Automatic wharf outer container truck full-path guiding system for container

Abstract

The invention relates to the technical field of image processing, and discloses a full-path guiding system for an outer container truck of a container automation wharf. The method comprises an obtaining module used for obtaining an outer container truck wharf driving path image and generating lane line coordinate data and interactive operation area identification position information; the extraction module is used for inputting the lane line coordinate data into a ResNet-50 convolutional neural network for feature extraction and outputting a guide marking feature vector; the fusion module is used for performing outer container truck position fusion calculation through adaptive Kalman filtering; the prediction module is used for inputting the historical passing record of the outer container truck into the Markov chain model to predict the arrival time; and the planning module is used for performing path re-planning through the improved ant colony algorithm and outputting a steering angle correction instruction. The technical problems that in an existing container automation wharf outer container truck guiding system, multi-sensor fusion positioning precision is insufficient, and dynamic environment adaptability and an intelligent path optimization mechanism are lacked are solved.

Description

Technical Field

[0001] This application relates to the field of image processing technology, and particularly to a full-path guidance system for external container trucks in a container automated terminal. Background Art

[0002] As an important part of modern port logistics, a container automated terminal uses automated equipment and intelligent control systems to achieve efficient loading, unloading, and transportation of containers. In the prior art, after an external container truck enters the automated terminal, it mainly relies on the GPS positioning system and preset static paths for guidance, and conducts basic traffic dispatching and operation arrangements through the terminal operating system (TOS). The current external container truck guidance technology generally adopts a positioning method based on magnetic nail embedding or single GPS navigation, combined with fixed traffic signals and manual command to achieve the path guidance of external container trucks.

[0003] However, the existing external container truck guidance technology has significant technical defects. The construction cost of the traditional magnetic nail positioning system is high and it is difficult to maintain. In a complex terminal environment, the positioning accuracy is easily interfered. The GPS single-point positioning has insufficient accuracy in the complex port area environment and cannot meet the requirement of centimeter-level precise path tracking. Secondly, the existing technology lacks the ability to adapt to dynamic environments and mainly relies on preset static path planning. It cannot be dynamically adjusted according to real-time traffic flow and congestion conditions, resulting in external container trucks having to wait for manual intervention when encountering temporary obstacles or traffic congestion.

[0004] The existing guidance system lacks the real-time positioning ability of multi-sensor fusion and the dynamic decision-making mechanism based on artificial intelligence. The traditional method cannot organically combine visual perception, position fusion, traffic flow prediction, and path optimization, and lacks the intelligent perception and adaptive response ability to the complex dynamic environment of the terminal. Especially in dealing with the correction of the path deviation of external container trucks, real-time obstacle avoidance, and multi-vehicle collaborative optimization, the existing technology is difficult to achieve intelligent guidance for the full path, which affects the overall operation efficiency and safety of the automated terminal. Summary of the Invention

[0005] This application provides a full-path guidance system for external container trucks in a container automated terminal, which is used to solve the technical problems of insufficient multi-sensor fusion positioning accuracy, lack of dynamic environment adaptation ability, and intelligent path optimization mechanism in the existing external container truck guidance system for container automated terminals.

[0006] This application provides a full-path guidance system for external container trucks in a container automated terminal. The full-path guidance system for external container trucks in the container automated terminal includes: an acquisition module, which is used to perform real-time image acquisition and processing on the driving path of the external container truck in the terminal through a vehicle-mounted multi-modal image acquisition unit, and generate lane line coordinate data and interactive operation area identification position information; an extraction module, which is used to input the lane line coordinate data into a ResNet-50 convolutional neural network for feature extraction processing, output a guiding marking feature vector and identify path reference point coordinates; a fusion module, which is used to perform external container truck position fusion calculation processing through adaptive Kalman filtering according to the path reference point coordinates and the guiding marking feature vector, and calculate the lateral deviation amount and the heading angle deviation degree; a prediction module, which is used to input the historical passing records of the external container truck combined with the interactive operation area identification position information into a Markov chain model for arrival time prediction processing, predict the transfer probabilities of each functional area and calculate the future queue length; a planning module, which is used to perform path replanning processing through an improved ant colony algorithm according to the transfer probabilities and the lateral deviation amount, generate a congestion avoidance and detour node sequence based on the heading angle deviation degree and output a steering angle correction instruction.

[0007] In the technical solution provided by this application, real-time image acquisition and processing of the driving path of the external container truck in the terminal is realized through a vehicle-mounted multi-modal image acquisition unit, and lane line coordinate data and interactive operation area identification position information are generated. Compared with the traditional single-sensor positioning method, the multi-modal image fusion technology significantly improves the perception accuracy and environmental adaptability in the complex terminal environment. By inputting the lane line coordinate data into a ResNet-50 convolutional neural network for feature extraction processing, outputting a guiding marking feature vector and identifying path reference point coordinates, the deep learning feature extraction algorithm can effectively handle the problems of light changes, weather interference and occlusion in the terminal environment, and has stronger robustness and accuracy compared with the traditional template matching method. According to the path reference point coordinates and the guiding marking feature vector, external container truck position fusion calculation processing is performed through adaptive Kalman filtering, and the lateral deviation amount and the heading angle deviation degree are calculated. The multi-sensor data fusion technology realizes centimeter-level positioning accuracy and solves the problem of insufficient accuracy of single GPS positioning in the port area environment. Inputting the historical passing records of the external container truck combined with the interactive operation area identification position information into a Markov chain model for arrival time prediction processing, predicting the transfer probabilities of each functional area and calculating the future queue length, the traffic flow prediction model based on probability statistics can accurately predict the terminal traffic conditions and provide a scientific basis for dynamic path planning. According to the transfer probabilities and the lateral deviation amount, path replanning processing is performed through an improved ant colony algorithm, a congestion avoidance and detour node sequence is generated based on the heading angle deviation degree and a steering angle correction instruction is output. The intelligent path optimization algorithm realizes the dynamic congestion avoidance and real-time path adjustment of the external container truck, and significantly improves the terminal traffic efficiency and safety.

[0008] In the specific application field of the full-path guidance of external container trucks in container automated terminals, the technical solution of this application fully exploits the domain specialization advantages of each core algorithm. In the lane line recognition task in the terminal environment by the ResNet-50 convolutional neural network, its residual learning mechanism can effectively extract the deep semantic features of the terminal road markings, and the recognition accuracy is significantly improved compared with traditional image processing methods under complex lighting and occlusion conditions. The adaptive Kalman filter algorithm optimizes the parameters according to the motion characteristics of large external container trucks. When fusing GPS and visual positioning data, it can dynamically adjust the fusion weights according to the terminal environment to achieve the optimal position estimation effect. The Markov chain model makes full use of the regular characteristics of the passage of external container trucks outside the terminal. By modeling the three-state transition probability, it can accurately capture the flow pattern of external container trucks between the gate area, buffer area, and interactive operation area, and the prediction accuracy is significantly improved compared with simple statistical methods. The improved ant colony algorithm combines the topological features of the terminal road network and the motion constraints of external container trucks, and realizes the rapid search of congestion avoidance paths through dynamic adjustment of pheromones and multi-objective optimization. Both the algorithm convergence speed and solution quality meet the requirements of terminal real-time scheduling. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings. Figure 1 FIG. is a schematic diagram of an embodiment of the full-path guidance system for external container trucks in a container automated terminal according to an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0010] The embodiments of the present application provide a full-path guidance system for external container trucks in a container automated terminal. The terms "first", "second", "third", "fourth", etc. (if any) in the specification, claims, and above-mentioned drawings of the present application are used to distinguish similar objects and do not have to be used to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments described here can be implemented in an order other than that illustrated or described here. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products, or devices.

