An AI vision-based robot movement path optimization method
By introducing a candidate path probability field into DDRNet for path prior modulation and risk adaptive fusion, the problem of independent visual perception and path decision-making in robot path planning is solved, and the collaborative optimization of path trend and local risk is achieved, thereby improving the path stability and safety of the robot in complex environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- URUMQI TIANYAO WEIYE INFORMATION TECH SERVICE CO LTD
- Filing Date
- 2026-03-13
- Publication Date
- 2026-06-09
AI Technical Summary
In existing robot path planning methods, the visual network is independent of the path planning process and lacks a feedback mechanism. This results in the visual perception results not being optimized for path decision-making, and the cost field construction lacks dynamic scene feedback, leading to insufficient path stability and safety.
By introducing a candidate path probability field into the improved DDRNet, path prior modulation and risk adaptive fusion are performed to construct a risk adaptive fusion mechanism, thereby achieving collaborative optimization of path trend modeling and local risk refinement, and forming a closed-loop mechanism for path decision-making and risk response.
It improves the directional stability and global consistency of robot path search, enhances obstacle avoidance accuracy and safety in complex environments, and strengthens the robot's operational stability and path execution accuracy in dynamic scenarios.
Smart Images
Figure CN122176256A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer vision technology, and in particular to a method for optimizing robot movement paths based on AI vision. Background Technology
[0002] Autonomous robot mobility technology is widely used in warehousing and transportation, indoor inspection, unmanned delivery and industrial production scenarios. Existing path planning methods are mostly based on LiDAR to build an occupied grid map and generate feasible paths on a static cost map through search algorithms. Visual assistance solutions use convolutional neural networks for semantic segmentation, extract passable areas from images and map them to the robot coordinate system for path search. Some technologies introduce dual-resolution networks or multi-scale segmentation networks to improve semantic recognition accuracy and perform path planning and obstacle avoidance control on the generated cost map.
[0003] Existing technologies suffer from the following shortcomings: Firstly, visual networks are only used to output semantic segmentation results, and the path planning process is independent of the visual feature extraction process. Candidate path information cannot influence visual feature calculation in reverse, resulting in visual perception results not being optimized for path decision-making. Secondly, cost field construction is mostly based on single forward inference results, lacking a candidate path probability field feedback mechanism. Path search and risk assessment cannot form a closed-loop update, leading to insufficient path stability in dynamic scenes. While some methods employ dual-resolution structures to improve detail representation, they do not functionally reconstruct the dual-resolution branches. The cross-resolution interaction process lacks a risk statistical modulation mechanism, making it difficult to coordinate and unify global trends and local risks. In the path cost calculation stage, most schemes simply accumulate grid cost values without jointly weighting and controlling path smoothness and risk intensity. The motion control sequence lacks an adjustment mechanism dynamically linked to the cost field, affecting the robot's stability and safety in complex environments.
[0004] Therefore, how to provide a robot movement path optimization method based on AI vision is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0005] One objective of this invention is to propose a robot movement path optimization method based on AI vision. This invention injects the candidate path probability field feedback into a dual-resolution visual network and constructs a risk adaptive fusion mechanism to achieve collaborative optimization of path trend modeling and local risk refinement. It has the advantages of high path decision stability, timely risk response and strong environmental adaptability.
[0006] A robot movement path optimization method based on AI vision according to an embodiment of the present invention includes the following steps:
[0007] Collect robot visual image data and preprocess it to generate a standardized visual data set;
[0008] The standardized visual data set is input into the improved DDRNet. The improved DDRNet maintains the dual-resolution parallel backbone structure and cross-resolution interaction structure. Candidate path probability fields are injected into the low-resolution branch and path prior modulation is performed. Path neighborhood risk gradient refinement is performed in the high-resolution branch. The output path trend feature map and local risk feature map are then output.
[0009] A semantic risk cost map is constructed based on the path trend feature map and the local risk feature map. The semantic risk cost map is then mapped to the robot coordinate system to generate a planarable spatial cost field.
[0010] A candidate path set is generated based on a planarable spatial cost field. Each candidate path in the candidate path set is rasterized and encoded to generate a candidate path probability field.
[0011] The candidate path probability field is fed back into the low-resolution branch of the improved DDRNet. Risk adaptive fusion is performed through a cross-resolution interactive structure to update the path trend feature map and local risk feature map, and to regenerate the semantic risk cost map and the planarable spatial cost field.
[0012] The comprehensive cost of each candidate path in the candidate path set is calculated based on the updated planarable spatial cost field. The candidate path with the minimum comprehensive cost is selected as the optimal path, and the corresponding motion control sequence is generated.
[0013] The motion control sequence is sent to the robot actuator to drive the robot to move along the optimal path. During the movement, the candidate path generation and risk adaptive fusion steps are repeatedly executed until the target position is reached.
[0014] Optionally, the step of preprocessing the acquired robot visual image data to generate a standardized visual data set is as follows: acquiring continuous frame visual image data through a camera installed on the robot body and recording the corresponding timestamp information and robot pose data; performing time alignment and resampling processing on the continuous frame visual image data based on the timestamp information to generate a time-aligned visual image sequence; performing distortion correction, resolution unification, and scale normalization processing on the time-aligned visual image sequence to generate a scale-unified visual image sequence; performing pixel intensity normalization and channel normalization processing on the scale-unified visual image sequence to generate a standardized visual image sequence; and associating the standardized visual image sequence with the time-aligned robot pose data in spatial coordinates to generate a standardized visual data set.
[0015] Optionally, the improved DDRNet includes a low-resolution branch, a high-resolution branch, a cross-resolution interaction unit, a path prior injection unit, a path neighborhood risk refinement unit, and a risk adaptive fusion unit. The low-resolution branch and the high-resolution branch maintain a parallel backbone structure and propagate forward step-by-step according to preset stages. The cross-resolution interaction unit is positioned between each stage. The path prior injection unit is positioned at the input end of each stage of the low-resolution branch. The path neighborhood risk refinement unit is positioned in the middle and later stages of the high-resolution branch. The risk adaptive fusion unit is positioned at the cross-resolution interaction unit, which performs weighted fusion of the low-resolution branch feature map and the high-resolution branch feature map based on a fusion gating coefficient, outputting a path trend feature map and a local risk feature map.
