AI Vision-Based Dynamic Path Optimization Control System for Tracked Vehicles

Abstract

This invention relates to the field of intelligent driving in rail transit, and discloses a dynamic optimization control system for the travel path of a railcar based on AI vision. The system includes: a data acquisition module for real-time acquisition of track environment images, railcar motion state data, and auxiliary information; a data preprocessing module for standardizing the acquired data; an environment modeling module for constructing a dynamic track environment model to predict the future position of the railcar; a model optimization module for extracting direction perception features and generating a fused feature map through a three-motion attention mechanism in the diagonal, forward, and backward directions, and extracting path key points based on an attention-guided fast path search method, then using a graph search algorithm to search for the optimal path among the key points; and a control execution module for generating control signals based on the optimal path and driving the railcar. This invention achieves stable operation and intelligent decision-making for the railcar in complex dynamic environments.

Description

Technical Field

[0001] This invention relates to the field of intelligent driving technology for rail transit, and in particular to a dynamic optimization control system for the driving path of a railcar based on AI vision. Background Technology

[0002] With the development of artificial intelligence and computer vision technologies, rail transit systems are rapidly evolving towards greater intelligence and autonomy. As crucial automated execution units in scenarios such as industrial logistics, tunnel inspection, and intelligent transportation, the planning and control of railcars directly impacts the overall efficiency and safety of the system.

[0003] In traditional track vehicle path control methods, path planning often relies on static maps and rule-based control logic, making it difficult to cope with dynamic environmental changes, such as interference from temporary obstacles, abrupt changes in path structure, and multiple vehicle convergence. At the same time, existing methods often neglect the track vehicle's ability to perceive the direction of the environmental structure during operation, resulting in blind spots in path selection, slow response, and even risks such as collisions and path interruptions.

[0004] Furthermore, in path search algorithms, how to enhance local diversity while ensuring global optimality, and avoid getting trapped in early convergence and local optima, is also a key challenge in the field of intelligent path optimization.

[0005] Therefore, there is an urgent need for an efficient path optimization and control mechanism that has direction perception capabilities and can integrate visual features to achieve stable operation and intelligent decision-making of the track vehicle in complex dynamic environments. Summary of the Invention

[0006] This invention aims to provide a dynamic optimization control system for the travel path of a track vehicle based on AI vision. The system integrates an attention mechanism with a fast path search method guided by attention, and achieves collaborative optimization in the path perception, planning and control stages. It solves the problems that traditional methods cannot cope with dynamic environmental changes, such as interference from temporary obstacles, sudden changes in path structure, and convergence of multiple vehicles. It also addresses the risks of blind spots in path selection, slow response, and even collisions and path interruptions in existing methods, thus enabling stable operation and intelligent decision-making of the track vehicle in complex dynamic environments.

[0007] To achieve the above objectives, the following technical solution is adopted:

[0008] A dynamic optimization control system for the travel path of a track vehicle based on AI vision, comprising:

[0009] The data acquisition module is used to acquire real-time images of the track environment, motion status data of the track trolley, and auxiliary environmental information including lighting conditions and track markings;

[0010] The data preprocessing module standardizes the collected data and outputs standardized sensor data.

[0011] The environment modeling module is used to build a dynamic track environment model with environmental perception and behavior prediction interfaces based on the standardized number of sensors, so as to predict the future position of the track vehicle.

[0012] The model optimization module extracts direction-aware features and generates a fused feature map through a three-motion attention mechanism in the diagonal, forward, and backward directions. It also extracts path key points based on an attention-guided fast path search method and then uses a graph search algorithm to search for the optimal path among the key points.

[0013] The control execution module generates control signals based on the optimal path and drives the track trolley to move.

[0014] Furthermore, the environment modeling module includes:

[0015] An environmental sensing unit is used to divide the track area into a grid matrix, with each grid containing an environmental attribute identifier, including obstacles, track markings, and hazardous conditions.

[0016] The trajectory prediction unit is used to predict the position of the next K time steps based on the current position and velocity vector using a kinematic model.

[0017] The behavior prediction interface unit is used to integrate environmental and state data and output structured information, including: static obstacle set, dangerous area set, moving obstacle parameters, track vehicle state, and load factor.

[0018] Furthermore, the model optimization module includes:

[0019] The three motion direction modeling units are used to model the diagonal direction to detect the oblique trend of the track and obstacles; model the forward direction to enhance the accessibility analysis of the path ahead; and model the backward direction to identify obstacles and path loop features behind.

[0020] The orientation masking suppression unit sets mask matrices for three motion directions, including diagonal orientation mask matrix, forward mask matrix and backward mask matrix, to suppress attention scores in irrelevant directions.

[0021] Furthermore, the orientation mask suppression unit specifically includes:

[0022] The diagonal mask matrix is ​​set to 0 when the feature point positions i and j are different, and negative infinity when they are the same; the reverse mask matrix is ​​set to 0 when the feature point position i is greater than j, and negative infinity otherwise; the forward mask matrix is ​​set to 0 when the feature point position i is less than j, and negative infinity otherwise.

[0023] Furthermore, the model optimization module also includes:

[0024] The three-directional feature pair scoring unit is used to apply the corresponding mask matrix to any two feature points in the orbit image according to their positional relationship. It performs direction-aware transformation through a learnable mapping matrix, adds the transformed feature vectors and adds a learnable bias vector, enhances the non-linear interaction between features through the hyperbolic tangent activation function, and superimposes the mask values ​​of the corresponding directions to generate directional attention scores.

[0025] The probability distribution calculation unit is used to calculate the attention score for each feature channel in three motion directions, and obtain the weight probability distribution of the three directions by normalization through the Softmax function.