[0011] For ease of understanding, the specific process of the embodiments of the present application will be described below. Please refer to Figure 1, an embodiment of the full-path guidance system for external container trucks in a container automated terminal in an embodiment of the present application includes: An acquisition module 101, configured to perform real-time image acquisition processing on the driving path of the external container truck outside the terminal through an on-vehicle multimodal image acquisition unit, and generate lane line coordinate data and interactive operation area identification position information; An extraction module 102, configured to input the lane line coordinate data into a ResNet-50 convolutional neural network for feature extraction processing, output a guiding marking feature vector, and identify the coordinates of a path reference point; A fusion module 103, configured to perform external container truck position fusion calculation processing through adaptive Kalman filtering according to the path reference point coordinates and the guiding marking feature vector, and calculate a lateral deviation amount and a heading angle offset degree; A prediction module 104, configured to input the historical passing records of the external container truck and the interactive operation area identification position information into a Markov chain model for arrival time prediction processing, predict the transfer probabilities of each functional area, and calculate the future queuing length; A planning module 105, configured to perform path replanning processing through an improved ant colony algorithm according to the transfer probability and the lateral deviation amount, generate an anti-blocking detour node sequence based on the heading angle offset degree, and output a steering angle correction instruction.

[0012] It can be understood that the execution subject of the present application can be a full-path guidance system for external container trucks in a container automated terminal, or a terminal or a server. Specifically, it is not limited here. An embodiment of the present application is described by taking the server as the execution subject as an example.

[0013] Specifically, the acquisition module 101 realizes real-time visual perception of the path of the external container truck outside the terminal through an on-vehicle multimodal image acquisition unit. This module starts a CCD camera to continuously scan the terminal road surface, and identifies the lane boundary through the pixel-level gray value difference. The Canny edge detection operator calculates the gradient amplitude and direction at each pixel point, and marks it as a boundary point when the gradient amplitude exceeds the set threshold, forming a lane line boundary pixel coordinate sequence. The Hough transform algorithm converts these discrete pixel points into line parameters, and determines the slope and intercept of the lane line by finding the peak in the parameter space, generating standardized lane line coordinate data. At the same time, an infrared thermal imager captures the temperature distribution image of the interactive operation area, and uses the thermal radiation difference of different materials and devices for area identification. The temperature threshold is set to 30-40 degrees Celsius to segment the concrete ground and metal equipment areas, and identify the boundary contour coordinates of the interactive operation area. The morphological closing operation fills the holes and breaks in the contour through the operations of dilation first and then erosion, ensuring that a complete and continuous area boundary is obtained, and generating interactive operation area identification position information.

[0014] The extraction module 102 inputs the lane line coordinate data into the ResNet-50 convolutional neural network for deep feature learning, normalizes the input coordinate data, scales the coordinate values to the range of 0-1, and rearranges them into the standard format of 224×224 pixels. The first convolutional block uses a 7×7 convolutional kernel to extract features from the input data. When the convolutional kernel slides on the image, it moves 2 pixel positions each time, calculates the weighted sum of the local area as the output feature value, and generates a primary feature map of 112×112 pixels. The batch normalization layer calculates the mean and variance of the current batch of data, normalizes each feature value by subtracting the mean and dividing by the standard deviation to prevent the problems of gradient vanishing and explosion. Four residual blocks process the feature data in sequence. Each residual block contains three convolutional layers and a skip connection. The skip connection directly adds the input feature to the convolutional output, learning the residual features while retaining the original information. The global average pooling compresses the feature maps of different scales into a fixed-length 2048-dimensional vector, reducing the data dimension by taking the average of all pixels in each feature map. The fully connected layer performs a linear transformation on this 2048-dimensional feature vector. The output of each neuron is equal to the dot product of the input vector and the weight vector plus the bias term. After passing through the softmax activation function, the output is converted into a probability distribution, and the position with the highest probability value corresponds to the coordinate of the path reference point.

[0015] The fusion module 103 realizes the optimal fusion of multi-sensor data through adaptive Kalman filtering, constructs a 6-dimensional state vector including the position coordinates, speed components, and heading angle of the external truck, and establishes the initial state in combination with the GPS positioning data. In the state prediction stage, the position and attitude at the next moment are predicted according to the vehicle motion model. The prediction equation considers the turning radius and speed constraints of the external truck, and outputs the predicted state vector and the corresponding covariance matrix. In the observation update stage, the guiding line feature vector is converted into observation data, the residual between the predicted value and the actual observation value is calculated, and the Kalman gain determines the fusion weight according to the ratio of the predicted error covariance and the observation noise covariance. The adaptive mechanism dynamically adjusts the noise covariance parameters according to the GPS signal strength and the confidence level of visual observations. When the GPS signal is weak, the weight of visual observations is increased, and vice versa, the weight of GPS data is increased. The state update fuses the predicted value and the observation value by weighted average to obtain the optimal position estimation result. The deviation calculation performs geometric operations on the fused accurate position and the center line of the guiding path. The lateral deviation is equal to the perpendicular distance from the current position of the external truck to the path center line, and the heading angle deviation is equal to the angle between the vehicle orientation and the tangent direction of the path.

[0016] The prediction module 104 uses a Markov chain model to predict and analyze the arrival time and traffic flow of external container trucks. The terminal is divided into three state spaces: the gate area, the buffer area, and the interactive operation area. The residence time and transfer times of external container trucks in each area are statistically recorded from historical passing records. The state transition probability is calculated by the maximum likelihood estimation method. The transition probability from the gate area to the buffer area is equal to the number of times this transition occurs divided by the total number of departures from the gate area. Similarly, the transition probabilities between other states are calculated to form a 3×3 transition probability matrix. The Poisson distribution modeling models the arrival time interval of external container trucks as a stochastic process, and the arrival rate parameter is estimated based on the historical arrival frequencies in different time periods. The arrival rate is higher during the morning rush hour and lower during the night time. The Markov chain state equation predicts the state distribution at each future moment through matrix multiplication operations. The current state vector is multiplied by the power of the transition probability matrix to obtain the future state probability distribution. The queue length prediction calculates the average queue length based on the vehicle arrival rate and service rate in each area. When the arrival rate approaches the upper limit of the service rate, the queue length increases sharply, forming a congestion warning.

[0017] The planning module 105 uses an improved ant colony algorithm for dynamic path optimization. The terminal road network is modeled as a weighted directed graph, where the nodes in the graph represent intersections and key locations, and the edges represent road segments. The weights of the edges comprehensively consider distance, travel time, and congestion level. The weight calculation formula performs a weighted sum of the distance cost, time cost, and congestion cost. The congestion cost is dynamically adjusted according to the predicted transition probability, and a high congestion probability corresponds to a high weight value. In the initialization stage of the ant colony algorithm, pheromones are evenly distributed on all road segments. The pheromone concentration is adjusted according to the lateral deviation amount. The pheromone concentration of the path segments with a large deviation decreases, guiding the virtual ants to avoid the paths with serious deviations. 50 virtual ants simultaneously search for paths in the road network. Each ant selects the next node according to the pheromone concentration and heuristic information, and the selection probability is proportional to the pheromone concentration and path quality. In the path evaluation stage, the total cost function of each candidate path is calculated, including path length, expected travel time, and the number of turns. An additional penalty term is added to the path segments where the heading angle deviation exceeds the threshold. The pheromone update adjusts the pheromone concentration of each path segment according to the path quality. The pheromone of the high-quality paths increases, and the pheromone of the low-quality paths evaporates. After multiple rounds of iteration, it converges to the optimal solution. The steering control calculation is based on the node sequence of the optimal path. The azimuth angle difference between adjacent nodes is calculated by geometric methods, and the steering angle and turning radius are determined according to the vehicle kinematic constraints, generating accurate steering angle correction instructions to guide the external container truck to drive along the optimized path.