[0016] Optionally, the step of injecting candidate path probability fields into the low-resolution branch and performing path prior modulation includes:
[0017] Discretize the candidate path set under a unified grid coordinate system and accumulate and normalize the path coverage times of the grid cells to generate a candidate path probability field with values ranging from 0 to 1.
[0018] At each stage input of the low-resolution branch of the improved DDRNet, the candidate path probability field is downsampled according to the spatial size of the low-resolution feature map of the current stage, and concatenated with the low-resolution feature map of the current stage in the channel dimension to generate a concatenated feature tensor.
[0019] Perform convolution operation on the spliced feature tensor to generate a path modulation weight tensor, and perform normalization processing on the path modulation weight tensor;
[0020] The path modulation weight tensor is multiplied element-wise with the current stage low-resolution feature map to obtain a weighted low-resolution feature map. The weighted low-resolution feature map is then added element-wise with the current stage low-resolution feature map to generate a path modulation low-resolution feature map. The high-probability grid region corresponding to the candidate path probability field has a higher feature response value in the path modulation low-resolution feature map than the low-probability grid region.
[0021] Optionally, the step of performing path neighborhood risk gradient refinement in the high-resolution branch includes:
[0022] The candidate path probability field is upsampled according to the spatial size of the current high-resolution feature map to generate a high-resolution path probability map. The high-resolution path probability map is then binarized according to a preset probability threshold to generate a path neighborhood mask.
[0023] The path neighborhood mask is multiplied element-wise with the current high-resolution feature map to extract the features of the path neighborhood region and generate the path neighborhood feature map.
[0024] The horizontal and vertical gradient operators are performed on the path neighborhood feature map in the spatial dimension to generate the path neighborhood gradient response map. The path neighborhood gradient response map is then subjected to amplitude synthesis processing to generate the path neighborhood gradient intensity map.
[0025] The gradient intensity map of the path neighborhood is superimposed with the high-resolution feature map of the current stage on an element-by-element weighted basis to generate a risk-refined high-resolution feature map, so that the path neighborhood region forms a gradient enhancement response in the risk-refined high-resolution feature map.
[0026] Optionally, the step of generating a programmable spatial cost field includes:
[0027] Upsampling is performed on the path trend feature map to make its spatial size more similar to the local risk features. Figure 1 Generate an alignment path trend feature map;
[0028] The alignment path trend feature map and the local risk feature map are concatenated along the channel dimension to generate a fused feature tensor. A convolution operation is then performed on the fused feature tensor to generate a semantic risk response map.
[0029] Normalization is performed on the semantic risk response map to generate a semantic risk cost map with values ranging from 0 to 1, so that each pixel value in the semantic risk cost map represents the risk cost value of the corresponding spatial location;
[0030] Based on the robot's current pose data and the camera's extrinsic parameter matrix, the semantic risk cost map is mapped from the image coordinate system to the robot coordinate system. The mapped cost is then rasterized according to a preset raster resolution to generate a planarable spatial cost field. Null value filling and boundary smoothing are then performed on the planarable spatial cost field to obtain the final planarable spatial cost field.
[0031] Optionally, the step of generating the candidate path probability field specifically includes:
[0032] In the programmable spatial cost field, the robot's starting grid cell and target grid cell are determined. A grid connectivity graph is constructed based on the grid resolution of the programmable spatial cost field, and each grid cell is assigned a cost value provided by the programmable spatial cost field.
[0033] On the grid connectivity graph, perform several path searches based on different cost weight combinations to obtain several discrete grid paths from the starting grid cell to the target grid cell. Perform smoothing processing on each discrete grid path to generate continuous candidate paths, forming a candidate path set.
[0034] Project the candidate path set onto a unified grid coordinate system, perform discrete sampling on each candidate path in the candidate path set according to a preset step size to obtain the candidate path sampling point sequence, and map the candidate path sampling point sequence to the corresponding grid cell.
[0035] Using grid cells as the statistical object, the counts of sampling points mapped to grid cells for all candidate paths in the candidate path set are accumulated to generate a path coverage count matrix;
[0036] The path coverage count matrix is normalized according to the number of candidate paths in the candidate path set to generate a candidate path probability field with values ranging from 0 to 1.
[0037] Optionally, the step of regenerating the semantic risk cost map and the programmable spatial cost field includes:
[0038] The candidate path probability field is injected again as feedback input into the input of each stage of the low-resolution branch of the improved DDRNet, and a path modulation low-resolution feature map is generated according to the path prior modulation steps.
[0039] In the improved high-resolution branch of DDRNet, a path neighborhood mask is generated based on the candidate path probability field, and a risk-refined high-resolution feature map is generated according to the path neighborhood risk gradient refinement step.
[0040] At each cross-resolution interaction node, the risk mean and risk variance are calculated within the path neighborhood mask region based on the risk-refined high-resolution feature map, and path neighborhood risk statistics are generated. Based on the path neighborhood risk statistics, risk adaptive fusion coefficients are generated through a gating function.
[0041] At each cross-resolution interaction node, the path modulation low-resolution feature map and the risk refinement high-resolution feature map are weighted and fused using the risk adaptive fusion coefficient to generate a fused feature map. The fused feature map is then fed back to the low-resolution branch and the high-resolution branch as input for the next stage to continue forward propagation.
[0042] After completing all stage calculations, the path trend feature map and local risk feature map are updated based on the fused feature map. The semantic risk cost map and the planarable spatial cost field are then regenerated according to the semantic risk cost map construction and coordinate mapping steps.
[0043] Optionally, the step of calculating the comprehensive cost of each candidate path in the candidate path set includes:
[0044] Each candidate path in the candidate path set is discretely sampled according to a preset step size to obtain a path sampling point sequence. The path sampling point sequence is then mapped to the grid coordinates in the updated planarable spatial cost field. The grid cell sequence corresponding to the path sampling point sequence is determined based on the grid coordinates.
[0045] For each grid cell in the grid cell sequence, read the updated cost value of the planarable spatial cost field, accumulate the cost value according to the path sampling point sequence to obtain the path risk cost, calculate the grid distance for the grid coordinates corresponding to adjacent path sampling points and accumulate them to obtain the path length cost.
[0046] The change in steering angle is calculated based on the displacement vectors of adjacent sampling points in the path sampling point sequence, and the path smoothing cost is accumulated. The path risk cost, path length cost, and path smoothing cost are multiplied by preset weights and then summed to obtain the comprehensive cost. The corresponding comprehensive cost is generated for each candidate path in the candidate path set.