[0026] Furthermore, the model optimization module also includes:

[0027] The feature fusion unit is used to aggregate the adjacent features of the corresponding directions according to the weight probability distribution of the three directions to generate a direction-aware feature representation; a gating mechanism is used to weight and combine the original visual features and the direction-aware features to generate a fused feature map representation; wherein, the gating mechanism learns the feature dependency degree through the Sigmoid function and adaptively adjusts the fusion ratio of the original information and the direction information.

[0028] Furthermore, the attention-guided fast path search method includes:

[0029] Key point extraction steps: Extract high attention regions as key points from the fusion feature map output by the three motion attention mechanisms, and add the current position and target position of the track car to form a set of key points;

[0030] Path graph construction steps: Key points are treated as nodes, neighboring nodes are connected to form edges, and each edge is assigned a weight, which is determined by the distance between nodes, the attention value at the node, and the risk value of the area traversed by the edge.

[0031] Graph search steps: On the constructed path graph, use the A* search algorithm to find the optimal path from the starting point to the ending point.

[0032] Furthermore, the key point extraction step specifically includes:

[0033] The fused feature map is compressed into a single-channel two-dimensional attention map by taking the maximum or average value.

[0034] Based on threshold segmentation, positions in the attention map that are greater than or equal to τ are marked as candidate keypoints;

[0035] Non-maximum suppression is applied to the candidate keypoints to obtain a sparse set of keypoints;

[0036] The current position and target position of the trolley are also added to the key point set, and finally the key points of the track trolley are extracted.

[0037] Furthermore, the control execution module includes:

[0038] The path smoothing unit is used to smooth the optimal path obtained by search using cubic spline interpolation or B-spline curve to obtain a smooth path.

[0039] The speed planning unit is used to plan the driving speed based on the curvature of each point on the smooth path and the corresponding attention value.

[0040] The control signal generation unit is used to generate steering and acceleration control commands based on the smoothed path and the planned driving speed.

[0041] The real-time feedback unit is used to collect the deviation between the actual position of the track vehicle and the nearest point on the planned path at the current moment. When the triggering condition is met, a replanning instruction is triggered.

[0042] Furthermore, the data preprocessing module performs the following operations: image denoising, brightness equalization, and edge enhancement.

[0043] Compared with the prior art, the present invention achieves the following beneficial effects:

[0044] 1. This invention is based on the Three-Motion Attention Mechanism (TMA): it simulates the human visual system's ability to selectively focus on dynamic features in different directions, and constructs attention perception models for three motion directions: diagonal, forward, and backward. This guides the track vehicle to focus on the "forward path", "potential backtracking trajectory", and "track oblique structure" at the visual level, effectively improving the ability to understand the directional environment and extract features, and solving the problem of blind spots in the directional perception of traditional methods.

[0045] 2. This invention proposes a directional masking suppression mechanism: A masking matrix is ​​introduced for multi-directional attention, by setting the attention score of irrelevant directions to... In the Softmax operation, invalid directions are suppressed, allowing attention to be focused more on the track area that is consistent with or related to the direction of the vehicle's movement, thereby improving the effectiveness and selectivity of path perception.

[0046] 3. This invention proposes a three-directional feature pair scoring method: constructing a directional feature scoring function to score the orbital feature points in any pair of images in a directional manner, capturing their spatial correlation in the current direction, and providing basic directional score support for subsequent probability calculation and feature fusion.

[0047] 4. This invention proposes to use a gating mechanism to achieve feature fusion: the original visual features and the orientation enhancement features are weighted and combined, an adaptive gating structure is introduced, and the degree of dependence of the region on the original information and orientation information is learned through the Sigmoid function, so as to realize a more refined feature fusion strategy and enhance the environmental adaptability of the vehicle path judgment.

[0048] 5. This invention employs Attention-Guided Fast Path Search (AGFPS), which extracts key points from high-attention regions to construct a sparse path graph. By combining the A* algorithm with a directional heuristic function, it shortens the time required for long path planning, significantly improves the dynamic obstacle response speed, and avoids local optima, thus solving the problem of dynamic obstacle response delay in traditional methods.

[0049] 6. This invention combines dynamic replanning and multi-factor adaptive strategies to ensure driving smoothness and safety. It integrates path curvature, attention risk value and load factor to dynamically adjust speed, and uses pure tracking steering PID speed tracking to reduce trajectory tracking error and achieve high-precision stable control under complex working conditions.

[0050] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0051] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the invention. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0052] Figure 1 This is a schematic diagram of a module of a dynamic optimization control system for the travel path of a track vehicle based on AI vision, according to an embodiment of the present invention.

[0053] Figure 2 This is a flowchart illustrating the model optimization module of a dynamic optimization control system for a track vehicle's travel path based on AI vision, according to an embodiment of the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0056] Figure 1 This is a schematic diagram of a module of a dynamic optimization control system for a track vehicle's travel path based on AI vision, according to an embodiment of the present invention. Figure 1 As shown, a dynamic optimization control system 100 for the travel path of a track vehicle based on AI vision includes:

[0057] The data acquisition module 110 is used to acquire real-time images of the track environment, trolley motion status data, and auxiliary information.