[0018] In a specific embodiment, the acquisition module 101 is used for: Continuously scanning and photographing the terminal road surface through a CCD camera, performing edge detection processing based on pixel gray value analysis, and extracting the pixel coordinate sequence of the lane line boundary; Input the lane line boundary pixel coordinate sequence into the Hough transform algorithm for straight line fitting processing, calculate the lane line slope and intercept parameters, and generate lane line coordinate data; Use an infrared thermal imager to scan the temperature distribution of the interactive operation area, perform regional segmentation processing based on the temperature difference threshold, and identify the boundary contour coordinates of the interactive operation area; Perform morphological closing operation filtering on the boundary contour coordinates of the interactive operation area, eliminate noise interference, extract the complete regional contour, and generate the identification position information of the interactive operation area; Perform coordinate system normalization processing on the lane line coordinate data and the identification position information of the interactive operation area, establish a global coordinate reference system for the dock, and output path environment data in a unified coordinate format.

[0019] Specifically, the acquisition module 101 realizes continuous image acquisition of the dock road surface through a CCD camera. The camera scans the road surface in front of the external container truck at a frequency of 30 frames per second. The obtained original images contain road surface textures, lane markings, and various interference information. For edge detection processing, the Sobel operator is used to calculate the gradient of each pixel point. The operator calculates the gray difference values in the horizontal and vertical directions respectively. When the gray difference value of a certain pixel point exceeds the set threshold with adjacent pixels, it is marked as an edge point. The lane lines, as high-contrast linear features on the road surface, are preferentially recognized. During the analysis of pixel gray values, the gray values of white lane lines are usually in the range of 200 - 255, while the gray values of dark road surfaces are in the range of 50 - 100. This significant gray difference enables the edge detection algorithm to accurately locate the boundaries of lane lines. The extracted sequence of pixel coordinate points of the lane line boundaries contains the coordinate information of all edge points on the left and right lane lines. Each coordinate point records its horizontal and vertical coordinate positions in the image coordinate system. The Hough transform algorithm converts these discrete pixel coordinate points into a straight-line representation in the parameter space. The algorithm accumulates the vote counts for each possible straight line in the parameter space, and the parameter combination with the highest vote count corresponds to the real lane line. For straight-line fitting processing, the least squares method is used to calculate the slope and intercept parameters of the lane line. The slope of the left lane line is negative, indicating an inclination direction from the upper left to the lower right, and the slope of the right lane line is positive, indicating an inclination direction from the lower left to the upper right. The intercept parameter determines the intersection position of the lane line and the image boundary. The calculated slope and intercept parameters are converted into lane line coordinate data in the actual road coordinate system through coordinate transformation, forming standardized road geometric information. The infrared thermal imager measures the temperature distribution by detecting the infrared radiation emitted by the target object. The interactive operation area contains objects of different materials such as concrete ground, metal gantry cranes, and containers. There are significant differences in the thermal radiation characteristics of various materials. During the temperature distribution scanning process, the temperature of the concrete ground is usually close to the ambient temperature. Due to good thermal conductivity, the temperature of metal equipment changes relatively quickly. As a large metal structure, the temperature of the container increases significantly under sunlight. The region segmentation processing divides the thermal imaging image into different temperature regions based on the temperature difference threshold. The set temperature threshold is dynamically adjusted according to different seasons and time periods. The threshold is set to 35 degrees Celsius in summer and 25 degrees Celsius in winter. The temperature difference threshold segmentation algorithm traverses each pixel point in the thermal imaging image, marks the pixels with temperatures higher than the threshold as the equipment area, and marks the pixels with temperatures lower than the threshold as the ground area, forming a binary region segmentation image. The boundary contour coordinates of the interactive operation area are obtained through the contour extraction algorithm. The algorithm traces the continuous boundary pixel points along the region boundary and records the coordinate positions of each boundary point, forming a closed sequence of contour coordinates.

[0020] Morphological closing filtering is used to eliminate noise and correct the shape of the boundary contour of the extracted interactive operation area. Closing operation consists of two steps: dilation first and then erosion. The dilation operation uses a structuring element to expand the contour. The structuring element is a 3×3 square kernel. During dilation, the 8 neighboring pixels around each contour pixel point are marked as contour points, so that the originally broken contour line segments are connected. The erosion operation uses the same structuring element to contract the dilated contour. During erosion, only when a pixel point and its 8 neighboring pixels all belong to the contour will the pixel point maintain the contour mark, and other pixel points will have their contour marks removed. Noise interference elimination is achieved through connected component analysis. The algorithm calculates the area of each connected region, and the connected regions with an area smaller than the set threshold are considered as noise points and removed from the contour. The extraction of the complete regional contour ensures that the boundary of the interactive operation area is continuous and closed. The algorithm checks the connectivity of the contour, and for the contour with breakpoints, the linear interpolation method is used to connect them, generating the identification position information that describes the exact position and shape of the interactive operation area.

[0021] Coordinate system normalization solves the problem of coordinate inconsistency between different sensor data. The lane line coordinate data comes from the image coordinate system of the CCD camera, and the identification position information of the interactive operation area comes from the temperature image coordinate system of the infrared thermal imager. There are differences in the origin position, axis direction, and scale ratio of the two coordinate systems. The normalization process establishes the sensor calibration relationship, and the transformation matrix between the two coordinate systems is obtained by setting a calibration board in the dock environment. The transformation matrix contains rotation, translation, and scaling parameters. Coordinate transformation calculation converts the pixel coordinates in the image coordinate system into the physical coordinates in the actual road coordinate system. The parameters such as the installation height, pitch angle, and azimuth angle of the camera are considered during the transformation process. The dock global coordinate reference system is established with the dock gate as the origin. The horizontal axis points to the inside direction of the dock, and the vertical axis points to the transverse direction of the dock. All sensor data are uniformly converted to this global coordinate system. The path environment data with a unified coordinate format includes the start and end coordinates of the lane line, lane width information, vertex coordinates of the interactive operation area, and area information of the region. The data format is stored in the standard JSON format, which is convenient for data reading and processing by subsequent modules.

[0022] Illustrate the complete working process of the acquisition module as follows: After the external container truck enters the terminal, the CCD camera starts to scan the road surface ahead and captures a road surface image containing a double-lane. The left lane line consists of a series of pixel coordinate points, including point A with coordinates of 100 pixels in the horizontal direction and 200 pixels in the vertical direction, point B with coordinates of 150 pixels in the horizontal direction and 300 pixels in the vertical direction, etc., totaling 200 boundary points. The edge detection algorithm calculates the gradient value of each pixel point and finds that the gradient value in the lane line area is significantly higher than that in the road surface area, successfully extracting the boundary pixel coordinate sequences of the left and right lane lines. The Hough transform algorithm counts the straight line votes in the parameter space and finds that the straight line with a slope of 0.8 and an intercept of 50 obtains the highest number of votes, corresponding to the actual left lane line, and the slope of the right lane line is 0.7 and the intercept is 400. At the same time, the infrared thermal imager scans the interactive operation area and detects that the temperature of the gantry crane equipment is 40 degrees Celsius and the ground temperature is 30 degrees Celsius. Region segmentation is performed according to the threshold of 35 degrees Celsius, and the boundary contour of the gantry crane equipment is identified to contain 8 key vertices. Morphological closing operation processing finds that there are two breakpoints in the original contour. The dilation operation connects the breakpoints, and the erosion operation removes the redundant protrusions to obtain a complete rectangular contour. Coordinate unification processing converts the image coordinates into road coordinates. The road coordinates of the starting point of the left lane line are 2 meters eastward and 0 meters southward, the end point coordinates are 50 meters eastward and 8 meters southward, the road coordinates of the center position of the interactive operation area are 30 meters eastward and 15 meters southward, and the area of the region is 400 square meters.