[0047] Optionally, the motion control sequence is a control command sequence consisting of linear velocity commands and angular velocity commands arranged in chronological order. The motion control sequence is generated based on the path sampling point sequence corresponding to the optimal path. Linear velocity commands are generated based on the displacement distance between adjacent path sampling points, and angular velocity commands are generated based on the directional angular difference between adjacent path sampling points. The sequence is then time-discretized according to a preset control cycle to form continuous control commands.
[0048] The beneficial effects of this invention are:
[0049] (1) This invention introduces a candidate path probability field into the improved DDRNet and performs path prior modulation in the low-resolution branch, so that the visual feature extraction process and the path decision process are coupled. The path trend feature map has already fused candidate path information in the generation stage, avoiding the problem that the visual perception result and path planning are independent of each other, and improving the directional stability and global consistency of path search.
[0050] (2) In this invention, the path neighborhood risk gradient is refined in the high-resolution branch, and the risk adaptive fusion coefficient is generated based on the path neighborhood risk statistics at the cross-resolution interaction node. This realizes the dynamic weighted fusion of high-resolution detailed information and low-resolution global information, so that the local risk response and the global path trend are coordinated and unified, and the obstacle avoidance accuracy and safety in complex environments are improved.
[0051] (3) Based on the updated planarable spatial cost field, the present invention performs multi-dimensional comprehensive cost calculation on the candidate path set, and dynamically adjusts the motion control sequence according to the risk threshold and position error, so that path evaluation, cost field update and motion control form a closed loop mechanism, thereby improving the robot's running stability and path execution accuracy in dynamic scenarios. Attached Figure Description
[0052] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:
[0053] Figure 1 This is an overall flowchart of a robot movement path optimization method based on AI vision proposed in this invention;
[0054] Figure 2 This is a schematic diagram of the structure of an improved DDRNet based on AI vision for optimizing robot movement paths, as proposed in this invention.
[0055] Figure 3 This is a schematic diagram illustrating the generation of a candidate path set, calculation of the comprehensive cost, and generation of a motion control sequence based on a planarable spatial cost field proposed in this invention. Detailed Implementation
[0056] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.
[0057] refer to Figures 1-3 A robot movement path optimization method based on AI vision includes the following steps:
[0058] Collect robot visual image data and preprocess it to generate a standardized visual data set;
[0059] The standardized visual data set is input into the improved DDRNet. The improved DDRNet maintains the dual-resolution parallel backbone structure and cross-resolution interaction structure. Candidate path probability fields are injected into the low-resolution branch and path prior modulation is performed. Path neighborhood risk gradient refinement is performed in the high-resolution branch. The output path trend feature map and local risk feature map are then output.
[0060] A semantic risk cost map is constructed based on the path trend feature map and the local risk feature map. The semantic risk cost map is then mapped to the robot coordinate system to generate a planarable spatial cost field.
[0061] A candidate path set is generated based on a planarable spatial cost field. Each candidate path in the candidate path set is rasterized and encoded to generate a candidate path probability field.
[0062] The candidate path probability field is fed back into the low-resolution branch of the improved DDRNet. Risk adaptive fusion is performed through a cross-resolution interactive structure to update the path trend feature map and local risk feature map, and to regenerate the semantic risk cost map and the planarable spatial cost field.
[0063] The comprehensive cost of each candidate path in the candidate path set is calculated based on the updated planarable spatial cost field. The candidate path with the minimum comprehensive cost is selected as the optimal path, and the corresponding motion control sequence is generated.
[0064] The motion control sequence is sent to the robot actuator to drive the robot to move along the optimal path. During the movement, the candidate path generation and risk adaptive fusion steps are repeatedly executed until the target position is reached.
[0065] In this embodiment, the steps of collecting robot visual image data and preprocessing it to generate a standardized visual data set are as follows: acquiring continuous frame visual image data through a camera installed on the robot body and recording the corresponding timestamp information and robot pose data; performing time alignment and resampling processing on the continuous frame visual image data based on the timestamp information to generate a time-aligned visual image sequence; performing distortion correction, resolution unification, and scale normalization processing on the time-aligned visual image sequence to generate a scale-unified visual image sequence; performing pixel intensity normalization and channel normalization processing on the scale-unified visual image sequence to generate a standardized visual image sequence; and associating the standardized visual image sequence with the time-aligned robot pose data in spatial coordinates to generate a standardized visual data set.
[0066] In this embodiment, the improved DDRNet includes a low-resolution branch, a high-resolution branch, a cross-resolution interaction unit, a path prior injection unit, a path neighborhood risk refinement unit, and a risk adaptive fusion unit. The low-resolution branch and the high-resolution branch maintain a parallel backbone structure and propagate forward step-by-step according to preset stages. The cross-resolution interaction unit is located between each stage to achieve feature exchange between the low-resolution and high-resolution branches. The path prior injection unit is located at the input end of each stage of the low-resolution branch and is used to adjust the candidate path probability field to the same spatial size as the low-resolution feature map of the current stage before stitching them together in the channel dimension. The path neighborhood risk refinement unit is located in the middle and later stages of the high-resolution branch and is used to generate a path neighborhood mask based on the candidate path probability field and perform spatial gradient enhancement on the high-resolution feature map. The risk adaptive fusion unit is located at the cross-resolution interaction unit and is used to generate a fusion gating coefficient based on the risk statistics of the path neighborhood region, and perform weighted fusion of the low-resolution branch feature map and the high-resolution branch feature map based on the fusion gating coefficient, outputting a path trend feature map and a local risk feature map.
[0067] In this embodiment, the improved DDRNet enhances the structure of the low-resolution branch, high-resolution branch, and cross-resolution interaction node while keeping the original DDRNet dual-resolution parallel backbone structure and cross-resolution interaction structure unchanged.
[0068] Both the low-resolution branch and the high-resolution branch are constructed using a staged stacked residual block structure. The spatial resolution of the low-resolution branch is 1 / 8 to 1 / 32 of the input image, while the spatial resolution of the high-resolution branch is maintained at 1 / 4 of the input image. The stages exchange features bidirectionally through cross-resolution interaction units, which include upsampling convolutional layers and downsampling convolutional layers.
