A robot travel route planning and obstacle avoidance system based on AI vision detection

Abstract

This invention discloses a robot path planning and obstacle avoidance system based on AI vision detection, comprising: an image acquisition and preprocessing module that acquires camera image data to obtain a preprocessed image sequence; an improved BEVFormer perception module that constructs an improved BEVFormer model and outputs a bird's-eye view occupied grid sequence and a risk grid sequence; a safety expansion processing module that obtains a local traversable region; a polar coordinate mapping module that constructs a polar coordinate traversable region representation; a vector field histogram planning module that, based on the vector field histogram method, obtains a set of traversable angle intervals and a channel center ridge sequence; a route generation and control module that calculates control commands; and a validity verification and execution module that drives the robot's movement through the control commands. This invention, by combining the improved BEVFormer model and the vector field histogram method, achieves stable route planning and smooth obstacle avoidance closed-loop control for the robot.

Description

Technical Field

[0001] This invention relates to the field of AI vision technology, and in particular to a robot path planning and obstacle avoidance system based on AI vision detection. Background Technology

[0002] Existing mobile robots in warehousing, park inspection, and indoor service scenarios typically need to achieve autonomous movement, path planning, and obstacle avoidance in complex environments. Current solutions often use LiDAR or ultrasonic sensors to construct local occupancy grids and combine them with local planning algorithms to output control commands. In vision-only solutions, obstacle information is often obtained through visual detection methods such as object detection and semantic segmentation, and the detection results are then converted into grid maps or cost maps for planning. However, due to the influence of lighting variations, reflections, occlusion, texture loss, and motion blur on camera imaging, visual detection results are prone to jitter, false positives, and false negatives. Furthermore, detection results are usually output as bounding boxes or pixel masks, lacking a stable, structured representation for planning. This leads to discontinuous or unstable traversable areas during the conversion from visual detection results to an environment representation usable for path planning, resulting in robot path jitter, misrouting, and even planning failures in narrow passages or areas with dense obstacles.

[0003] In dynamic scenarios with pedestrians, vehicles, or other moving objects, the environmental occupancy state changes rapidly over time, requiring real-time closed-loop coordination between planning, obstacle avoidance, and control. Existing visual navigation systems typically generate the current occupancy representation based on single-frame or short-term observations, followed by real-time decision-making by local obstacle avoidance algorithms. This approach struggles to effectively represent occupancy changes and risk evolution over a future period. Regarding local obstacle avoidance, common reactive methods using vector field histograms often rely on angle histogram thresholds to define passable directions. These methods are susceptible to perceptual noise and discretization, leading to directional jumps, obstacle-hugging maneuvers, or local oscillations. Furthermore, they may exhibit reaction lag and unsmooth obstacle avoidance when facing dynamic obstacles. When perceptual uncertainty increases or control execution is delayed, the lack of effective verification and replanning trigger mechanisms for future occupancy and risk assessment further increases collision risk and reduces system reliability and safety.

[0004] Therefore, how to provide a robot path planning and obstacle avoidance system based on AI visual detection 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 path planning and obstacle avoidance system based on AI visual detection. This invention comprehensively utilizes multi-camera visual perception, bird's-eye view spatial encoding, local path planning, and control closed-loop verification technologies to form a complete process from camera image acquisition and preprocessing, improved BEVFormer perception modeling, local traversable domain generation, polar coordinate mapping, vector field histogram planning, route and control command generation to validity determination and execution. In terms of model structure, it innovatively adopts corridor anchor point query, safety boundary driven sampling zone, and temporal occupancy prediction attention output bird's-eye view occupancy grid sequence and risk grid sequence. In terms of local planning, it uses angle interval map and polar coordinate distance field ridge line extraction to obtain the set of traversable angle intervals and the channel center ridge line sequence, realizing a stable transformation of visual detection results into a planned usable environment representation and real-time closed-loop coordination of planning, obstacle avoidance, and control under dynamic obstacles and perception uncertainties. Compared with existing technologies, this invention has the advantages of more stable route output, smoother dynamic obstacle avoidance, lower collision risk, and strong engineering feasibility.