[0023] In a specific embodiment, the extraction module 102 includes: A processing unit for inputting the lane line coordinate data into the input layer of the ResNet-50 network for data preprocessing, normalizing the coordinate data and converting it into a standard input format of 224×224 pixels; An operation unit for performing convolution operation with a 7×7 convolution kernel on the data in the standard input format, setting the stride to 2 and performing batch normalization processing to extract a primary feature map; A generation unit for sequentially inputting the primary feature map into 4 residual blocks for deep feature extraction processing. Each residual block contains 3 convolutional layers and a skip connection structure to generate a multi-scale deep feature map; A dimensionality reduction unit for performing global average pooling dimensionality reduction processing on the multi-scale deep feature map, compressing the feature map into a 2048-dimensional feature vector and generating a guiding marking feature vector; A transformation unit for performing linear transformation processing on the guiding marking feature vector through a fully connected layer, calculating the probability distribution of each pixel point belonging to the path reference point through the softmax activation function, identifying the highest probability position and outputting the path reference point coordinates.

[0024] Specifically, after receiving the lane line coordinate data transmitted by the acquisition module, the processing unit performs data preprocessing operations. The original format of the lane line coordinate data contains an irregular sequence of pixel coordinate points, and each coordinate point records its horizontal and vertical position information in the image. The normalization process converts the original coordinate values from the pixel scale to a standardized numerical range. The specific operation is to subtract the minimum value of the dimension from each coordinate value and then divide it by the numerical range of the dimension, ensuring that all coordinate values are distributed between 0 and 1. The data format conversion rearranges the one-dimensional coordinate sequence into a two-dimensional matrix structure. The number of rows and columns of the matrix are both set to 224, forming a standard square input format. When the number of original coordinate points is insufficient, the missing positions are filled by zero padding. When the number of coordinate points exceeds, key points are selected by equidistant sampling to generate data in the standard input format of 224×224 pixels. Each matrix element in the standard input format data represents the feature intensity at the corresponding position. The element value at the position where the lane line passes is 1, and the element value in the background area is 0, forming a binary feature representation.

[0025] The operation unit performs preliminary feature extraction on the standard input data through the first convolutional block. The 7×7 convolutional kernel contains 49 weight parameters. When the convolutional kernel slides on the input data, the product sum of all elements within the coverage area and the corresponding weights is calculated each time. The setting of the stride of 2 means that the convolutional kernel moves 2 pixel positions each time. After the convolutional operation on the 224×224 input data, the output size becomes 112×112. During the convolutional operation, the value at each output position is equal to the dot product result of all input pixels within the coverage area of the convolutional kernel and the convolutional kernel weights, and then the bias term is added to obtain the feature value. The batch normalization process calculates the mean and variance of all samples in the current batch on each feature channel, and then subtracts the mean from each feature value, divides it by the standard deviation, multiplies it by the learnable scaling parameter, and adds the learnable offset parameter. The normalization operation solves the problem of internal covariate shift during the training process of the deep network, making the network training more stable and efficient. The primary feature map contains 64 feature channels, and each channel captures different feature patterns of the input data, such as basic geometric features like horizontal edges, vertical edges, and diagonal edges.

[0026] The generation unit sequentially feeds the primary feature maps into 4 residual blocks for deep feature learning. Each residual block adopts a skip connection structure to solve the problem of gradient vanishing in deep networks. The first residual block contains 3 convolutional layers. The first layer uses a 1×1 convolutional kernel for channel dimension compression, the second layer uses a 3×3 convolutional kernel for spatial feature extraction, and the third layer uses a 1×1 convolutional kernel for channel dimension restoration. The skip connection adds the input of the residual block directly to the output of the third convolutional layer. In this way, the network learns the residual mapping instead of the direct mapping, reducing the learning difficulty. The 4 residual blocks process feature information of different scales respectively. The first residual block focuses on local detail features, and the subsequent residual blocks gradually expand the receptive field to focus on context information in a larger range. During the deep feature extraction process, the spatial resolution of the feature maps gradually decreases while the number of channels gradually increases, forming a multi-scale deep feature map that contains low-level edge texture features and high-level semantic concept features. The multi-scale feature map can capture both the local geometric shape and the global spatial layout information of the lane line, laying a feature foundation for subsequent path reference point positioning.

[0027] The dimensionality reduction unit performs global average pooling on the multi-scale deep feature map, compressing the feature information in the spatial dimension into a feature vector of a fixed length. Global average pooling calculates the average value of each feature channel over the entire spatial range. Assuming that the spatial size of a certain feature channel is height H and width W, the pooling result of this channel is equal to the arithmetic mean of all H×W elements of this channel. The dimensionality reduction process converts the originally spatially structured feature map into a one-dimensional feature vector with a fixed length of 2048 dimensions, and each dimension represents the intensity value of a high-level semantic feature. The guiding marking feature vector contains all the key feature information of the input lane line data. These features have been transformed from the original pixel-level information into a high-level semantic-level representation through successive abstractions in the deep network. The elements of different dimensions in the feature vector correspond to different visual patterns, such as straight line features, curve features, intersection features, etc. These features combined completely describe the geometric and semantic attributes of the lane line.

[0028] The transformation unit performs a linear transformation on the guiding marking feature vector through a fully-connected layer. The fully-connected layer contains 2048×224×224 weight parameters, and the value at each output position is equal to the inner product of the input feature vector and the corresponding weight vector. The linear transformation maps the high-dimensional feature vector back to the spatial coordinate domain, and the output size is 224×224, which is consistent with the spatial resolution of the original input. The softmax activation function performs probability normalization on the output of the linear transformation, calculates the probability value of each spatial position belonging to the path reference point, and the sum of the probability values of all positions is equal to 1. During the probability distribution calculation process, the softmax function calculates the exponential function of the output value at each position, and then divides the exponential value at each position by the sum of the exponential values at all positions to obtain the probability value at each position. The path reference point recognition is achieved by finding the maximum value position in the probability distribution. The position with the highest probability value corresponds to the key feature point of the lane line, usually located at the center or turning point of the lane line. The coordinate value of the highest probability position is converted from the normalized coordinate to the actual image coordinate through inverse coordinate transformation, and the output path reference point coordinates include the exact horizontal and vertical coordinate positions of the point in the image.

[0029] In a specific embodiment, the operation unit is used for: Performing a sliding window scan process on the standard input format data with a 7×7 convolution kernel, moving 2 pixel positions each time and calculating the convolution value to generate a convolution output matrix of 112×112 pixels; Inputting the convolution output matrix into a batch normalization layer for data standardization processing, calculating the batch mean and variance and performing a normalization transformation to output a standardized feature matrix; Performing a non-linear transformation on the standardized feature matrix with a ReLU activation function, setting negative values to zero and keeping positive values unchanged to generate an activated feature map; Performing downsampling on the activated feature map through a 3×3 max pooling layer, setting the stride to 2 and selecting the maximum value within the pooling window to output a pooling feature map of 56×56 pixels; Performing channel dimension reorganization on the pooling feature map, adjusting the depth dimension of the feature map to 64 channels and keeping the spatial resolution to generate a primary feature map.

[0030] Specifically, when the arithmetic unit performs a sliding window scan on the standard input format data using a 7×7 convolution kernel, the convolution kernel scans row by row starting from the upper left corner of the input data. The convolution value at each position is obtained by calculating the element-wise product of the convolution kernel weights and the corresponding input pixel values and then summing them up. During the sliding window scan, the convolution kernel moves 2 pixel positions to the right each time. When it reaches the end of the row, it moves 2 pixel positions down and returns to the beginning of the row to continue scanning. This moving method with a stride of 2 reduces the output size to 112×112 after scanning the 224×224 input data. The calculation of the convolution value involves multiplying 49 weight parameters with the corresponding 49 input pixel values. After all the product results are accumulated, the bias parameter is added to obtain the final convolution output value at this position. The value at each output position during the scanning process reflects the feature intensity of the input data in that area. The convolution value in the lane line edge area is relatively large, while the convolution value in the flat road surface area is relatively small. The resulting 112×112 pixel convolution output matrix contains the local feature information of the input data. The batch normalization layer calculates the statistical information of all samples in the current batch on each feature channel after receiving the convolution output matrix. The data normalization process requires calculating the batch mean and batch variance separately. The batch mean is obtained by summing the values at the corresponding positions of all samples in the current batch and then dividing by the number of samples. The batch variance is obtained by calculating the sum of the squares of the differences between the values at the corresponding positions of all samples and the batch mean and then dividing by the number of samples. The normalization transformation performs the operation of subtracting the batch mean and dividing by the square root of the batch variance on each feature value. This transformation adjusts the feature value distribution to a standard normal distribution with a mean close to zero and a variance close to one. The normalized values are then multiplied by the learnable scaling parameter gamma and added with the learnable offset parameter beta. These two parameters allow the network to adaptively adjust the scale and position of the feature distribution during the training process. The output normalized feature matrix maintains the same spatial size of 112×112 as the input, but the numerical distribution is more stable and is beneficial for the training and convergence of subsequent network layers.