[0069] The path prior injection unit is set at the input end of each stage of the low-resolution branch; in the specific implementation, the candidate path probability field first performs bilinear interpolation downsampling according to the spatial size of the low-resolution feature map of the current stage, and adjusts the candidate path probability field to match the low-resolution feature map of the current stage. Figure 1 The spatial dimensions are consistent, and then a splicing operation is performed in the channel dimension. The spliced feature tensor is compressed by a 1×1 convolutional layer, and then fused by a 3×3 convolutional layer to generate a path-modulated low-resolution feature map.
[0070] The path neighborhood risk refinement unit is set in the middle and later stages of the high-resolution branch. In the specific implementation, the candidate path probability field is segmented by a threshold to generate a path neighborhood mask. The path neighborhood mask is multiplied element-wise with the high-resolution feature map to retain the features of the path neighborhood region. Then, the spatial gradient operator is performed on the retained region to obtain the gradient response map. The gradient response map is then superimposed element-wise with the original high-resolution feature map to generate the risk refinement high-resolution feature map.
[0071] The risk adaptive fusion unit is set at each cross-resolution interaction node. Before fusion, the path neighborhood region features are extracted from the risk refinement high-resolution feature map, and the risk mean and risk variance of the path neighborhood region are calculated. Based on the risk mean and risk variance, fusion gate coefficients are generated through a fully connected layer and a sigmoid activation function. The fusion gate coefficients are used to adjust the weighting ratio of the low-resolution branch feature map and the high-resolution branch feature map during fusion to achieve adaptive weighted fusion. The fused feature map is fed back to the low-resolution branch and the high-resolution branch respectively to continue forward propagation.
[0072] After all stages of computation, the low-resolution branch outputs a path trend feature map, and the high-resolution branch outputs a local risk feature map. The path trend feature map and the local risk feature map are used together to construct the semantic risk cost map and generate the planarable spatial cost field.
[0073] In this embodiment, the steps of injecting candidate path probability fields into the low-resolution branch and performing path prior modulation include:
[0074] Discretize the candidate path set under a unified grid coordinate system and accumulate and normalize the path coverage times of the grid cells to generate a candidate path probability field with values ranging from 0 to 1.
[0075] At each stage input of the low-resolution branch of the improved DDRNet, the candidate path probability field is downsampled according to the spatial size of the low-resolution feature map of the current stage, and concatenated with the low-resolution feature map of the current stage in the channel dimension to generate a concatenated feature tensor.
[0076] Perform convolution operation on the spliced feature tensor to generate a path modulation weight tensor, and perform normalization processing on the path modulation weight tensor;
[0077] The path modulation weight tensor is multiplied element-wise with the current stage low-resolution feature map to obtain a weighted low-resolution feature map. The weighted low-resolution feature map is then added element-wise with the current stage low-resolution feature map to generate a path modulation low-resolution feature map. The high-probability grid region corresponding to the candidate path probability field has a higher feature response value in the path modulation low-resolution feature map than the low-probability grid region.
[0078] In this embodiment, the path-modulated low-resolution feature map is input into the next stage of the low-resolution branch to participate in cross-resolution interaction and path trend feature map generation. After the path-modulated low-resolution feature map is generated, it is used as the output feature of the current stage's low-resolution branch and input to the next stage's residual convolution unit to perform convolution and nonlinear activation operations, generating the next stage's low-resolution feature map. At the cross-resolution interaction nodes of each stage, the next stage's low-resolution feature map is upsampled and mapped to the current high-resolution feature map. Figure 1 After achieving the desired spatial dimensions, feature fusion is performed, and the current high-resolution feature map is downsampled and mapped to the low-resolution feature map of the next stage. Figure 1 After achieving the desired spatial dimensions, feature fusion is performed. The fused low-resolution feature map continues to be used as the input for the next stage, propagating it step by step so that the candidate path probability field forms a continuous modulation effect within the low-resolution branch. At the end of the low-resolution branch, a path trend feature map is generated through convolution mapping.
[0079] In this embodiment, the steps of performing path neighborhood risk gradient refinement in the high-resolution branch include:
[0080] The candidate path probability field is upsampled according to the spatial size of the current high-resolution feature map to generate a high-resolution path probability map. The high-resolution path probability map is then binarized according to a preset probability threshold to generate a path neighborhood mask.
[0081] The path neighborhood mask is multiplied element-wise with the current high-resolution feature map to extract the features of the path neighborhood region and generate the path neighborhood feature map.
[0082] The horizontal and vertical gradient operators are performed on the path neighborhood feature map in the spatial dimension to generate the path neighborhood gradient response map. The path neighborhood gradient response map is then subjected to amplitude synthesis processing to generate the path neighborhood gradient intensity map.
[0083] The gradient intensity map of the path neighborhood is superimposed with the high-resolution feature map of the current stage on an element-by-element weighted basis to generate a risk-refined high-resolution feature map, so that the path neighborhood region forms a gradient enhancement response in the risk-refined high-resolution feature map.
[0084] In this embodiment, the risk-refined high-resolution feature map is input into the next stage of the high-resolution branch to participate in cross-resolution interaction and local risk feature map generation. After the risk-refined high-resolution feature map is generated, it is used as the output feature of the current stage's high-resolution branch and input to the next stage's convolutional unit to perform convolution and nonlinear activation operations, generating the next stage's high-resolution feature map. At the cross-resolution interaction nodes of each stage, the next stage's high-resolution feature map is downsampled and mapped to the current low-resolution feature map. Figure 1 After achieving the desired spatial dimensions, feature fusion is performed, and the current low-resolution feature map is upsampled and mapped to the high-resolution feature map of the next stage. Figure 1 After achieving the desired spatial dimensions, feature fusion is performed. The fused high-resolution feature map continues to propagate step by step within the high-resolution branch, so that the risk gradient information in the path neighborhood retains spatial details and is integrated with global semantic information during the multi-stage convolution process. At the end of the high-resolution branch, a local risk feature map is generated through convolution mapping.
[0085] In this embodiment, the step of generating a planarable spatial cost field includes:
[0086] Upsampling is performed on the path trend feature map to make its spatial size more similar to the local risk features. Figure 1 Generate an alignment path trend feature map;
[0087] The alignment path trend feature map and the local risk feature map are concatenated along the channel dimension to generate a fused feature tensor. A convolution operation is then performed on the fused feature tensor to generate a semantic risk response map.