[0058] In a track-guided vehicle operating system, data acquisition is a fundamental step in achieving dynamic path optimization. This system deploys high-definition wide-angle cameras at key nodes along the track (such as curves, intersections, and areas with frequent obstacles), while integrating forward-looking and top-looking dual-camera modules on the vehicle itself, forming a multi-angle, all-around visual perception network. Preferably, the high-definition wide-angle cameras are deployed at intervals ≤5 meters at track curves / intersections / slope change points, with the top-looking module vertically positioned 0.8-1.2 meters from the track, and the forward-looking module tilted at an angle of 15°±3°. Visual positioning uses ORB-SLAM3 fused with IMU data, with a positioning error ≤2cm. The following information is collected in real time through the high-definition wide-angle cameras and dual-camera modules:

[0059] (1) Track environment images: including track structure, obstacles, and traffic conditions (such as water accumulation, damage, etc.);

[0060] (2) Car motion status data: including the car's position on the track (using visual positioning fusion IMU), real-time speed, acceleration, etc.;

[0061] (3) Auxiliary information: such as changes in lighting, interference from people or other vehicles.

[0062] These raw data are preprocessed by local edge processing modules (such as Jetson or RK3588 platforms) and then uploaded to the central control unit, providing a high-quality, low-latency data source for subsequent processing.

[0063] The data preprocessing module 120 standardizes the collected data and constructs a dynamic track environment model with environmental perception and behavior prediction interfaces based on the standardized data to predict the future position of the track vehicle.

[0064] The data preprocessing module first standardizes the raw images and sensor data collected by the 120 lectures through the image preprocessing module, including image denoising (such as wavelet transform or Gaussian filtering), brightness equalization, edge enhancement and other operations, in order to improve the accuracy of environmental feature recognition.

[0065] Through the above data cleaning, standardization, and preliminary feature extraction (such as image denoising, brightness equalization, and edge enhancement), this module outputs standardized sensor data.

[0066] Environment Modeling Module 130: Constructs a dynamic orbital environment model with interfaces for environmental perception and behavior prediction based on preprocessed data;

[0067] The environment modeling module 130 integrates the real-time position information and velocity vector data of the track vehicle to construct a dynamic track environment model. This model has the following functions: environmental perception: identifying obstacles, road markings, and temporary changes in working conditions in the track; synchronous modeling of track vehicle status: synchronously superimposing the movement trajectory of the track vehicle in the environment to predict its future position; behavior prediction interface: providing input to the model optimization module 140, including dynamic information such as obstacle positions, movement trends, and changes in track vehicle load.

[0068] The dynamic orbital environment model specifically includes:

[0069] The environmental sensing unit 131 is used to divide the track area into a grid matrix, each grid containing an environmental attribute identifier, the environmental attributes including obstacles, track markings and hazardous conditions;

[0070] Specifically, the orbital region is divided into an M×N grid matrix, and an environmental attribute matrix is ​​generated by identifying the environmental attributes within each grid cell using a target detection algorithm. ,in:

[0071] When there are no obstacles in the grid (i', j'), i.e., a safe area (unmarked), =0;

[0072] When there is a static obstacle in the grid (i', j'), =1;

[0073] When the grid (i', j') has track markers (such as stop lines), =2;

[0074] When the grid (i', j') has a temporary hazardous condition (water accumulation or oil stains), =3.

[0075] Raster coordinate transformation satisfies: ,

[0076] Wherein, the raster coordinates (i', j') represent the raster index after the orbital region is divided, i' is the row index, and j' is the column index; (x, y) are the world coordinate system coordinates; , ) represents the origin of the orbital region; r represents the grid resolution, and the grid resolution r satisfies 0.05m≤r≤0.2m. Target detection is implemented using a lightweight convolutional neural network.

[0077] The trajectory prediction unit 132 is used to predict the future trajectory of the track vehicle, including: calculating the position of the track vehicle in the next K time steps based on the current position and velocity vector of the track vehicle through a kinematic model;

[0078] Specifically, based on the current position of the track vehicle ( , ) and velocity vector ( , ) Calculate the position at the next K time steps using a kinematic model, including:

[0079] k=1,2,…,K, the prediction step K satisfies 3≤K≤5, and the prediction period Δt is consistent with the system control period.

[0080] The behavior prediction interface unit 133 is used to integrate environmental and state data and output structured information to subsequent modules (model optimization module 140 and control execution module 150), including: static obstacle set, dangerous area set, moving obstacle parameters, track vehicle state and load factor.

[0081] static obstacle coordinate set ;

[0082] dangerous area coordinate set ;

[0083] Set of parameters for moving obstacles The trend vector ;

[0084] Specifically, the trend vector of moving obstacles is calculated using the inter-frame differencing method: , t is the current control cycle, t-1 is the previous control cycle, ( ) represents the world coordinate system position of the moving obstacle in the current control period t (obtained through raster coordinate mapping transformation), The value represents the world coordinate system position of the moving obstacle in the previous control period t⁻¹ (obtained through raster coordinate mapping). This differential calculation reflects the displacement of the moving obstacle within one control period Δt (optionally 0.1s).

[0085] The car's current state D = {current position (i', j'), predicted path {( , }};

[0086] Loading factor , The load for the current control cycle. The load is the load of the previous control cycle t-1.

[0087] The output of this module (environmental state matrix) The data from the behavior prediction interface is input into the model optimization module 140, which is used to determine the risk value of the edge, the current state of the car (starting point) and the target position (end point), and to adjust the direction mask matrix. The area of ​​interest. The load factor λ of the output of the prediction interface unit is input to the control execution module 150 to participate in speed planning.

[0088] The model optimization module 140 extracts direction-aware features and generates a fusion feature map through a three-motion attention mechanism in the diagonal, forward, and backward directions. It also extracts path key points based on an attention-guided fast path search method and then uses a graph search algorithm to search for the optimal path among the key points.

[0089] The model optimization module 140 is mainly used to implement model training and optimization.