[0006] According to an embodiment of the present invention, a robot path planning and obstacle avoidance system based on AI visual detection includes: The image acquisition and preprocessing module acquires camera image data from the robot, preprocesses the camera image data, and obtains a preprocessed image sequence. Improve the BEVFormer perception module, construct an improved BEVFormer model, perform feature extraction and bird's-eye view spatial encoding on the preprocessed image sequence, perform corridor-level feature aggregation based on corridor anchor point query, introduce safety boundary-driven sampling band to determine cross-view sampling position, use temporal occupancy prediction attention to perform temporal fusion and prediction, and output bird's-eye view occupancy grid sequence and risk grid sequence. The safety expansion processing module expands the impassable area based on the robot's safety margin to obtain a locally passable area; The polar coordinate mapping module maps the local traversable region to a polar coordinate grid with the robot as the origin, constructs a polar coordinate traversable region representation, and determines the target direction. The vector field histogram planning module, based on the vector field histogram method, uses an angle interval map to construct a set of passable angle intervals and interval connectivity by representing the passable domain in polar coordinates. It also constructs a polar coordinate distance field by extracting the ridge line of the polar coordinate distance field, thus obtaining the set of passable angle intervals and the sequence of channel center ridge lines. The route generation and control module generates a local travel route and calculates control commands based on the target direction, the set of passable angle intervals, and the sequence of the central ridge line of the channel. The validity verification execution module performs real-time validity determination of the local travel route based on the bird's-eye view occupied grid sequence and risk grid sequence, and drives the robot to move through control commands.

[0007] Optionally, the camera image data includes original camera image frames, corresponding timestamp information, camera intrinsic data, camera extrinsic data, and camera image acquisition parameters.

[0008] Optionally, obtaining the preprocessed image sequence includes: The system acquires raw image frames from the camera, reads the timestamp information corresponding to each frame, performs frame-level synchronization and pairing of raw image frames from different cameras according to their timestamps, generates a multi-camera synchronized image frame group arranged in chronological order, and removes image frames whose timestamp deviation exceeds the synchronization threshold. Distortion correction and pixel coordinate remapping are performed on each original image frame in the camera synchronization image frame group. The image is cropped and scaled according to the input resolution. The image is color space consistent to obtain a corrected image frame with a unified color channel expression. The pixel values ​​of the calibrated image frames are normalized, the data format is converted, the image data is converted into a preset tensor format, the channel order is arranged, the corresponding camera intrinsic and extrinsic data are associated with each calibrated image frame, and the preprocessed image sequence is output.

[0009] Optionally, the improved BEVFormer perception module includes: An improved BEVFormer model is constructed, which includes corridor anchor point query, safety boundary driven sampling band and temporal occupancy prediction attention structure. Camera features are extracted from the preprocessed image sequence to obtain the image feature sequence arranged by camera channel, and a set of bird’s-eye grid query vectors is established in the bird’s-eye coordinate system. Based on the corridor anchor point query, a set of corridor query vectors for candidate travel corridors is generated. The candidate travel corridors are represented in a parametric way and are determined by the parameters of corridor centerline, corridor width, corridor orientation and corridor curvature. The set of corridor query vectors and the set of bird’s-eye view query are used together as the query input for bird’s-eye view spatial encoding and corridor-level feature aggregation is performed. Cross-view sampling positions are determined by using a safety boundary-driven sampling band. The safety boundary-driven sampling band is generated by the safety margin and distributed along the robot's safety boundary contour and forward fan-shaped region in the bird's-eye view space. The cross-view sampling positions are projected onto the pixel coordinate system of the image feature sequence of each camera to obtain the set of sampling points for each camera. Key vectors and value vectors are extracted from the image feature sequence using the sampling point set of each camera. Spatial cross-attention fusion is performed using the bird’s-eye view query set and the corridor query vector set to obtain the bird’s-eye view feature map at the current moment. Temporal occupancy prediction attention is used to perform temporal fusion of the current bird's-eye view feature map and the historical bird's-eye view feature map. Based on the temporal fusion result, occupancy prediction output is generated, and the output consists of a bird's-eye view occupancy raster sequence and a bird's-eye view risk raster sequence arranged by time step.