[0088] Normalization is performed on the semantic risk response map to generate a semantic risk cost map with values ranging from 0 to 1, so that each pixel value in the semantic risk cost map represents the risk cost value of the corresponding spatial location;
[0089] Based on the robot's current pose data and the camera's extrinsic parameter matrix, the semantic risk cost map is mapped from the image coordinate system to the robot coordinate system. The mapped cost is then rasterized according to a preset raster resolution to generate a planarable spatial cost field. Null value filling and boundary smoothing are then performed on the planarable spatial cost field to obtain the final planarable spatial cost field.
[0090] In this embodiment, the steps for generating the candidate path probability field specifically include:
[0091] In the programmable spatial cost field, the robot's starting grid cell and target grid cell are determined. A grid connectivity graph is constructed based on the grid resolution of the programmable spatial cost field, and each grid cell is assigned a cost value provided by the programmable spatial cost field.
[0092] On the grid connectivity graph, perform several path searches based on different cost weight combinations to obtain several discrete grid paths from the starting grid cell to the target grid cell. Perform smoothing processing on each discrete grid path to generate continuous candidate paths, forming a candidate path set.
[0093] Project the candidate path set onto a unified grid coordinate system, perform discrete sampling on each candidate path in the candidate path set according to a preset step size to obtain the candidate path sampling point sequence, and map the candidate path sampling point sequence to the corresponding grid cell.
[0094] Using grid cells as the statistical object, the counts of sampling points mapped to grid cells for all candidate paths in the candidate path set are accumulated to generate a path coverage count matrix;
[0095] The path coverage count matrix is normalized according to the number of candidate paths in the candidate path set to generate a candidate path probability field with values ranging from 0 to 1.
[0096] In this embodiment, the steps of regenerating the semantic risk cost map and the programmable spatial cost field include:
[0097] The candidate path probability field is injected again as feedback input into the input of each stage of the low-resolution branch of the improved DDRNet, and a path modulation low-resolution feature map is generated according to the path prior modulation steps.
[0098] In the improved high-resolution branch of DDRNet, a path neighborhood mask is generated based on the candidate path probability field, and a risk-refined high-resolution feature map is generated according to the path neighborhood risk gradient refinement step.
[0099] At each cross-resolution interaction node, the risk mean and risk variance are calculated within the path neighborhood mask region based on the risk-refined high-resolution feature map, and path neighborhood risk statistics are generated. Based on the path neighborhood risk statistics, risk adaptive fusion coefficients are generated through a gating function.
[0100] Specifically, calculating the risk mean and risk variance involves: first, obtaining a refined high-resolution feature map of the risk and a path neighborhood mask, where a value of 1 in the path neighborhood mask represents a path neighborhood region; then, copying and expanding the path neighborhood mask in the channel dimension to match the number of channels in the refined high-resolution feature map of the risk; and finally, performing element-wise multiplication with the refined high-resolution feature map of the risk to obtain a path neighborhood feature tensor that contains only the feature responses of the path neighborhood region.
[0101] Within the spatial range where the path neighborhood mask value is 1, the path neighborhood feature tensor is statistically calculated for each channel to obtain the mean and variance values of each channel within the path neighborhood region. The mean values of all channels are combined to form a risk mean vector, and the variance values of all channels are combined to form a risk variance vector. The risk mean vector and the risk variance vector together constitute the path neighborhood risk statistic, which is used to calculate the risk adaptive fusion coefficient.
[0102] In this embodiment, the path neighborhood risk statistics are used as the input to the gating function. The path neighborhood risk statistics include the risk mean and risk variance of each channel within the path neighborhood. First, the risk mean vector and the risk variance vector are concatenated by channel to form a risk statistical feature vector. Then, the risk statistical feature vector is input to the fully connected mapping layer to perform a linear transformation to obtain the fusion modulation vector. A nonlinear activation operation is performed on the fusion modulation vector to limit the output value to the range of 0 to 1, thereby generating risk adaptive fusion coefficients.
[0103] The risk adaptive fusion coefficient is used to represent the adjustment weight of the current path neighborhood risk intensity on the dual-resolution feature fusion ratio. When the path neighborhood risk statistic increases, the risk adaptive fusion coefficient approaches the range of values that enhance the weight of high-resolution features; when the path neighborhood risk statistic decreases, the risk adaptive fusion coefficient approaches the range of values that enhance the weight of low-resolution features.
[0104] The generated risk-adaptive fusion coefficients are input to the cross-resolution interaction node to perform a weighted fusion operation on the path-modulated low-resolution feature map and the risk-refined high-resolution feature map.
[0105] At each cross-resolution interaction node, the path modulation low-resolution feature map and the risk refinement high-resolution feature map are weighted and fused using the risk adaptive fusion coefficient to generate a fused feature map. The fused feature map is then fed back to the low-resolution branch and the high-resolution branch as input for the next stage to continue forward propagation.
[0106] After completing all stage calculations, the path trend feature map and local risk feature map are updated based on the fused feature map. The semantic risk cost map and the planarable spatial cost field are then regenerated according to the semantic risk cost map construction and coordinate mapping steps.
[0107] In this embodiment, the fused feature maps output by the cross-resolution interactive nodes are input into the low-resolution branch and the high-resolution branch respectively for further forward propagation. At the end of each branch, a new path trend feature map and a new local risk feature map are generated through a convolutional mapping layer. The new path trend feature map is then superimposed with the previous path trend feature map on an element-wise weighted basis to generate an updated path trend feature map. Similarly, the new local risk feature map is superimposed with the previous local risk feature map on an element-wise weighted basis to generate an updated local risk feature map. The weighting coefficients are determined by the risk adaptive fusion coefficients, enabling the feedback effect of the candidate path probability field to participate in the iterative update of the path trend feature map and the local risk feature map through the fused feature map.
[0108] In this embodiment, the step of calculating the comprehensive cost of each candidate path in the candidate path set includes:
[0109] Each candidate path in the candidate path set is discretely sampled according to a preset step size to obtain a path sampling point sequence. The path sampling point sequence is then mapped to the grid coordinates in the updated planarable spatial cost field. The grid cell sequence corresponding to the path sampling point sequence is determined based on the grid coordinates.
[0110] For each grid cell in the grid cell sequence, read the updated cost value of the planarable spatial cost field, accumulate the cost value according to the path sampling point sequence to obtain the path risk cost, calculate the grid distance for the grid coordinates corresponding to adjacent path sampling points and accumulate them to obtain the path length cost.