[0090] To achieve dynamic optimization of the track vehicle's travel path, the model optimization module 140 integrates a three-motion attention mechanism with an attention-guided fast path search method. Through the synergistic effect of feature enhancement and path search, it constructs a model system with environmental perception and intelligent decision-making capabilities. Specifically, it consists of the following two core components:

[0091] (a) Introducing the three-motor attention mechanism

[0092] To enhance the track vehicle's ability to perceive key path information in complex environments, this system introduces a three-motion attention mechanism (TMA). This mechanism simulates how the human visual system perceives dynamic features in different directions, guiding the model to focus on information such as the "forward path," "potential backtracking trajectory," and "track oblique structure," thereby enhancing the directionality of track environment perception and improving feature discrimination capabilities.

[0093] Figure 2 This is a flowchart illustrating the model optimization module of a dynamic optimization control system for a track vehicle's travel path based on AI vision, according to an embodiment of the present invention. Figure 2 As shown, the model optimization module 140 includes:

[0094] The three motion direction modeling unit 141 is used to model the diagonal direction to detect the oblique trend of the track and obstacles; model the forward direction to enhance the accessibility analysis of the path ahead; and model the backward direction to identify obstacles and path loop features behind.

[0095] Specifically, the modeling process for the three motion directions of the three motion direction modeling unit 141 is as follows:

[0096] Diagonal direction (d=1): Focuses on the diagonal movement of tracks or obstacles, suitable for detecting obstacles such as ramps, curves, or diagonally intruding obstacles;

[0097] Forward direction (d=2): Simulates the normal driving direction of the car (from the upper left to the lower right), used to enhance the analysis of the accessibility of the path ahead;

[0098] Backward direction (d=3): Focus on back paths or abnormal trajectories to improve the ability to identify obstacles or closed-loop features of the path behind.

[0099] The directional masking suppression unit 142 sets mask matrices for three motion directions, including a diagonal direction mask matrix, a forward mask matrix, and a backward mask matrix, to suppress attention scores in irrelevant directions.

[0100] Introducing a directional masking suppression mechanism: by introducing a mask matrix Attention suppression is applied to irrelevant directions, and the suppression term is set to... This causes the score in that direction to tend to 0 during the Softmax operation.

[0101]

[0102] Where i, j: feature point indices at different locations in the track image, used to represent position pairs within the grid map under the visual perception of the track vehicle). : Diagonal mask matrix, used to shield feature points from their own ineffective attention; : Inverse mask matrix, used to preserve feature information (historical trajectory) from the current position before; : Forward mask matrix, used to preserve feature information (predicted path) from the current position onwards; This indicates that the direction is valid and attention should be retained. : This indicates that the direction is invalid, and attention is suppressed to 0; : Represents the mask matrix corresponding to the direction of motion (can be ) , or ), used in Softmax to force the attention weights of irrelevant directions to be zero.

[0103] In this orientation mask suppression unit 142, the output of the behavior prediction interface unit is used to adjust the orientation mask matrix. The area of ​​interest. Here, obstacle information (static and moving) is primarily used to adjust the mask, suppressing the directional mask in the obstacle area (set to...). ) or Enhance (set to 0). Specifically, adjust the mask using obstacle information (S, H, Y). The specific logic is as follows:

[0104] For any pair of positions (i, j), if the grid (i', j') corresponding to position i or position j belongs to the set of static obstacles S or the set of dangerous areas H, then... Set to -∞ for all directions d (or a specific direction);

[0105] For the current position (i', j') of the moving obstacle Y, the position of position i or j corresponding to the position of this grid is also masked. Set to -∞. Additionally, based on the trend vector of the moving obstacle. Predict the grid coordinates it may move to in the near future (e.g., within the next control cycle or path search estimation time). , ),in , , This indicates the prediction duration. These predicted raster coordinates are also considered temporary danger zones, and the position pair mask corresponding to these predicted rasters at position i or j is set to -∞.

[0106] When adjusting the orientation mask using obstacle information, it is necessary to map the feature map position indices i and j to their corresponding physical space regions. Using the known mapping relationship between the feature map spatial coordinate system and the physical world coordinate system, the physical coordinate range or representative point coordinates corresponding to positions i and j are calculated. , Then, according to the raster coordinate transformation rules (see environment perception unit 131), these physical coordinates are converted into corresponding raster coordinates. , Finally, check whether these grid coordinates belong to the static obstacle set S, the danger zone set H, or the moving obstacle set Y.

[0107] By setting mask matrices for the three motion directions—diagonal mask matrix, forward mask matrix, and backward mask matrix—information about irrelevant directions in the track image is masked, allowing the attention mechanism to focus more on the current motion direction of the track vehicle.

[0108] The three-directional feature pair scoring unit 143 is used to apply the corresponding mask matrix to any two feature points in the orbit image according to their positional relationship, perform direction-aware transformation through a learnable mapping matrix, add the transformed feature vectors and add a learnable bias vector, enhance the nonlinear interaction between features through the hyperbolic tangent activation function, and superimpose the mask value of the corresponding direction to generate a directional attention score.

[0109] For any pair of feature points in the orbit images and The relevance score of its attention direction is:

[0110] in, : Represents the directional attention score at the i-th and j-th positions in the track image, used to measure whether the track vehicle should focus on the passage information in that direction on the track. : The visual feature vectors at the i-th and j-th positions of the track image represent the spatial point states perceived by the track vehicle; : A learnable direction-aware mapping matrix used to perform direction encoding transformations on the start and end features respectively; : Learnable bias vectors used to adjust the nonlinear expressive power of the fused features; : Scaling factor, which is an empirical parameter with a value range of 0.1 ≤ q ≤ 1.0, and a default value of q=0.5, to prevent... Nonlinear functions tend to saturate when the input value is too large, and are used to improve the stability of gradient propagation; Hyperbolic tangent activation function, used to enhance nonlinear interactions between feature pairs; The value at the corresponding position in the mask matrix controls whether to retain the attention score for that pair of features; : A vector (or scalar) of all ones used to mask The weighting is applied to the score in an explicit form, thereby achieving attention shielding from irrelevant directions.