[0031] The ReLU activation function performs a non-linear transformation on the normalized feature matrix. The calculation rule of this function is to keep the original value unchanged when the input value is greater than zero, and set it to zero when the input value is less than or equal to zero. The non-linear transformation process traverses each element in the normalized feature matrix. The operation of setting negative values to zero eliminates negative feature responses and only positive feature responses are retained for subsequent processing. The property of keeping positive values unchanged ensures that effective feature information is completely retained. The non-linear property of the ReLU function introduces non-linear modeling ability to the deep network, enabling the network to learn complex feature mapping relationships. In the activated feature map, the zero-value area corresponds to the position where the features in the input data are not obvious, and the non-zero value area corresponds to the position where the features are significant. This sparse feature representation reduces the computational complexity and highlights the key feature information.

[0032] The max pooling layer performs downsampling on the activated feature map. When the 3×3 pooling window slides on the feature map, it covers 9 adjacent pixel positions each time. The maximum value selection within the pooling window is achieved by comparing these 9 pixel values and selecting the maximum value as the pooling output at that position. The maximum value selection retains the strongest feature responses within the local area and suppresses the weaker noise information. The setting of the stride of 2 causes the pooling window to move 2 pixel positions each time. This downsampling method compresses the 112×112 input feature map to an output size of 56×56. The spatial resolution is reduced by half while retaining the most important feature information. The downsampling process reduces the amount of data for subsequent calculations and expands the receptive field of the network, enabling the subsequent network layers to capture a larger range of context information. Each pixel position in the pooled feature map represents a feature summary of a larger area in the original input.

[0033] The channel dimension reorganization process adjusts the depth structure of the pooled feature map. The number of channels in the original pooling output is related to the channel configuration of the input data. The reorganization process normalizes the number of channels to 64 feature channels. The depth dimension adjustment is achieved by increasing or decreasing the number of feature channels. When the original number of channels is less than 64, the number of channels is increased by replicating the existing channels or zero-padding. When the original number of channels is more than 64, the number of channels is reduced by channel selection or linear combination. Maintaining the spatial resolution ensures that the reorganized feature map maintains a spatial size of 56×56. Each spatial position contains a 64-dimensional feature vector describing various feature attributes at that position. The primary feature map contains the basic visual features extracted from the input lane line data after convolution, normalization, activation, and pooling processes.

[0034] In a specific embodiment, the fusion module 103 is used for: Initialize the state vector by constructing it with the path reference point coordinates and the current GPS position data of the external truck. Set the position, speed, and heading angle of the external truck in the dock coordinate system as state variables to generate the motion state vector of the external truck; Perform state prediction processing by establishing a vehicle motion prediction model in the dock environment based on the motion state vector of the external truck. Consider the motion characteristics of the external truck turning and going straight in the dock, and output the predicted position coordinates and the predicted error covariance; Convert the guiding marking feature vector into the observation data of the external truck relative to the center line of the lane for observation update processing. Calculate the observation residual of the external truck deviating from the guiding path to generate the Kalman gain coefficient; Perform adaptive weight fusion processing on the predicted position of the external truck based on the Kalman gain coefficient. Dynamically adjust the fusion weights in combination with the GPS positioning accuracy and the visual observation confidence, and output the positioning result of the external truck in the dock; Based on the positioning result of the external container truck, deviation analysis and processing are carried out with the center line of the wharf guiding path, the vertical distance of the external container truck deviating from the lane center line and the included angle between the vehicle head direction and the path direction are calculated, and the lateral deviation amount and the heading angle offset degree are generated.

[0035] Specifically, the fusion module 103 receives the path reference point coordinates output by the extraction module and the current position data of the external container truck obtained by the vehicle-mounted GPS receiver, and the state vector construction process converts these two types of spatial position information into a unified data structure. The initialization process converts the GPS longitude and latitude coordinates into plane coordinates in the local coordinate system of the wharf through coordinate system transformation. The wharf coordinate system takes the gate position as the origin, the eastward direction as the positive direction of the X axis, the northward direction as the positive direction of the Y axis, and the elevation as the positive direction of the Z axis. The state variable setting includes six dimensions such as the position coordinates, speed components, and heading angle of the external container truck. The position coordinates record the X and Y coordinate values of the external container truck in the wharf coordinate system, the speed components record the instantaneous speeds of the external container truck in the X and Y directions, and the heading angle records the angle of the vehicle head direction of the external container truck relative to the X axis of the coordinate system. The motion state vector of the external container truck organizes these six state variables into a column vector. The first and second elements of the vector are the position coordinates, the third and fourth elements are the speed components, the fifth element is the heading angle, and the sixth element is the angular velocity, forming a mathematical representation describing the complete motion state of the external container truck. The vehicle motion prediction model establishes a mathematical motion equation based on the motion state vector of the external container truck, and the state prediction process considers the motion constraint conditions of the external container truck as a large vehicle in the wharf environment. The motion characteristics in the wharf environment include physical parameters such as turning radius limit, maximum acceleration constraint, and angular velocity limit. When turning, the external container truck needs a large turning radius to avoid collision with roadside facilities, and when moving straight, the external container truck maintains a constant speed or uniform acceleration motion. The motion prediction model adopts a discretized state transition equation. The position coordinates at the next moment are equal to the current position plus the product of the speed and the time interval, the speed at the next moment is equal to the current speed plus the product of the acceleration and the time interval, and the heading angle at the next moment is equal to the current heading angle plus the product of the angular velocity and the time interval. The predicted position coordinates are calculated by the motion equation to obtain the expected position of the external container truck at the next time step. The prediction error covariance describes the uncertainty degree of the prediction result. The diagonal elements of the covariance matrix represent the prediction variances of each state variable, and the non-diagonal elements represent the correlations between different state variables.

[0036] The observation update process converts the guiding marking feature vector output by the extraction module into the observation information of the relative position of the external truck. The guiding marking feature vector contains the position and direction information of the lane center line, and the relative relationship between the current position of the external truck and the lane center line is calculated through geometric transformation. The observation data conversion process extracts the key geometric parameters in the guiding marking feature vector, including the equation parameters of the lane center line and the position information of the external truck in the image, and then converts the image coordinates into the physical distance and angle in the road coordinate system through coordinate transformation. The observation residual calculation compares the difference between the predicted position of the external truck and the actual position obtained by observation. Each element of the residual vector represents the prediction deviation in different dimensions. The Kalman gain coefficient is determined according to the relative magnitudes of the predicted error covariance and the observation noise covariance. When the prediction error is large, the Kalman gain increases and more relies on the observation data for state update. When the observation noise is large, the Kalman gain decreases and more relies on the prediction result.

[0037] The adaptive weight fusion process dynamically adjusts the fusion weight according to the GPS positioning accuracy and the visual observation confidence. The GPS positioning accuracy is evaluated by the satellite signal strength and the geometric distribution factor. When the signal strength is high and the satellite distribution is uniform, the GPS accuracy is high; otherwise, the accuracy is low. The visual observation confidence is evaluated by the image quality and the feature matching degree. When the image is clear and the lane line features are obvious, the confidence is high; when the image is blurred or seriously occluded, the confidence is low. The weight adjustment mechanism dynamically allocates the fusion weight according to the real-time performance of the two sensors. When the GPS accuracy is higher than the visual observation, the weight of the GPS data is increased. When the visual observation quality is better than the GPS, the weight of the visual data is increased. The fusion calculation combines the predicted state and the observation state through weighted average. The final state estimate is equal to the sum of the predicted state and the product of the Kalman gain and the observation residual. The accurate positioning result of the external truck in the wharf output contains the optimal estimated values of the position coordinates, speed, and heading angle.