[0010] Optionally, the security expansion processing module includes: Read the current-time grid cells from the bird's-eye view occupied grid sequence and risk grid sequence, convert the occupied status and risk level of each cell in the grid into passable and impassable markers, and generate a grid set of impassable areas; The expansion radius is determined based on the robot's safety margin. The expansion radius is equal to the sum of the robot's outer radius and the minimum safe distance, plus the distance margin corresponding to the pose uncertainty. In the bird's-eye view grid coordinate system, the grid set of the impassable area is subjected to morphological expansion processing to obtain the expanded impassable area. A locally traversable region is constructed using grid cells within the bird's-eye view grid coverage area, excluding the expanded impassable areas. The grid representation of the locally traversable region and its corresponding spatial boundary information are then output.

[0011] Optionally, the polar coordinate mapping module includes: A polar coordinate grid is established with the bird's-eye view coordinates corresponding to the robot's current pose as the origin of the polar coordinate grid. The polar coordinate grid consists of an angle dimension and a distance dimension. The angle dimension is divided into angle sectors, and the distance dimension is divided into distance loops. The bird's-eye view grid cells in the local traversable domain are converted into polar coordinate cells. For each bird's-eye view grid cell, the distance and azimuth values ​​from the polar coordinate origin to the center point of the bird's-eye view grid cell are calculated. The traversable markers of the bird's-eye view grid cells are mapped to polar coordinate cells containing the current distance and azimuth values ​​to generate a polar coordinate traversable domain representation. The target direction is calculated based on the coordinates of the target point in the bird's-eye view coordinate system and the coordinates of the robot's current pose in the bird's-eye view coordinate system. The target direction is the azimuth angle from the robot's current position to the target point. The target direction is mapped to the angle dimension of the polar coordinate grid to determine the target angle sector where the target direction is located.

[0012] Optionally, the vector field histogram planning module includes: Read the polar coordinate passable domain representation and scan the polar coordinate passable domain representation along the angular dimension. Determine the minimum impassable distance from the polar coordinate origin in each angular sector. Convert the angular range corresponding to the minimum impassable distance into a blocking angular interval and generate a set of blocking angular intervals. An angle interval graph is used to perform interval merging and interval adjacency establishment on the set of blocked angle intervals. The set of blocked angle intervals is deduplicated and merged in the angle dimension to form the connection relationship of blocked intervals. The set of passable angle intervals and the connection relationship of passable angle intervals are generated on the complement of the connection relationship of blocked intervals. A polar coordinate distance field is constructed within the polar coordinate traversable domain defined by the set of traversable angle intervals. The distance field value of each polar coordinate unit in the polar coordinate distance field is defined as the shortest distance from the polar coordinate unit to the boundary of the impassable region. The shortest distance is the minimum value of the Euclidean distance from the center point of the polar coordinate unit to any point on the boundary of the impassable region. The channel center ridge sequence is determined by extracting the polar coordinate distance field ridge. The channel center ridge sequence consists of polar coordinate units that satisfy the local maximum condition. The local maximum condition is that the distance field value of the polar coordinate unit is greater than the distance field value of the adjacent polar coordinate unit in the angular dimension, and is also greater than the distance field value of the adjacent polar coordinate unit in the distance dimension. The channel center ridge sequence is mapped to the angular dimension to obtain the channel center direction sequence corresponding to the channel center ridge sequence.

[0013] Optionally, the route generation control module includes: Receive the target direction, the set of passable angle intervals, the connectivity of the passable angle intervals, and the channel center ridge sequence. Determine the target candidate interval containing the target direction from the set of passable angle intervals. When the target direction does not fall into any passable angle interval, determine the candidate interval with the smallest angular deviation from the target direction based on the connectivity of the passable angle intervals. A local travel route is generated based on the channel center ridge sequence. The local travel route is composed of the point sequence of the channel center ridge sequence in the bird's-eye view coordinate system. The continuous path curve is obtained by curve fitting of the point sequence. The curvature of the continuous path curve is determined by the rate of change of the tangent direction of the path curve at the corresponding position. The control commands are calculated based on the local travel path. The control commands include linear velocity and angular velocity. The linear velocity is the smaller of the preset maximum linear velocity and the upper limit of the safe linear velocity calculated based on the nearest obstacle distance on the local travel path. The angular velocity is the product of the linear velocity and the curvature of the local travel path at the robot's current position.