[0111] This formula is used to score the directional degree of visual association between any two points in a track image, and is key to the generation of attention weights.

[0112] The probability distribution calculation unit 144 is used to calculate the attention score for each feature channel in three motion directions. The attention probability distribution in the three directions is obtained by normalization through the Softmax function, so that the attention weight of the invalid direction is reduced to zero.

[0113] Specifically, the calculation process of probability distribution calculation unit 144 includes the following:

[0114] Fractions calculated separately for each of the three directions of motion It needs to be standardized to a probabilistic form:

[0115]

[0116] On the m-th feature channel, the i-th and j-th feature points are in the same direction. Attention scores are used to measure whether the track vehicle should prioritize traffic features in this direction; : Movement direction type index, d=1 indicates diagonal direction, d=2 indicates forward direction, d=3 indicates backward direction; : The learnable mapping matrix of the m-th feature dimension under direction d, used to transform the source point With the target point Channel characteristics, simulating directional correlation; : The bias vector (length C) of the m-th channel under direction d, used to enhance the directional sensitivity of feature fusion; : Scaling factor, used to avoid The activation function can prevent gradient vanishing when the input amplitude is too large, thus improving learning stability; : The mask matrix for the d-th direction, used to suppress illegal or invalid attention values ​​(e.g., no passageway) in that direction before the Softmax operation. It is a scalar value, calculated based on the position pair (i, j) and direction d (mainly adjusted based on the environment model output); A vector of dimension m consisting entirely of 1s, used in conjunction with a mask. Element-wise multiplication forces the attention value for invalid directions to zero.

[0117] The attention scores for the three motion directions are input into Softmax to calculate the dependency probabilities:

[0118]

[0119] This represents the weighted probability of aggregating information from direction d for feature channel m when fusing features at position j. This value indicates the strength of the likelihood of the track vehicle traveling in that direction, and its range is [value missing]. ; The exponential function is used to convert the original scores into positive values ​​and amplify the score differences, thereby enhancing the contrast of directional choices. : Used to normalize the scores of all directions, so that the result is expressed in probabilistic form; d represents the index of the direction currently being calculated. This represents the index of the direction variable used to iterate through the denominator of the Softmax function, which is used to calculate the total score for all directions.

[0120] Among them, the diagonal direction probability represents the degree of attention to the diagonal structure, the forward direction probability represents the degree of attention to the forward path, and the backward direction probability represents the degree of attention to the backward trajectory.

[0121] The feature fusion unit 145 is used to weight and aggregate the adjacent features of the corresponding directions according to the attention probability distribution of the three directions to generate a direction-aware feature representation; a gating mechanism is used to weight and combine the original visual features and the direction-aware features to generate a fused feature map representation; wherein, the gating mechanism learns the feature dependency degree through the Sigmoid function and adaptively adjusts the fusion ratio of the original information and the direction information.

[0122] Specifically, for each feature point, its neighboring features from each direction are aggregated based on the probability weighting of directional attention to generate a direction-aware representation. :

[0123]

[0124] : Indicates the direction from the d-th direction to the feature point It extracts direction-aware features to capture structural and semantic information in specific directions; : Represents the directional perception representation after the fusion of the j-th track feature point, which integrates its information weighted in three directions, helping the track vehicle to make a comprehensive judgment on the "upstream and downstream path structure" and "diagonal channel potential" at the current node.

[0125] The final By integrating structural information from three directions, the track trolley can dynamically assess the upstream and downstream structures and the oblique passageway structure.

[0126] Furthermore, a gating mechanism is used to achieve a weighted combination of the original features and the directional fusion features:

[0127]

[0128] The fusion feature map representation integrates the original visual features and directional structural features as input for the path decision of the track vehicle, thereby improving its dynamic response capability to complex track environments. The original local feature representation in the track image is generated by the previous feature extraction network and mainly reflects the basic visual features of the track point where the track car is located. The three motion direction feature representation, which integrates forward, backward and diagonal structural information, is used to enhance the perception ability of the track vehicle of the upstream and downstream track environment and channel structure. The directional feature mapping matrix in the gating mechanism is used to... The feature channels are weighted and transformed to adapt to the fusion scenario. The original feature mapping matrix in the gating mechanism is used to... Perform a mapping transformation to align with and fuse the orientation-enhanced features. : Gating bias term, controls the degree of offset of the gating output, and adjusts the feature activation threshold during the fusion process. The Sigmoid activation function compresses the weighted feature result to... The interval, as a gating vector for fusion weights, is used to adaptively control the fusion ratio of the original features and the directional enhancement features. By fusing gating vectors, we can automatically learn which areas in a track scene rely more on raw visual perception and which areas rely more on orientation-enhanced perception. Hadamard product (element-wise multiplication) is used for gated weighting of feature channels. The complementary weighting factor for the directional features, relative to the gating vector, is used to adjust the directional sensing features. The degree of participation; F is the gating vector, i.e. .

[0129] By introducing a three-motor attention mechanism, the system can explicitly model the structural orientation of the orbital environment at the visual level, providing a more discriminative and directional feature base for the subsequent path planning stage.