[0038] Deviation analysis and processing geometrically compares the fused external container truck’s precise positioning results with the predefined terminal guidance path. The centerline of the guidance path is defined by a series of control points and connecting curves, describing the ideal driving trajectory that the external container truck should follow. The vertical distance calculation is implemented by the point-to-line distance formula or the point-to-curve distance algorithm. The algorithm finds the point on the guidance path that is closest to the current position of the external container truck, and then calculates the vertical distance from the external container truck’s position to that point as the lateral deviation. The angle between the vehicle head and the path direction is calculated by the vector angle formula. The heading angle of the external container truck represents the vehicle head pointing direction, and the tangent direction of the guidance path at the nearest point represents the desired driving direction. The angle between the two direction vectors is the heading angle deviation degree. When the lateral deviation is positive, it means that the external container truck deviates to the right side of the path, and when it is negative, it means that it deviates to the left side of the path. When the heading angle deviation degree is positive, it means that the vehicle head is deviating to the right, and when it is negative, it means that the vehicle head is deviating to the left. These two deviation parameters provide key feedback information for the subsequent path replanning algorithm.

[0039] In a specific embodiment, the prediction module 104 is used to: The historical passage records of foreign container trucks are divided into functional areas according to the gate area, buffer area and interactive operation area, and the residence time and transfer times of foreign container trucks in each area are counted to generate the historical data of regional transfer of foreign container trucks; According to the historical data of the regional transfer of the external container truck, a three-state Markov chain transfer matrix is constructed to perform probability calculation processing, calculate the state transfer probability of the external container truck from the gate area to the buffer area and from the buffer area to the interactive operation area, and output the inter-region transfer probability matrix; The interactive operation area identification location information is used as a constraint condition to perform Poisson distribution modeling on the arrival time interval of external container trucks, and the arrival rate parameters are calculated according to the historical arrival frequency of different time periods to generate the arrival time distribution model of external container trucks; Based on the inter-regional transfer probability matrix and the external container truck arrival time distribution model, the future traffic state prediction calculation process is performed, and the change in the number of external container trucks in each region within 15 minutes is predicted through the Markov chain state equation, and the transfer probability of each functional area is output; The queue length is predicted based on the transfer probability of each functional area and the number of container trucks outside each area, and the cumulative number and average waiting time of vehicles in the buffer area and the interactive operation area are calculated to generate the future queue length.

[0040] Specifically, after receiving the historical passing data of external container trucks recorded by the terminal monitoring system, the prediction module 104 performs functional area division processing according to the spatial layout of the terminal. The gate area is the first area where external container trucks enter the terminal, including the areas for identity verification and vehicle information registration. The buffer area is located between the gate area and the interactive operation area, and is a temporary parking area for external container trucks to wait for operation arrangements and route allocations. The interactive operation area is the core area where external container trucks perform container loading and unloading operations with automated equipment. The functional area division is achieved by analyzing the GPS trajectory data of external container trucks, and each trajectory point is assigned to the corresponding functional area according to the spatial coordinate range. The coordinate range of the gate area is a rectangular area near the terminal entrance, the coordinate range of the buffer area is a strip-shaped area in the middle of the terminal, and the coordinate range of the interactive operation area is a rectangular area within the operation range of the gantry crane. The residence time statistics are obtained by calculating the time span of consecutive trajectory points of external container trucks in each area. When an external container truck moves from one area to another, a regional transfer event is recorded once. The transfer times statistics are obtained by accumulating the total number of regional transfer events of all external container trucks within a specified time period. The historical data of external container truck regional transfers includes the entry time, departure time, residence duration of each external container truck in each area, as well as the transfer path and transfer moment between areas. These data are cleaned and formatted to form a structured historical data set.

[0041] The construction of the three-state Markov chain transition matrix is based on probability calculation processing of the historical data of external container truck regional transfers. The Markov chain abstracts the movement of external container trucks in the terminal as a random transfer process between three discrete states. The state transition probability calculation is obtained by dividing the frequency of an external container truck transferring from one state to another by the total number of transfers starting from that state. The transfer probability from the gate area to the buffer area is equal to the number of times all external container trucks enter the buffer area from the gate area divided by the total number of times starting from the gate area. The transfer probability from the buffer area to the interactive operation area is equal to the number of times all external container trucks enter the interactive operation area from the buffer area divided by the total number of times starting from the buffer area. The transition matrix is a three-row and three-column square matrix, and the matrix elements represent the probability of transferring from the state corresponding to the row to the state corresponding to the column. The sum of the elements in each row of the matrix is equal to one, indicating the completeness of the probability distribution. The inter-regional transfer probability matrix not only includes the transfer probabilities between directly adjacent regions, but also includes the probabilities of cross-regional transfers. For example, the probability of an external container truck directly entering the interactive operation area from the gate area or returning from the interactive operation area to the gate area. The zero elements in the matrix indicate the absence of corresponding transfer paths.

[0042] The Poisson distribution modeling process takes the identification location information of the interactive operation area as a constraint condition to statistically model the arrival time interval of outbound container trucks. The Poisson distribution is a probability distribution that describes the number of random events occurring per unit time. The identification location information of the interactive operation area defines the spatial scope of the arrival events of outbound container trucks. Only the outbound container trucks entering the specified interactive operation area are counted as arrival events. The arrival events in different interactive operation areas are modeled separately to reflect the usage characteristics of each area. The historical arrival frequency is calculated by statistically averaging the number of arrivals of outbound container trucks at the interactive operation area during different time periods. The time period division includes morning rush hour, daytime regular period, evening rush hour, and night period, etc. The arrival rate parameter for each time period is calculated by dividing the total number of arrivals during that time period by the duration of the time period. The arrival rate parameter reflects the density of the arrival of outbound container trucks. A higher arrival rate during the morning rush hour indicates that the arrival of outbound container trucks is more frequent, and a lower arrival rate during the night period indicates that the arrival of outbound container trucks is relatively sparse. The arrival time distribution model of outbound container trucks includes the Poisson distribution parameters for different time periods and the corresponding time window definitions. The model can predict the probability distribution of the arrival of outbound container trucks at any time point.

[0043] The future traffic state prediction calculation process is based on a comprehensive analysis of the inter-regional transfer probability matrix and the arrival time distribution model of outbound container trucks. The Markov chain state equation describes the evolution law of the system state over time. The basic form of the state equation is that the state probability distribution at the next moment is equal to the state probability distribution at the current moment multiplied by the transfer probability matrix. The state distributions at future moments are obtained through iterative calculations. The prediction calculation divides the 15-minute prediction time window into multiple time steps, each time step being 1 minute. The changing trend of the number of outbound container trucks in each area within 15 minutes is obtained through 15 iterative calculations. The change in the number of outbound container trucks in each area is the combined effect of two factors. One is the transfer flow of existing outbound container trucks between areas, and the other is the incremental contribution of newly arrived outbound container trucks to the number in each area. The output of the transfer probabilities for each functional area includes the real-time transfer probabilities between each area for each time step. These probability values are dynamically adjusted according to the current system state and historical transfer patterns.