[0014] Optionally, the validity verification execution module includes: Receive the bird's-eye view occupied grid sequence, risk grid sequence and local travel route, discretize the local travel route into a path sampling point sequence in the bird's-eye view coordinate system, and configure the path sampling point sequence with a time index that increases according to the control period to cover the current time to the future time range; For each path sampling point in the path sampling point sequence, based on the time index, select the corresponding time step of the occupancy grid and risk grid from the bird's-eye view occupancy grid sequence and the risk grid sequence, determine the occupancy status of the path sampling point in the corresponding occupancy grid, read the risk value of the path sampling point in the corresponding risk grid, and generate the path occupancy determination sequence and the path risk value sequence. When there are no impassable markers in the path occupancy determination sequence and the maximum risk value in the path risk value sequence is less than the preset risk threshold, the local travel route is determined to be valid, and control commands are output to drive the robot to move. When the conditions are not met, the local travel route is determined to be invalid, and preset safety control commands are output to regenerate the local travel route.

[0015] The beneficial effects of this invention are: This invention improves the structure, spatial consistency, and temporal continuity of environmental representation by preprocessing image acquisition, improving BEVFormer bird's-eye view coding, and applying safety dilation. It stably maps camera visual information from pixel-level observations to bird's-eye view occupancy and risk grid sequences, further forming local traversable regions. Compared to existing schemes that primarily rely on detection boxes, semantic segmentation results, or single-frame occupancy inference to reconstruct cost maps, this invention uses corridor anchor point queries to achieve feature aggregation oriented towards the travel corridor, making the environmental representation more aligned with planning requirements. It employs safety boundary-driven sampling bands to enhance cross-view sampling consistency in vehicle safety-related areas, reducing representation fluctuations caused by lighting changes, occlusion, and false detections. Furthermore, it uses temporal occupancy prediction attention to output occupancy and risk information for multiple future time steps, mitigating the impact of single-frame noise and short-term missing data on planning. This results in a more stable traversable region, significantly reducing route jitter, detour errors, and planning infeasibility in obstacle-heavy or narrow passages.

[0016] This invention utilizes polar coordinate mapping and vector field histogram planning. In the local planning stage, it constructs a set of passable angle intervals and their connectivity using angle interval graphs, avoiding directional jumps and valley value misjudgments caused by traditional histogram threshold segmentation. By extracting channel center ridge sequences through polar coordinate distance field ridge extraction, the local travel route becomes more centered and continuous, improving the smoothness and passability of the obstacle avoidance process. In terms of control loop, this invention combines route generation control with validity verification execution. It uses bird's-eye view occupied grid sequences and risk grid sequences to perform real-time validity determination of local travel routes. When a route is deemed invalid, a safety control command is output and an update recalculation is triggered. This ensures rapid response and a safety safety net even when encountering dynamic obstacles, experiencing increased perception uncertainty, or undergoing rapid environmental changes. Compared to existing weakly coupled perception, planning, and control schemes, this invention reduces reaction lag and collision risk, improving stability, robustness, and engineering usability in dynamic scenarios. Attached Figure Description

[0017] 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: Figure 1 This is a flowchart of a robot path planning and obstacle avoidance system based on AI visual detection proposed in this invention; Figure 2 This is a structural block diagram of the improved BEVFormer model of a robot path planning and obstacle avoidance system based on AI vision detection proposed in this invention. Figure 3 This is a functional schematic diagram of the vector field histogram method for a robot path planning and obstacle avoidance system based on AI visual detection proposed in this invention. Detailed Implementation

[0018] 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.