[0130] (ii) Attention-guided Fast Path Search (AGFPS) method for searching the optimal path

[0131] Fusion feature map of output from three motor attention mechanisms ,

[0132] Input: fused feature map (The size is H×W×C, but usually the average or maximum value is taken in the channel dimension to obtain a two-dimensional attention map with a size of H×W), and the behavior prediction interface data output by the dynamic track environment model (including obstacle position, current state of the track vehicle, target position, etc.).

[0133] The process of searching for the optimal path includes the following steps:

[0134] S1: Key point extraction steps: Extract high attention regions as key points from the fusion feature map output by the three motion attention mechanisms, and add the current position and target position of the track car to form a set of key points;

[0135] In step S1, the system performs high-confidence path node extraction, specifically including:

[0136] S11: Perform channel compression on the fused feature map, compressing it into a single-channel two-dimensional attention map by taking the maximum value or average value;

[0137] Specifically, the 3D feature map (H×W×C) is compressed into a 2D attention heatmap (H×W). Preferably, channel-dimensional max pooling is used. .

[0138] S12: Based on threshold segmentation, positions in the attention map that are greater than or equal to τ are marked as candidate keypoints;

[0139] Set a dynamic threshold τ (usually 0.5-0.7) and mark positions in the attention map that are greater than or equal to τ as candidate keypoints to filter out high-attention regions:

[0140] .in, It is a compressed 2D attention map The pixel coordinates on the screen.

[0141] S13: Perform non-maximum suppression (NMS) based on spatial distance on the candidate keypoints to obtain a sparse set of keypoints;

[0142] Non-maximum suppression (NMS): To avoid overly dense adjacent points, non-maximum suppression is applied to candidate keypoints. Specifically, a suppression radius is set. (Unit: pixels), and the suppression radius of NMS is determined using the raster resolution r and the desired physical spacing ∆min. For each candidate point, in its Within the neighborhood, only retain The point with the largest value is used to suppress other points. This results in a sparse keypoint set. , u=1,2,…,N; u is the index of the key point, and the spacing in physical space is at least ∆min.

[0143] S14: Add the current position (starting point) and target position (end point) of the track car to the key point set, and finally extract the key points of the track car.

[0144] In step S14, the current position of the track trolley is set. and target location Forced inclusion into the keypoint set to enhance keypoints.

[0145] Add start and end points: Add the current position (start) and target position (end) of the track vehicle to the key point set to obtain the key point set. (Including the starting point and the ending point).

[0146] S15: Keypoint Feature Index Mapping: For a set of keypoints Each key point in Using known physical coordinates ( The mapping relationship between the feature map and the feature map spatial coordinate system (this relationship is determined by the structure or calibration parameters of the feature extraction network) is calculated in the original feature map (i.e., the generated feature map). Feature map, size is Corresponding position index (or coordinates) , Store the index. The feature vector used for subsequent access to the key point .

[0147] S2: Path graph construction steps: Treat key points as nodes, connect neighboring nodes to form edges, and assign weights to each edge. The weights are determined by the distance between nodes, the attention value at the node, and the risk value of the area traversed by the edge.

[0148] Construct a weighted directed graph G=(V,E) as the basis for path search:

[0149] Node set: ;

[0150] Edge generation rules: For each node Connect all nodes within its spatial neighborhood. The connection radius R is dynamically adjusted (3-8m) according to the complexity of the environment.

[0151] The above process indicates that for each pair of nodes ( , If the Euclidean distance between them If the radius is smaller than the preset radius R (e.g., R = 5 meters), then an edge is created between the two nodes. That is, for each node... Connect nodes whose Euclidean distance is less than the radius R. (v≠u), forming an edge .

[0152] For each key point ,Pick The value at that position is as ,Right now (x', y') are key points. Position coordinates on a 2D attention map. Keypoints. Due to its The pixel coordinates (x', y') on the keypoint are defined. The physical location (world coordinates) corresponding to this keypoint needs to be obtained through coordinate transformation. Assume... The pixel coordinate system has a mapping relationship with the original perceived image / raster map (e.g., determined by camera intrinsic and extrinsic parameters or raster map generation parameters), and this mapping relationship is part of the system calibration. For simplicity, the scheme can assume that this mapping relationship is known, and key points... physical coordinates ( It can be obtained by looking up a table or by calculation.

[0153] Feature map coordinates Corresponding attention value: .

[0154] Edge weight calculation:

[0155] in, Each pair of nodes and The Euclidean distance between them; :node Normalized attention values ​​(obtained from the 2D attention map); :node Normalized attention values ​​(obtained from the 2D attention map). : indicates an edge The maximum risk level corresponding to all the grids traversed (obtained through discretization), specifically: the risk level is based on the environmental state matrix. The mapping yields: if =1 (static obstacle), then the risk level = (e.g., 1.0); if If the risk level is >=3 (temporary hazardous conditions), then the risk level is = (e.g., 0.6); others () If the risk level is 0 or 2, then the risk level is = (e.g., 0.0), then take the maximum value of all grid risks after discretizing the entire edge as... α, β, γ: Weighting coefficients, satisfying α + β + γ = 1. Specific values ​​can be adjusted according to the actual scenario, for example, α = 0.5, β = 0.3, γ = 0.2. Weights The smaller the value, the better the edge (because it has a shorter distance, higher attention value, and lower risk).

[0156] Furthermore, if the edge direction (from) arrive If the angle between the vector and the direction of the car's current velocity is less than 45 degrees, then the weights are discounted. .

[0157] S3: Graph search steps: On the constructed path graph, use the A* search algorithm to find the optimal path from the starting point to the ending point.