[0044] The queuing length prediction process performs queuing modeling calculations based on the transfer probabilities of each functional area and the current number of external container trucks in each area. Queuing theory models the waiting process of external container trucks in each area as a queuing system. The vehicle cumulative quantity is calculated by accumulating the number of external container trucks entering the area minus the number of external container trucks leaving the area in each time step. The cumulative quantity in the buffer reflects the queuing situation of external container trucks waiting for job arrangements, and the cumulative quantity in the interactive operation area reflects the number of external container trucks undergoing loading and unloading operations. The average waiting time is calculated based on Little's Law of queuing theory. The average waiting time is equal to the average queuing length in the system divided by the arrival rate of external container trucks. When the queuing length in the buffer increases, the average waiting time is correspondingly extended. When the service efficiency in the interactive operation area decreases, the waiting time further increases. The future queuing length is obtained by predicting the cumulative quantity of external container trucks in each area within the future time window. The prediction results include the queuing length values per minute for the buffer and the interactive operation area within the next 15 minutes.

[0045] In a specific embodiment, the planning module 105 is configured to: Perform graph theory modeling processing on the terminal road network using the transfer probability as the road weight parameter, assign dynamic weight values to road segments according to the congestion probability of each functional area, and generate a weighted terminal road network topology graph; Perform pheromone distribution setting processing by initializing the starting parameters of the ant colony algorithm based on the lateral deviation amount, reduce the pheromone concentration of path segments with large lateral deviations by 20%, and output the initial pheromone concentration matrix; Execute the ant colony search algorithm for path optimization processing according to the initial pheromone concentration matrix, set 50 virtual ants to explore paths on the terminal road network, calculate the total cost function of each path, and generate multiple candidate path schemes; Perform path evaluation processing on the multiple candidate path schemes in combination with the course angle deviation degree, add a penalty factor to path segments with a course angle deviation exceeding 5 degrees, screen out the optimal congestion avoidance path, and output the congestion avoidance detour node sequence; Perform steering control calculation processing based on the congestion avoidance detour node sequence and the current position of the external container truck, calculate the steering angle and turning radius between adjacent nodes through geometric methods, and generate a steering angle correction instruction.

[0046] Specifically, after receiving the transition probability data output by the prediction module, the planning module 105 performs graph theory modeling to abstract the terminal road network into a mathematical graph structure. The nodes in the graph represent key positions within the terminal, including intersections, turning points, and parking spaces, and the edges represent road segments connecting the nodes. In the process of allocating road weight parameters, the transition probabilities of each functional area calculated by the prediction module are converted into the passing costs of the corresponding road segments. Areas with high transition probabilities indicate a large external truck flow and are prone to congestion, and the corresponding road segments are assigned higher weight values to reflect the difficulty of passing. The congestion probability calculation is based on historical traffic data and current prediction results. When the number of external trucks in a certain functional area exceeds the designed capacity, the congestion probability of this area increases, and the weight of the road segment connecting this area is correspondingly increased. The dynamic weight value allocation is adjusted according to the real-time traffic conditions. During the morning rush hour, the weight of the road segment leading to the interactive operation area increases, and during the night time, the weights of all road segments decrease. The numerical range of the weight value is from 1 to 10, and the larger the value, the higher the passing cost. The weighted terminal road network topology graph contains the spatial coordinates of all nodes, the connection relationships between nodes, and the dynamic weight values of each edge. The graph structure provides a calculation basis for the path search of the subsequent ant colony algorithm.

[0047] The pheromone distribution setting process initializes the parameters of the ant colony algorithm based on the lateral deviation amount output by the fusion module. The pheromone concentration reflects the historical usage frequency and preference degree of the path segment. The initial parameter initialization sets the pheromone concentration of all road segments to the same reference value, and then adjusts it according to the current lateral deviation situation of the external truck. Road segments with a large lateral deviation indicate that the external truck is prone to deviate from the correct trajectory on this path, and the navigation quality of such road segments is poor, so their priority in path selection should be reduced. The pheromone concentration reduction process is achieved by multiplying the pheromone concentration of the road segment with a lateral deviation exceeding the threshold by a reduction coefficient of 0.8. The reduction amplitude of 20% ensures that the attractiveness of the deviation path segment decreases significantly. The construction of the pheromone concentration matrix is organized according to the road network topology structure. The rows and columns of the matrix correspond to the nodes in the road network respectively, and the matrix element value represents the pheromone concentration of the path segment between the corresponding nodes. After the initial pheromone concentration matrix is output, the ant colony algorithm will perform path search based on this matrix. Road segments with a high pheromone concentration are more likely to be selected by virtual ants, and the probability of selecting road segments with a low pheromone concentration is correspondingly reduced.

[0048] The execution process of the ant colony search algorithm performs path optimization based on the initial pheromone concentration matrix. Virtual ants are computational units that simulate the foraging behavior of biological ants. Each ant independently conducts path search and records the search results. At the beginning of path exploration, 50 virtual ants are randomly distributed at the current position of the out-gate truck. Each ant selects the next moving node according to the pheromone concentration and heuristic information. The node selection probability calculation comprehensively considers the pheromone concentration and distance factors. Nodes with high pheromone concentration and short distance have a greater probability of being selected, while nodes with low pheromone concentration or long distance have a smaller probability of being selected. During the path optimization process, each ant maintains a taboo list to record the visited nodes to avoid forming meaningless circular paths. When an ant reaches the target position, a complete path search is completed. The total cost function calculates the comprehensive cost of each path, which includes three components: the total path length, the expected travel time, and the congestion penalty. The total path length is obtained by accumulating the physical distances of each path segment. The expected travel time is calculated based on the path segment weights and vehicle speeds. The congestion penalty is determined based on the number of highly congested areas passed by the path. The generation process of multiple candidate path schemes collects all the valid paths found by the ants. After removing duplicate paths, they are sorted according to the total cost function values. The path with the minimum cost is ranked first as the preferred scheme.

[0049] The path evaluation process comprehensively analyzes the candidate path scheme and the course angle deviation degrees output by the fusion module. The course angle deviation reflects the difference between the current driving direction of the out-gate truck and the desired direction. The course angle of a path segment is calculated by analyzing the path geometry. The course angle of a straight path segment remains unchanged, the course angle of a curved path segment changes continuously, and the course angle of a sharp-turn path segment changes significantly. Path segments with a course angle deviation exceeding 5 degrees are considered to require large steering adjustments, and such path segments add additional penalty factors to reflect the complexity and risk of steering operations. The penalty factor calculation is determined according to the specific value of the course angle deviation. The larger the deviation degree, the higher the penalty factor. The penalty factor is doubled for path segments with a deviation exceeding 10 degrees, and path segments with a deviation exceeding 15 degrees are directly excluded. The comprehensive path score is obtained by adding the total path cost and the course angle penalty factor. The path with the lowest score is selected as the optimal congestion avoidance path, which has the minimum travel cost and steering complexity under the premise of reaching the target. The output of the congestion avoidance detour node sequence includes the coordinate information of all key nodes on the optimal path and the connection relationship between the nodes. The node sequence is arranged in the driving order, and the distance and direction information between adjacent nodes are clearly recorded.

[0050] The steering control calculation process performs geometric analysis based on the traffic avoidance and detour node sequence and the current position of the external container truck. The geometric method includes vector calculations and trigonometric operations to determine the steering parameters. The steering angle calculation between adjacent nodes is achieved by analyzing the angular relationship formed by three consecutive points. The current position of the external container truck, the next node, and the node after the next node form a triangle, and the steering angle is equal to the included angle between the current driving direction of the external container truck and the target direction. The turning radius calculation is based on the kinematic constraints of the external container truck and the road geometry conditions. The minimum turning radius of a large external container truck is limited by the wheelbase length and the steering angle. When turning on a narrow road, a larger turning radius is required to avoid collisions with roadside facilities. In the geometric method, the vector cross product calculation is used to determine the steering direction. A positive cross product indicates a left turn, a negative cross product indicates a right turn, and the absolute value of the cross product reflects the magnitude of the steering angle. The steering angle correction instruction generation process converts the calculated steering angle and turning radius into control instructions that can be executed by the driver of the external container truck or the autonomous driving system. The instructions include the specific value of the steering angle, a clear indication of the steering direction, and a recommended driving speed. The steering angle is accurate to one decimal place in degrees, the steering direction is clearly indicated as a left turn or a right turn, and the recommended speed is determined based on the turning radius and the road conditions.