[0111] The change in steering angle is calculated based on the displacement vectors of adjacent sampling points in the path sampling point sequence, and the path smoothing cost is accumulated. The path risk cost, path length cost, and path smoothing cost are multiplied by preset weights and then summed to obtain the comprehensive cost. The corresponding comprehensive cost is generated for each candidate path in the candidate path set.
[0112] In this embodiment, the motion control sequence is a control command sequence consisting of linear velocity commands and angular velocity commands arranged in chronological order. The motion control sequence is generated based on the path sampling point sequence corresponding to the optimal path. Linear velocity commands are generated based on the displacement distance between adjacent path sampling points, and angular velocity commands are generated based on the angular difference between adjacent path sampling points. The sequence is then time-discretized according to a preset control cycle to form continuous control commands.
[0113] Specifically, when the cost value of the corresponding grid cell ahead of the optimal path in the updated planarable space cost field is higher than the preset risk threshold, the value of the linear velocity command is reduced and the adjustment range of the angular velocity command is increased to slow down the robot and adjust its direction of travel; when the cost value of the corresponding grid cell ahead of the optimal path is lower than the preset risk threshold, the value of the linear velocity command is increased and the adjustment range of the angular velocity command is decreased to keep the robot moving steadily in a straight line.
[0114] When the position error between the robot's current position and the target sampling point corresponding to the optimal path is greater than a preset error threshold, the subsequent linear velocity and angular velocity commands are proportionally corrected to generate a corrected motion control sequence; when the position error is less than the preset error threshold, the original motion control sequence is executed until the target position is reached.
[0115] Example 1: To verify the feasibility of the present invention in practice, the present invention was applied to an autonomous mobile robot navigation scenario in a complex indoor passageway environment. The scenario includes narrow passages, dynamic pedestrian interference areas, temporary obstacle stacking areas, and wide open areas. The environment contains both static structural boundaries and randomly moving targets. The visual image simultaneously includes ground reflections, shadow changes, and partial occlusion. In this environment, the traditional path planning method based on constructing a cost graph from a single semantic segmentation result is prone to path oscillation, frequent replanning, and delayed obstacle avoidance response. Especially when a dynamic target suddenly appears in front of the path, the robot may not decelerate in time or take an excessively long detour.
[0116] In this embodiment, the robot is equipped with a forward-looking vision acquisition device to continuously acquire visual image data and generate a standardized visual data set. The standardized visual data set is input into the improved DDRNet. After receiving the candidate path probability field, the low-resolution branch performs path prior modulation to make the global semantic feature extraction process biased towards the candidate path coverage area. The high-resolution branch performs risk gradient refinement under the constraint of the path neighborhood mask to enhance the response of the path neighborhood boundary. The cross-resolution interaction structure generates risk adaptive fusion coefficients based on the path neighborhood risk statistics, adjusts the fusion ratio of low-resolution features and high-resolution features, and outputs a path trend feature map and a local risk feature map. A semantic risk cost map is constructed based on the two types of feature maps and mapped to the robot coordinate system to generate a planarable spatial cost field. A candidate path set is generated based on the cost field and a candidate path probability field is formed. The candidate path probability field is fed back to the improved DDRNet low-resolution branch to form a closed-loop update of path trend and risk details.
[0117] In continuous testing, three environmental complexity scenarios were set up: low-density obstacle scenario, medium-density dynamic interference scenario, and high-density dynamic intersection scenario. At least 100 path planning and execution tests were conducted for each scenario. Test results show that in the low-density obstacle scenario, the traditional visual segmentation plus cost accumulation method has an average path length of 18.6 meters, an average arrival time of 32.4 seconds, and an average number of replanning attempts of 2.8. Using the method of this invention, the average path length is 17.9 meters, the average arrival time is 29.7 seconds, and the average number of replanning attempts is 1.3. In the medium-density dynamic interference scenario, the traditional method has an average path deviation rate of 12.4% and an obstacle avoidance delay response time of 0.82 seconds. The method of this invention reduces the average path deviation rate to 6.7% and the obstacle avoidance delay response time to 0.41 seconds. In the high-density dynamic intersection scenario, the traditional method experiences local path oscillations in 18% of cases, with a speed fluctuation standard deviation of 0.56 during path execution. The method of this invention reduces the path oscillation rate to 4% and the speed fluctuation standard deviation to 0.21.
[0118] Regarding risk response accuracy, the accuracy of risk identification in the neighborhood of the statistical path is improved. Traditional methods achieve an accuracy of 83.5% for risk identification in the neighborhood of dynamic targets, while the method of this invention achieves 94.2%. The accuracy for risk identification in narrow-boundary areas is increased from 78.6% to 92.8%. Under the closed-loop update mechanism, after feedback from the candidate path probability field, a stable path converges in an average of two iterations, reducing the fluctuation range of the overall path cost from 15.2% in the original method to 4.3%.
[0119] During the motion control sequence execution phase, the fluctuation range of linear velocity commands and the number of steering corrections are statistically analyzed. Traditional methods, in high-density dynamic scenarios, have an average linear velocity fluctuation range of 0.45 m / s and 9.6 steering corrections. The method of this invention reduces the linear velocity fluctuation range to 0.18 m / s and the number of steering corrections to 4.1. Regarding position error control, the average endpoint arrival error is reduced from 0.27 m to 0.11 m.
[0120] Comprehensive testing shows that the present invention, through the candidate path probability field feedback mechanism and the risk adaptive fusion mechanism, enables the path trend modeling and local risk modeling to form a coupled update relationship, and the cost field is corrected in real time with changes in visual features, thereby improving path stability and dynamic environment adaptability.