[0158] Specifically: The A* algorithm is used to search for the optimal path (i.e. the path with the lowest cost) from the starting point to the ending point in the constructed path graph.

[0159] (1) Cost function: , where g(n) is the actual cost from the starting point start to node n (i.e., the sum of the weights of all edges on the path). w(e) is the weight of edge e, and Path(start→n) is the path from the starting point start to node n. h(n) is a heuristic function. dist(n,end) is the Euclidean distance from node n to the endpoint end. The average attention value of the grid cells within a 3×3 neighborhood of node n is derived from the fused attention map. Extracted from . φ is the attention weight coefficient, with a value range of [0.1, 0.3], preferably 0.2.

[0160] (2) Direction-guided node expansion: (2.1) Expansion priority: Nodes in the OpenSet are sorted in ascending order of f(n), and the node with the smallest f(n) is expanded first. (2.2) Attention-driven directional bias: For the neighbors v of the current node n, the directional angle from the current node n to each neighbor v is calculated. (The current velocity direction relative to the current node n) ): , in Let n be a vector pointing to v. Weight adjustment condition: If... Align with the high attention direction, i.e., the neighbor v satisfies ≥ (Attention threshold) If the edge weight is 0.6, then the edge weight will be temporarily adjusted. The weight adjustment factor ξ = 0.5 is used after adjustment. Participate in the calculation of g(n). The weight of high attention regions is reduced, increasing their selection priority. (2.3) Invalid direction suppression: If v is located in the mask matrix Labeled inhibition region ( If the value is -∞, then skip that neighbor and do not add it to the open set.

[0161] (3) Search termination conditions: Success: The target node is reached (the current node is the end point); Failure: The open set is empty or the maximum number of iterations is exceeded (e.g., 1000 times).

[0162] (4) Path backtracking: trace back from the end point end to the parent node and back to the starting point s.

[0163] (5) Generate the optimal path: trace the parent node pointers backward from the endpoint back to the starting point, generating a series of ordered key point sequences: .

[0164] The control execution module 150 generates control signals based on the optimal path and drives the track trolley to move.

[0165] After completing path planning, the system needs to generate control signals based on the optimal path, and then drive the track trolley to move precisely according to the control signals. Specifically, the control execution module 150 includes:

[0166] The path smoothing unit 151 is used to smooth the optimal path obtained through the search using cubic spline interpolation or B-spline curves to obtain a smooth path. Since the searched path is a polyline composed of a series of key points, smoothing is necessary for the smooth movement of the track vehicle. Here, cubic spline interpolation (or B-spline curve) is used to fit the polyline into a smooth curve, obtaining the smooth path C(s), where s is the arc length parameter. The control point spacing adaptively adjusts the path curvature to satisfy the maximum curvature constraint. .

[0167] Speed ​​planning unit 152: On a smooth path C(s), the driving speed is planned based on the curvature of each point on the smooth path and the corresponding attention value (safety level), specifically including the following steps:

[0168] Calculate the curvature κ(s) at each point on the smooth path C(s) (which can be obtained by differentiation), and add curvature constraints: , where δ is the load sensitivity coefficient; This represents the upper limit of the dynamic curvature corresponding to the arc length parameter s, which decreases as the load λ increases.

[0169] Calculate the speed at that point based on the curvature. :

[0170] in: The maximum speed set by the system; The maximum allowable curvature of the track trolley (exceeding this curvature may cause it to overturn). It is a preset small value (such as 0.001) to avoid division by zero; Attention value at that point (from nearest neighbor or bilinear interpolation) (obtained from the figure); λ is the load factor; is the sensitivity coefficient for the effect of load change on speed, with a value range of [0, 1]. =0 indicates that the effect of load changes is ignored; =1 indicates that the effect of load change on speed reaches its maximum (i.e., speed is scaled by a factor of 1 / λ); 0 < < 1 indicates that the impact of load changes on speed falls between these two values. This value can be set according to system characteristics, for example... = 0.5.

[0171] In addition, the risk of obstacles in a dynamic environment must be considered: if there are moving obstacles near the point, the speed should be further reduced.

[0172] The control signal generation unit 153 is used to generate steering and acceleration control commands based on the smoothed path and the planned driving speed, and specifically performs the following operations:

[0173] (1) Steering control: Pure Pursuit algorithm is adopted. During the movement of the track vehicle, a certain forward look-ahead distance from the current position of the track vehicle is selected on the path. Using the target point as the reference point, calculate the steering angle δ:

[0174] Where: L: wheelbase of the track trolley (distance between the front and rear wheels); The angle between the current heading of the track trolley and the vector from the track trolley to the target point is called the heading deviation angle. Forward sight distance, which is positively correlated with speed (e.g., , and (where v is a constant and v is the real-time speed of the track vehicle).

[0175] Speed ​​control: A PID controller is used to track the planned speed curve v(s). Where 'a' represents the acceleration command.

[0176] The real-time feedback unit 154 is used to collect the deviation between the actual position of the track trolley and the nearest point on the planned path at the current moment. When the trigger condition is met, a replanning instruction is triggered. Specifically:

[0177] Collect the actual position of the track trolley Distance on the planned path at the current moment nearest point Deviation:

[0178] When any of the following conditions are met, i.e. when When a preset threshold (e.g., 0.2m) is reached or a new obstacle is added, the safe distance is maintained. = 0.5v + 0.3 (meters) or the predicted path of the moving obstacle intersects with the current path, where v is the real-time speed of the track vehicle, triggering the following replanning instruction:

[0179] The vehicle's state and obstacle information in the dynamic track environment model (environment modeling module 130) are updated using the latest sensor data. Based on the updated environmental information, the orientation mask in the three-motion attention mechanism (model optimization module 140) is recalculated. Based on the updated environment model and attention mask, the path search process (AGFPS method) of the model optimization module 140 is re-executed to generate a new optimal path. Control execution module 150 based on the new path Re-perform path smoothing, speed planning, and generate control signals (steering angle δ and acceleration a).

