7 is completed, the next loading cycle is started, and steps 3-7 are repeated; if there is no next car detected, the chute control device drives the chute to swing and retract to the zero position, and the train loading is completed.

[0050] (1) Detect that the loading is not completed: after step 7 is completed, the next loading cycle is started, and steps 3-7 are repeated; (2) Detect that the loading is completed: after step 7 is completed, the chute control device drives the chute to swing and retract to the zero position, and the zero position is a position with a swing angle of 0 degrees and a retractable displacement of 0 mm. At this point, the train loading is completed.

[0051] Example two:

[0052] This embodiment is an improvement of the above-mentioned embodiment, which is a refinement of the above-mentioned embodiment about step 3, identifying the distance and height of the chute opening.

[0053] This embodiment uses the Kernelized Correlation Filter (KCF) algorithm to ensure the accuracy and real-time performance of tracking; uses the Kalman filter to optimize the tracking of the object; selects four reference points to calculate the auxiliary straight line equation, and uses the straight line equation to calculate the vertical distance between the bottom of the chute opening and the top plane of the train car. Real-time detection of this vertical distance realizes the visualization of the relative height between the bottom of the chute opening and the top plane of the train car, which is a measurement method for measuring whether the chute opening and the train car interfere in the height direction, solves the problem of not having a sensor to measure the relative distance between the two objects in motion, and can be used to guide the control of the chute opening movement mechanism, which is of great significance for the realization of intelligentization of the falling material loading equipment.

[0054] The algorithm used in the method described in this embodiment includes Kernelized Correlation Filter (KCF), Fourier transform, regularization, mapping processing, Gaussian kernel calculation, linear dynamic system of Gaussian process, Kalman filter, and calculation of the distance between the target tracking center point and the upper surface of the car. The selection of the target tracking center point needs to consider the situation that the lowest end of the chute opening will be inserted into the train car to cause obstruction, so the upper h0 part of the material dropping mechanism is selected as the target tracking area.

[0055] Step 3 described in this embodiment, identifying the distance and height of the chute opening, includes the following sub-steps:

[0056] Sub-step 1, target tracking of the chute opening: KCF algorithm is used for target tracking of the moving chute opening: first, convert the target image to be tracked into the Fourier domain, and train a nonlinear classifier; the goal of KCF training is to find a function f(z) = ω T z that minimizes the squared error between the sample x i and its regression target y i ; in the formula: ω is the model weight; T is the matrix transpose; z is the cropped image block;

[0057] In order to ensure the accuracy and real-time performance of tracking, the algorithm used is Kernelized Correlation Filter (KCF). The process of this algorithm is as follows: first, convert the target image to be tracked into the Fourier domain x i (w,h)∈{0,…,W-1}×{0,…,H-1} and train a nonlinear classifier. The pixel values inside the size WxH are obtained by circularly moving the target image, and the cropped image around the target center is obtained. The expected label y i (w,h) of each sample x i is represented by a Gaussian function, which ranges from 0 to 1. y i For a centered target, it is 1, and for other areas away from the target center, it is reduced to 0. The goal of KCF training is to find a function f(z) = ω T z that minimizes the squared error between the sample x i and its regression target y i . As shown below:

[0058]

[0059] In the formula: λ is a regularization parameter used to control overfitting.

[0060] Map x i to Hilbert space through kernel function. The optimization goal becomes:

[0061] min ω ∑ i |<φ(x i ,ω)>-y i | 2 +λ‖ω‖ 2 .

[0062] where φ is a mapping function with kernel κ, After mapping, ω can be expressed as ω = ∑a si φ(x i ). The problem becomes equivalent to solving the problem, i.e., turning to solve the classifier coefficients a s . According to the circulant matrix structure and convolution theorem, the solution of the coefficient a s is:

[0063]

[0064] where F is the Fourier transform, y = {y i (w,h) | (w,h) ∈ {0,…,W-1}×{0,…,H-1}}.

[0065] k x = κ(x,x') is calculated in the Fourier transform domain using a Gaussian kernel.

[0066] For the (t+1)th frame of the image block z with size WxH, where z is cropped in the search window around the object position, where z is cropped in the search window around the object position, the confidence response is calculated as:

[0067]

[0068] where ⊙ is the element product, k z = κ(z,x i ) is the kernel distance between the regression sample z and the learning object appearance x i . The final target position is determined by the position where the maximum response R occurs, and the response R is represented as:

[0069]

[0070] Sub-step 2: Optimize target tracking: further optimize the tracking of the object using a Kalman filter, and use Kalman filtering for prediction from the (t-1)th frame to the tth frame; where t is the tth frame of the video stream;

[0071] Since the movement trajectory of the chute port is relatively fixed, and the tracking noise of the image conforms to the Gaussian distribution, it conforms to the linear dynamic system (LDS) of the Gaussian process. Therefore, the Kalman filter is used to further optimize the tracking of the object. In the t-1 frame to the t frame, the Kalman filter is used for prediction. The state variable domain state variable covariance matrix can be calculated as:

[0072] Wherein: is the state variable, specifically (x t ,y t ,dx t ,dy t ), is the covariance matrix of the state variable, F k is the state transition matrix, B k is the control matrix, is the control vector, w k is the noise matrix of the state variable, Q k is the noise matrix of the covariance matrix. The prediction value and the measurement value mixing step is:

[0073]

[0074] Wherein: H k is the mapping matrix from the state space to the measurement space; is the Kalman gain; R k is the noise matrix of the observation matrix; is the observation matrix.