[0121] Table 1: Comparison of Path Optimization Performance in Complex Environments
[0122] Indicator Categories Traditional visual planning methods Method of the present invention Average path length (meters) in low-density scenes 18.6 17.9 Average arrival time (seconds) for low-density scenes 32.4 29.7 Average number of replanning operations in low-density scenarios 2.8 1.3 Path offset rate (%) for medium-density scenes 12.4 6.7 Obstacle avoidance delay response time (seconds) in medium-density scenes 0.82 0.41 Path oscillation rate in high-density scenarios (%) 18 4 Standard deviation of velocity fluctuation in high-density scenes 0.56 0.21 Accuracy rate of dynamic target neighborhood risk identification (%) 83.5 94.2 Accuracy rate of risk identification in narrow areas (%) 78.6 92.8 Overall cost fluctuation range (%) 15.2 4.3 Average number of iterations to converge 4.7 2.1 Linear velocity fluctuation range (meters per second) 0.45 0.18 Number of steering corrections 9.6 4.1 Endpoint arrival error (meters) 0.27 0.11
[0123] As shown in Table 1, in low-density scenarios, the method of this invention outperforms traditional visual planning methods in both average path length and average arrival time. The average path length is reduced from 18.6 meters to 17.9 meters, resulting in a more reasonable path; the average arrival time is reduced from 32.4 seconds to 29.7 seconds, indicating improved path decision-making efficiency. Meanwhile, the average number of replanning operations is reduced from 2.8 to 1.3, demonstrating that the fusion mechanism of path trend modeling and risk adaptive optimization reduces path oscillations and frequent recalculation, thereby improving path stability.
[0124] In medium-density dynamic interference scenarios, the path offset rate of the method of this invention decreased from 12.4% to 6.7%, with the offset amplitude reduced by nearly half, reflecting a significant enhancement in the constraint effect of the path trend feature map on the global direction. The obstacle avoidance delay response time decreased from 0.82 seconds to 0.41 seconds, a reduction of approximately 50%, demonstrating the rapid perception capability of the path neighborhood risk gradient refinement and risk statistics-driven fusion mechanism for dynamic target changes.
[0125] In high-density dynamic intersection scenarios, the path oscillation rate decreased from 18% to 4%, and the standard deviation of velocity fluctuation decreased from 0.56 to 0.21, indicating that the motion control sequence is smoother and more stable in complex environments. The accuracy of dynamic target neighborhood risk identification improved from 83.5% to 94.2%, and the accuracy of narrow region risk identification improved from 78.6% to 92.8%, demonstrating the ability of the local risk feature map to accurately represent boundaries and dynamic targets. The overall cost fluctuation amplitude decreased from 15.2% to 4.3%, and the average number of iterations for convergence decreased from 4.7 to 2.1, indicating that the candidate path probability field feedback mechanism accelerated the cost field convergence process.
[0126] In terms of online speed and control accuracy, the linear speed fluctuation range decreased from 0.45 m / s to 0.18 m / s, the number of steering corrections decreased from 9.6 to 4.1, and the endpoint arrival error decreased from 0.27 m to 0.11 m. This indicates that the integrated cost calculation and dynamic adjustment mechanism of the motion control sequence make the path execution more stable and accurate. The overall data shows that the present invention has an improvement effect on path stability, risk response speed, cost field convergence efficiency, and endpoint control accuracy.
[0127] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. An AI vision-based robot movement path optimization method, characterized by, Includes the following steps: Collect robot visual image data and preprocess it to generate a standardized visual data set; The standardized visual data set is input into the improved DDRNet. The improved DDRNet maintains the dual-resolution parallel backbone structure and cross-resolution interaction structure. Candidate path probability fields are injected into the low-resolution branch and path prior modulation is performed. Path neighborhood risk gradient refinement is performed in the high-resolution branch. The output path trend feature map and local risk feature map are then output. A semantic risk cost map is constructed based on the path trend feature map and the local risk feature map. The semantic risk cost map is then mapped to the robot coordinate system to generate a planarable spatial cost field. A candidate path set is generated based on a planarable spatial cost field. Each candidate path in the candidate path set is rasterized and encoded to generate a candidate path probability field. The candidate path probability field is fed back into the low-resolution branch of the improved DDRNet. Risk adaptive fusion is performed through a cross-resolution interactive structure to update the path trend feature map and local risk feature map, and to regenerate the semantic risk cost map and the planarable spatial cost field. The comprehensive cost of each candidate path in the candidate path set is calculated based on the updated planarable spatial cost field. The candidate path with the minimum comprehensive cost is selected as the optimal path, and the corresponding motion control sequence is generated. The motion control sequence is sent to the robot actuator to drive the robot to move along the optimal path. During the movement, the candidate path generation and risk adaptive fusion steps are repeatedly executed until the target position is reached.
2. The AI vision-based robot movement path optimization method of claim 1, wherein, The steps for preprocessing the acquired robot visual image data to generate a standardized visual data set are as follows: acquiring continuous frame visual image data using a camera installed on the robot body and recording corresponding timestamp information and robot pose data; performing time alignment and resampling processing on the continuous frame visual image data based on the timestamp information to generate a time-aligned visual image sequence; performing distortion correction, resolution unification, and scale normalization processing on the time-aligned visual image sequence to generate a scale-unified visual image sequence; performing pixel intensity normalization and channel normalization processing on the scale-unified visual image sequence to generate a standardized visual image sequence; and associating the standardized visual image sequence with the time-aligned robot pose data using spatial coordinates to generate a standardized visual data set. 3.The AI vision-based robot movement path optimization method of claim 2, wherein, The improved DDRNet includes a low-resolution branch, a high-resolution branch, a cross-resolution interaction unit, a path prior injection unit, a path neighborhood risk refinement unit, and a risk adaptive fusion unit. The low-resolution branch and the high-resolution branch maintain a parallel backbone structure and propagate forward step by step according to preset stages. The cross-resolution interaction unit is set between each stage. The path prior injection unit is set at the input end of each stage of the low-resolution branch. The path neighborhood risk refinement unit is set in the middle and later stages of the high-resolution branch. The risk adaptive fusion unit is set at the cross-resolution interaction unit, which performs weighted fusion of the feature maps of the low-resolution branch and the high-resolution branch based on the fusion gating coefficient, and outputs a path trend feature map and a local risk feature map.
4. The AI vision-based robot movement path optimization method of claim 3, wherein, The steps of injecting candidate path probability fields into the low-resolution branch and performing path prior modulation include: Discretize the candidate path set under a unified grid coordinate system and accumulate and normalize the path coverage times of the grid cells to generate a candidate path probability field with values ranging from 0 to 1. At each stage input of the low-resolution branch of the improved DDRNet, the candidate path probability field is downsampled according to the spatial size of the low-resolution feature map of the current stage, and concatenated with the low-resolution feature map of the current stage in the channel dimension to generate a concatenated feature tensor. Perform convolution operation on the spliced feature tensor to generate a path modulation weight tensor, and perform normalization processing on the path modulation weight tensor; The path modulation weight tensor is multiplied element-wise with the current stage low-resolution feature map to obtain a weighted low-resolution feature map. The weighted low-resolution feature map is then added element-wise with the current stage low-resolution feature map to generate a path modulation low-resolution feature map. The high-probability grid region corresponding to the candidate path probability field has a higher feature response value in the path modulation low-resolution feature map than the low-probability grid region.