[0180] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0181] It should also be noted that, in the embodiments of this application, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0182] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in the embodiments of this application may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown in this application, but is to be accorded the widest scope consistent with the principles and novel features disclosed in the embodiments of this application.

Claims

1. A dynamic optimization control system for the travel path of a track vehicle based on AI vision, characterized in that, include: The data acquisition module is used to acquire real-time images of the track environment, motion status data of the track trolley, and auxiliary environmental information including lighting conditions and track markings; The data preprocessing module standardizes the collected data and outputs standardized sensor data. The environment modeling module is used to build a dynamic track environment model with environmental perception and behavior prediction interfaces based on the standardized number of sensors, so as to predict the future position of the track vehicle. The model optimization module extracts direction-aware features and generates a fused feature map through a three-motion attention mechanism in the diagonal, forward, and backward directions. It also extracts path key points based on an attention-guided fast path search method and then uses a graph search algorithm to search for the optimal path among the key points. The model optimization module includes: The three motion direction modeling units are used to model the diagonal direction to detect the oblique trend of the track and obstacles; model the forward direction to enhance the accessibility analysis of the path ahead; and model the backward direction to identify obstacles and path loop features behind. The orientation masking suppression unit sets mask matrices for three motion directions, including diagonal orientation mask matrix, forward mask matrix and backward mask matrix, to suppress attention scores in irrelevant directions; The orientation mask suppression unit specifically includes: The diagonal mask matrix is ​​set to 0 when feature point positions i and j are different, and negative infinity when they are the same; the reverse mask matrix is ​​set to 0 when feature point position i is greater than j, and negative infinity otherwise; the forward mask matrix is ​​set to 0 when feature point position i is less than j, and negative infinity otherwise. The model optimization module also includes: The three-directional feature pair scoring unit is used to apply the corresponding mask matrix to any two feature points in the orbit image according to their positional relationship. It performs direction-aware transformation through a learnable mapping matrix, adds the transformed feature vectors and adds a learnable bias vector, enhances the non-linear interaction between features through the hyperbolic tangent activation function, and superimposes the mask values ​​of the corresponding directions to generate directional attention scores. The probability distribution calculation unit is used to calculate the attention score for each feature channel in three motion directions, and obtain the weight probability distribution of the three directions by normalization through the Softmax function; The feature fusion unit is used to aggregate the adjacent features of the corresponding directions according to the weight probability distribution of the three directions to generate a direction-aware feature representation; a gating mechanism is used to weight and combine the original visual features and the direction-aware features to generate a fused feature map representation; wherein, the gating mechanism learns the feature dependency degree through the Sigmoid function and adaptively adjusts the fusion ratio of the original information and the direction information. The control execution module generates control signals based on the optimal path and drives the track trolley to move.

2. The AI ​​vision-based dynamic optimization control system for the travel path of a track vehicle according to claim 1, characterized in that, The environment modeling module includes: An environmental sensing unit is used to divide the track area into a grid matrix, with each grid containing an environmental attribute identifier, including obstacles, track markings, and hazardous conditions. The trajectory prediction unit is used to predict the position of the next K time steps based on the current position and velocity vector using a kinematic model. The behavior prediction interface unit is used to integrate environmental and state data and output structured information, including: static obstacle set, dangerous area set, moving obstacle parameters, track vehicle state, and load factor.

3. The AI ​​vision-based dynamic optimization control system for the travel path of a track vehicle according to claim 1, characterized in that, The attention-guided fast path search method includes: Key point extraction steps: Extract high attention regions as key points from the fusion feature map output by the three motion attention mechanisms, and add the current position and target position of the track car to form a set of key points; Path graph construction steps: Key points are treated as nodes, neighboring nodes are connected to form edges, and each edge is assigned a weight, which is determined by the distance between nodes, the attention value at the node, and the risk value of the area traversed by the edge. Graph search steps: On the constructed path graph, use the A* search algorithm to find the optimal path from the starting point to the ending point.

4. The AI ​​vision-based dynamic optimization control system for the travel path of a track vehicle according to claim 3, characterized in that, in, The key point extraction step specifically includes: The fused feature map is compressed into a single-channel two-dimensional attention map by taking the maximum or average value. Based on threshold segmentation, positions in the attention map that are greater than or equal to τ are marked as candidate keypoints; Non-maximum suppression is applied to the candidate keypoints to obtain a sparse set of keypoints; The current position and target position of the trolley are also added to the key point set, and finally the key points of the track trolley are extracted.

5. The AI ​​vision-based dynamic optimization control system for the travel path of a track vehicle according to claim 1, characterized in that, The control execution module includes: The path smoothing unit is used to smooth the optimal path obtained by search using cubic spline interpolation or B-spline curve to obtain a smooth path. The speed planning unit is used to plan the driving speed based on the curvature of each point on the smooth path and the corresponding attention value. The control signal generation unit is used to generate steering and acceleration control commands based on the smoothed path and the planned driving speed. The real-time feedback unit is used to collect the deviation between the actual position of the track vehicle and the nearest point on the planned path at the current moment. When the triggering condition is met, a replanning instruction is triggered.

6. The AI ​​vision-based dynamic optimization control system for the travel path of a track vehicle according to claim 2, characterized in that, The data preprocessing module performs the following operations: image denoising, brightness equalization, and edge enhancement.

Images