[0075] Substep 3: Solve the vertical distance from the chute port bottom to the train compartment top plane: select four reference points to calculate the auxiliary straight line equation, select two points in the chute port side plane and two points on the edge of the train compartment top surface respectively; a straight line is determined by two points to obtain a straight line equation. Considering that the lowest end of the chute port will be inserted into the train compartment to cause occlusion, the selection of the target tracking center point is particularly important. The distance between the target tracking center and the train compartment top plane is calculated according to the geometric relationship. Finally, the calculated distance is corrected.

[0076] Select four reference points to calculate the auxiliary straight line equation, select two points in the chute port side plane and two points on the edge of the train compartment top surface respectively. Since in the camera plane, the chute port side plane is perpendicular to the ground and the camera plane is also perpendicular to the ground, the scaling coefficient can be simplified as:

[0077]

[0078] Wherein (x1, y1), (x0, y0) are two known points in the chute port side plane; Z rThe reference parameter is the chute opening width. The edge of the upper surface of the car is parallel to the plane, and the straight line equation can be calculated by the known two points (x2, y2) and (x3, y3):

[0079]

[0080] As shown in Figure 6 The target tracking center point is (cx, cy), and the point of the target tracking center point falling on the straight line equation in the vertical ground direction can be calculated by the straight line equation (cx, py). Therefore, the distance between the target tracking center and the upper surface of the car can be calculated Since the relative position of the rail and the upper material falling mechanism is determined, the distance between the target tracking center and the upper surface of the car can be finally calculated

[0081] Since the lowest end of the chute opening is inserted into the train car to cause shielding, the upper h0 part of the material falling mechanism is selected as the target tracking area. The finally corrected distance is:

[0082]

[0083] Finally, it should be noted that the above is only used to illustrate the technical solutions of the present application and is not limited. Although the present application is described in detail with reference to the preferred arrangement, those skilled in the art should understand that the technical solutions of the present application (such as the form of the loading station, the form of the chute, the use of various formulas, the order of steps, etc.) can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application.

Claims

1. An automated loading method for a loading station chute based on monocular vision recognition, wherein the system used in the method includes: The loading station's chute, chute control subsystem, and vision subsystem comprise two parts: a car height recognition module and a chute opening / car side relative position recognition module, as well as an image recognition subsystem. The method is characterized by the following steps: Step 1, Identify the car height: Before loading begins, identify the car height to obtain the initial swing angle signal of the chute and the coal leveling height; Step 2, chute swings to initial position: Based on the action signal identified in Step 1, the chute control system drives the chute to move to the initial loading position and waits for the next instruction. Step 3, Identify the distance and height of the chute opening: A lightweight single-view 3D reconstruction fusion straight line detection algorithm is used to identify the relative distance between the chute opening and the front and rear sides of the carriage, as well as the relative height with the roof plane. The carriage side is reconstructed in 3D to construct the plane and range of the carriage side opening. A trajectory tracking algorithm based on kinematic constraints is used, fused with Kalman filtering, to accurately estimate the position and attitude of the bottom of the chute opening and calculate the relative positional relationship between the two. Step 4, chute swings and extends to the coal pressing position: The distance between the chute's leading edge and the front side of the vehicle is detected in real time in Step 3. The detected data is compared in real time with the distance S between the chute's leading edge and the front side of the vehicle when the chute is extended. Once the condition S > L / cosβ - V0·L / V is met, the chute control device drives the chute to move. First, the chute swings at a set angle, then the chute extension section moves to the coal pressing position at a set speed and length. Where: V0 is the vehicle speed; L is the chute extension length; V is the chute extension speed; β is the chute swing angle. Step 5, the chute swings back to the level coal position: through the real-time detection in step 3, when the height of the bottom edge of the chute opening is at the level coal position height, the image processing device sends a signal to stop the movement of the chute extension section. Step 6, determine whether the chute opening scrapes against the car side: after the distance between the rear edge of the chute and the rear car side reaches the set command, the loading of one car is completed. If the lower coal level is lower than the car roof, the chute needs to be controlled to retract to the initial position. Step 7, determine if the loading of the entire train is complete: If it is detected that there is another car, then after completing Step 7, start the next loading cycle and repeat Steps 3-7; if it is not detected that there is another car, then end the loading.

2. The loading method according to claim 1, characterized in that, Step 3, identifying the chute opening distance and height, includes the following sub-steps: Sub-step 1, target tracking of the chute opening: The KCF algorithm is used to track the moving chute opening. First, the target image to be tracked is transformed into the Fourier domain, and a nonlinear classifier is trained. The goal of KCF training is to find a function f(z) = ω. T z, so that sample x i and its regression target y i The squared error between them is minimized; where: ω is the model weight; T is the matrix transpose; z is the cropped image patch; Sub-step 2: Optimize target tracking: Use a Kalman filter to further optimize object tracking, and use the Kalman filter for prediction from frame t-1 to frame t; where: t is the t-th frame of the video stream; Sub-step 3: Solve for the vertical distance from the bottom edge of the chute to the top plane of the train car: Select four reference points to calculate the auxiliary straight line equation. Select two known points in the side plane of the chute and two points on the edge of the upper surface of the train car. Determine a straight line through the two points to obtain the straight line equation. Considering that the lowest end of the chute will insert into the train car and cause obstruction, the selection of the target tracking center point is particularly important. Calculate the distance between the target tracking center and the upper plane of the train car based on geometric relationships. Finally, correct the calculated distance.

Images