5. The AI vision-based robot movement path optimization method of claim 4, wherein, The steps for performing path neighborhood risk gradient refinement in the high-resolution branch include: The candidate path probability field is upsampled according to the spatial size of the current high-resolution feature map to generate a high-resolution path probability map. The high-resolution path probability map is then binarized according to a preset probability threshold to generate a path neighborhood mask. The path neighborhood mask is multiplied element-wise with the current high-resolution feature map to extract the features of the path neighborhood region and generate the path neighborhood feature map. The horizontal and vertical gradient operators are performed on the path neighborhood feature map in the spatial dimension to generate the path neighborhood gradient response map. The path neighborhood gradient response map is then subjected to amplitude synthesis processing to generate the path neighborhood gradient intensity map. The gradient intensity map of the path neighborhood is superimposed with the high-resolution feature map of the current stage on an element-by-element weighted basis to generate a risk-refined high-resolution feature map, so that the path neighborhood region forms a gradient enhancement response in the risk-refined high-resolution feature map.
6. The robot movement path optimization method based on AI vision according to claim 5, characterized in that, The step of generating a programmable spatial cost field includes: Upsampling is performed on the path trend feature map to make its spatial size consistent with the local risk feature map, thus generating an aligned path trend feature map. The alignment path trend feature map and the local risk feature map are concatenated along the channel dimension to generate a fused feature tensor. A convolution operation is then performed on the fused feature tensor to generate a semantic risk response map. Normalization is performed on the semantic risk response map to generate a semantic risk cost map with values ranging from 0 to 1, so that each pixel value in the semantic risk cost map represents the risk cost value of the corresponding spatial location; Based on the robot's current pose data and the camera's extrinsic parameter matrix, the semantic risk cost map is mapped from the image coordinate system to the robot coordinate system. The mapped cost is then rasterized according to a preset raster resolution to generate a planarable spatial cost field. Null value filling and boundary smoothing are then performed on the planarable spatial cost field to obtain the final planarable spatial cost field.
7. The robot movement path optimization method based on AI vision according to claim 6, characterized in that, The steps for generating the candidate path probability field specifically include: In the programmable spatial cost field, the robot's starting grid cell and target grid cell are determined. A grid connectivity graph is constructed based on the grid resolution of the programmable spatial cost field, and each grid cell is assigned a cost value provided by the programmable spatial cost field. On the grid connectivity graph, perform several path searches based on different cost weight combinations to obtain several discrete grid paths from the starting grid cell to the target grid cell. Perform smoothing processing on each discrete grid path to generate continuous candidate paths, forming a candidate path set. Project the candidate path set onto a unified grid coordinate system, perform discrete sampling on each candidate path in the candidate path set according to a preset step size to obtain the candidate path sampling point sequence, and map the candidate path sampling point sequence to the corresponding grid cell. Using grid cells as the statistical object, the counts of sampling points mapped to grid cells for all candidate paths in the candidate path set are accumulated to generate a path coverage count matrix; The path coverage count matrix is normalized according to the number of candidate paths in the candidate path set to generate a candidate path probability field with values ranging from 0 to 1.
8. The robot movement path optimization method based on AI vision according to claim 7, characterized in that, The steps for regenerating the semantic risk cost map and the programmable spatial cost field include: The candidate path probability field is injected again as feedback input into the input of each stage of the low-resolution branch of the improved DDRNet, and a path modulation low-resolution feature map is generated according to the path prior modulation steps. In the improved high-resolution branch of DDRNet, a path neighborhood mask is generated based on the candidate path probability field, and a risk-refined high-resolution feature map is generated according to the path neighborhood risk gradient refinement step. At each cross-resolution interaction node, the risk mean and risk variance are calculated within the path neighborhood mask region based on the risk-refined high-resolution feature map, and path neighborhood risk statistics are generated. Based on the path neighborhood risk statistics, risk adaptive fusion coefficients are generated through a gating function. At each cross-resolution interaction node, the path modulation low-resolution feature map and the risk refinement high-resolution feature map are weighted and fused using the risk adaptive fusion coefficient to generate a fused feature map. The fused feature map is then fed back to the low-resolution branch and the high-resolution branch as input for the next stage to continue forward propagation. After completing all stage calculations, the path trend feature map and local risk feature map are updated based on the fused feature map. The semantic risk cost map and the planarable spatial cost field are then regenerated according to the semantic risk cost map construction and coordinate mapping steps.
9. The robot movement path optimization method based on AI vision according to claim 8, characterized in that, The step of calculating the comprehensive cost of each candidate path in the candidate path set includes: Each candidate path in the candidate path set is discretely sampled according to a preset step size to obtain a path sampling point sequence. The path sampling point sequence is then mapped to the grid coordinates in the updated planarable spatial cost field. The grid cell sequence corresponding to the path sampling point sequence is determined based on the grid coordinates. For each grid cell in the grid cell sequence, read the updated cost value of the planarable spatial cost field, accumulate the cost value according to the path sampling point sequence to obtain the path risk cost, calculate the grid distance for the grid coordinates corresponding to adjacent path sampling points and accumulate them to obtain the path length cost. The change in steering angle is calculated based on the displacement vectors of adjacent sampling points in the path sampling point sequence, and the path smoothing cost is accumulated. The path risk cost, path length cost, and path smoothing cost are multiplied by preset weights and then summed to obtain the comprehensive cost. The corresponding comprehensive cost is generated for each candidate path in the candidate path set.
10. The robot movement path optimization method based on AI vision according to claim 9, characterized in that, The motion control sequence is a sequence of control commands consisting of linear velocity commands and angular velocity commands arranged in chronological order. The motion control sequence is generated based on the path sampling point sequence corresponding to the optimal path. Linear velocity commands are generated based on the displacement distance between adjacent path sampling points, and angular velocity commands are generated based on the angular difference between adjacent path sampling points. The sequence is then discretized in time according to a preset control period to form continuous control commands.