[0051] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. An external truck full-path guidance system for a container automated terminal, characterized in that the system Including: An acquisition module, configured to perform real-time image acquisition and processing on the driving path of the external container truck terminal through the vehicle-mounted multi-modal image acquisition unit, and generate lane line coordinate data and interactive operation area identification position information; An extraction module, configured to input the lane line coordinate data into a ResNet-50 convolutional neural network for feature extraction processing, output a guiding marking feature vector, and identify the path reference point coordinates; A fusion module, configured to perform external container truck position fusion calculation processing through adaptive Kalman filtering according to the path reference point coordinates and the guiding marking feature vector, and calculate the lateral deviation amount and the heading angle offset degree; A prediction module, configured to input the historical passing records of the external container truck combined with the interactive operation area identification position information into a Markov chain model for arrival time prediction processing, predict the transfer probabilities of each functional area, and calculate the future queue length; A planning module, configured to perform path replanning processing through an improved ant colony algorithm according to the transfer probability and the lateral deviation amount, generate a congestion avoidance and detour node sequence based on the heading angle offset degree, and output a steering angle correction instruction.

2. The full-path guidance system for external container trucks at an automated container terminal according to claim 1, wherein The acquisition module is configured to: Continuously scan and photograph the terminal road surface through a CCD camera, perform edge detection processing based on pixel gray value analysis, and extract the lane line boundary pixel coordinate sequence; Input the lane line boundary pixel coordinate sequence into the Hough transform algorithm for straight line fitting processing, calculate the lane line slope and intercept parameters, and generate lane line coordinate data; Scan the temperature distribution of the interactive operation area through an infrared thermal imager, perform region segmentation processing based on the temperature difference threshold, and identify the interactive operation area boundary contour coordinates; Perform morphological closing operation filtering processing on the interactive operation area boundary contour coordinates, eliminate noise interference, and extract the complete region contour to generate interactive operation area identification position information; Perform coordinate system normalization processing on the lane line coordinate data and the interactive operation area identification position information, establish a global coordinate reference system for the terminal, and output path environment data in a unified coordinate format.

3. The full-path guidance system for external container trucks at an automated container terminal according to claim 1, wherein, The extraction module includes: A processing unit, configured to input the lane line coordinate data into the input layer of the ResNet-50 network for data preprocessing, perform normalization processing on the coordinate data, and convert it into a standard input format of 224×224 pixels; An operation unit, configured to perform 7×7 convolutional kernel convolution operation on the data in the standard input format through the first convolutional block, set the stride to 2, and perform batch normalization processing to extract the primary feature map; A generation unit, configured to sequentially input the primary feature map into 4 residual blocks for deep feature extraction processing, each residual block includes 3 convolutional layers and a skip connection structure, and generate a multi-scale deep feature map; A dimensionality reduction unit, configured to perform global average pooling dimensionality reduction processing on the multi-scale deep feature map, compress the feature map into a 2048-dimensional feature vector, and generate a guiding marking feature vector; A transformation unit is used to perform a linear transformation on the guiding marking feature vector through a fully connected layer, calculate the probability distribution of each pixel belonging to the path reference point through a softmax activation function, identify the position with the highest probability, and output the path reference point coordinates.

4. The full-path guidance system for external container trucks at an automated container terminal according to claim 3, wherein, The operation unit is used for: Performing a sliding window scan process on the standard input format data with a 7×7 convolution kernel, moving 2 pixel positions each time and calculating the convolution value to generate a convolution output matrix of 112×112 pixels; Inputting the convolution output matrix into a batch normalization layer for data standardization processing, calculating the batch mean and variance, and performing a normalization transformation to output a standardized feature matrix; Performing a non-linear transformation on the standardized feature matrix through a ReLU activation function, setting negative values to zero and keeping positive values unchanged to generate an activated feature map; Performing downsampling on the activated feature map through a 3×3 max pooling layer, setting the stride to 2 and selecting the maximum value within the pooling window to output a pooling feature map of 56×56 pixels; Performing a channel dimension reorganization process on the pooling feature map, adjusting the depth dimension of the feature map to 64 channels while maintaining the spatial resolution to generate a primary feature map.

5. The full-path guidance system for external container trucks at an automated container terminal according to claim 1, characterized in that, The fusion module is used for: Initializing the state vector constructed from the path reference point coordinates and the current GPS position data of the external truck, setting the position, speed, and heading angle of the external truck in the dock coordinate system as state variables to generate the motion state vector of the external truck; Establishing a vehicle motion prediction model in the dock environment based on the motion state vector of the external truck for state prediction processing, considering the motion characteristics of the external truck turning and going straight in the dock, and outputting the predicted position coordinates and the predicted error covariance; Converting the guiding marking feature vector into observation data of the external truck relative to the lane centerline for observation update processing, calculating the observation residual of the external truck deviating from the guiding path, and generating a Kalman gain coefficient; Performing an adaptive weight fusion process on the predicted position of the external truck based on the Kalman gain coefficient, dynamically adjusting the fusion weight by combining the GPS positioning accuracy and the visual observation confidence, and outputting the positioning result of the external truck in the dock; Performing a deviation analysis process on the positioning result of the external truck and the centerline of the dock guiding path, calculating the vertical distance of the external truck deviating from the lane centerline and the angle between the vehicle head direction and the path direction, and generating a lateral deviation amount and a heading angle offset degree.

6. The full-path guidance system for external container trucks at an automated container terminal according to claim 1, wherein The prediction module is used for: Dividing the historical passing records of the external truck into functional areas according to the gate area, buffer area, and interactive operation area, counting the residence time and transfer times of the external truck in each area, and generating the historical data of the external truck area transfer; Constructing a three-state Markov chain transition matrix based on the historical data of the external truck area transfer for probability calculation processing, calculating the state transition probabilities of the external truck from the gate area to the buffer area and from the buffer area to the interactive operation area, and outputting the inter-area transfer probability matrix; Modeling the arrival time interval of the external truck according to the Poisson distribution with the position information of the interactive operation area identifier as a constraint condition, calculating the arrival rate parameter according to the historical arrival frequency in different time periods, and generating the arrival time distribution model of the external truck. Based on the inter-region transfer probability matrix and the external container truck arrival time distribution model, future traffic state prediction calculation and processing are carried out. The change in the number of external container trucks in each region within 15 minutes is predicted through the Markov chain state equation, and the transfer probability of each functional region is output; Based on the transfer probability of each functional region and the current number of external container trucks in each region, queue length prediction processing is carried out. The cumulative number of vehicles and the average waiting time in the buffer area and the interactive operation area are calculated to generate the future queue length.

7. The full-path guidance system for external container trucks at an automated container terminal according to claim 1, wherein The planning module is used for: Taking the transfer probability as the road weight parameter, graph theory modeling processing is carried out on the terminal road network. According to the congestion probability of each functional region, dynamic weight values are assigned to road segments, and a weighted terminal road network topology map is generated; Based on the lateral deviation amount, the starting parameters of the ant colony algorithm are initialized for pheromone distribution setting processing. The pheromone concentration of the path segment with a large lateral deviation is reduced by 20%, and the initial pheromone concentration matrix is output; According to the initial pheromone concentration matrix, the ant colony search algorithm is executed for path optimization processing. 50 virtual ants are set to explore paths on the terminal road network, and the total cost function of each path is calculated to generate multiple candidate path schemes; The multiple candidate path schemes are combined with the course angle offset degree for path evaluation processing. A penalty factor is added to the path segment with a course angle offset exceeding 5 degrees, and the optimal congestion avoidance path is selected, and the congestion avoidance detour node sequence is output; According to the congestion avoidance detour node sequence and the current position of the external container truck, steering control calculation and processing are carried out. The steering angle and turning radius between adjacent nodes are calculated by geometric methods, and a steering angle correction instruction is generated.

Images