[0019] refer to Figure 1 , Figure 2 and Figure 3 A robot path planning and obstacle avoidance system based on AI visual detection includes: The image acquisition and preprocessing module acquires camera image data from the robot, preprocesses the camera image data, and obtains a preprocessed image sequence. Improve the BEVFormer perception module, construct an improved BEVFormer model, perform feature extraction and bird's-eye view spatial encoding on the preprocessed image sequence, perform corridor-level feature aggregation based on corridor anchor point query, introduce safety boundary-driven sampling band to determine cross-view sampling position, use temporal occupancy prediction attention to perform temporal fusion and prediction, and output bird's-eye view occupancy grid sequence and risk grid sequence. The safety expansion processing module expands the impassable area based on the robot's safety margin to obtain a locally passable area; The polar coordinate mapping module maps the local traversable region to a polar coordinate grid with the robot as the origin, constructs a polar coordinate traversable region representation, and determines the target direction. The vector field histogram planning module, based on the vector field histogram method, uses an angle interval map to construct a set of passable angle intervals and interval connectivity by representing the passable domain in polar coordinates. It also constructs a polar coordinate distance field by extracting the ridge line of the polar coordinate distance field, thus obtaining the set of passable angle intervals and the sequence of channel center ridge lines. The route generation and control module generates a local travel route and calculates control commands based on the target direction, the set of passable angle intervals, and the sequence of the central ridge line of the channel. The validity verification execution module performs real-time validity determination of the local travel route based on the bird's-eye view occupied grid sequence and risk grid sequence, and drives the robot to move through control commands.

[0020] In this embodiment, the camera image data includes the original camera image frames, corresponding timestamp information, camera intrinsic data, camera extrinsic data, and camera image acquisition parameters.

[0021] In this embodiment, obtaining the preprocessed image sequence includes: The system acquires raw image frames from the camera, reads the timestamp information corresponding to each frame, performs frame-level synchronization and pairing of raw image frames from different cameras according to their timestamps, generates a multi-camera synchronized image frame group arranged in chronological order, and removes image frames whose timestamp deviation exceeds the synchronization threshold. Specifically, the removal of image frames with timestamp deviation exceeding the synchronization threshold involves: For each moment, the timestamp of the main camera image frame is used as the reference timestamp. The frame with the smallest absolute difference from the reference timestamp is selected from the remaining camera image frames as the pairing frame. The time deviation between each pairing frame and the reference timestamp is calculated. The time deviation is equal to the absolute difference between the pairing frame timestamp and the reference timestamp. The maximum value among all camera time deviations is recorded as the maximum time deviation of the synchronized image frame group. When the maximum time deviation is greater than the synchronization threshold, the synchronized image frame group is determined to be invalid, and camera image frames with time deviations exceeding the synchronization threshold are removed. At the same time, the synchronized image frame group is discarded, and pairing is re-executed with the next frame of the main camera. The synchronization threshold is set to 33 milliseconds. Distortion correction and pixel coordinate remapping are performed on each original camera image frame in the camera-synchronized image frame group. The image is then cropped and scaled according to the input resolution, and its color space is standardized to obtain a corrected image frame with a unified color channel representation. The distortion correction and pixel coordinate remapping process is performed as follows: For each original image frame from the camera, the camera's intrinsic parameter matrix and distortion coefficients are read. The distortion coefficients include radial distortion coefficients and tangential distortion coefficients. The pixel coordinates of the corrected image are used as the target coordinates. The target coordinates are normalized to obtain normalized coordinates. The normalized coordinates are substituted into the radial distortion coefficients and tangential distortion coefficients to calculate the distortion offset. The distortion offset is superimposed on the normalized coordinates to obtain the distortion coordinates. The distortion coordinates are mapped back to the original image coordinate system through the intrinsic parameter matrix to obtain the source pixel coordinates. Bilinear interpolation is used to sample the pixel values ​​of the target coordinates at the source pixel coordinates. The corrected image frame is generated by traversing all target coordinates. The image undergoes color space uniformity processing, specifically as follows: the current color channel format of the calibration image frame is read and uniformly converted to a preset color space, which is set to a red, green and blue three-channel color space. When the calibration image frame is in blue-green-red channel order, the channels are rearranged to obtain the red-green-blue channel order. When the calibration image frame is a single grayscale channel, the red, green and blue three channels are generated by copying the grayscale value. When the calibration image frame is a four-channel format with an alpha channel, the alpha channel is discarded and the red, green and blue three channels are retained. The converted three-channel pixel values ​​are cropped according to the value range of 0 to 255 for each channel. The cropping rule is to set values ​​less than 0 to 0 and values ​​greater than 255 to 255. Pixel value normalization is performed on the corrected image frames, data format conversion is completed, the image data is converted into a preset tensor format, the channel order is arranged, and each corrected image frame is associated with corresponding camera intrinsic and extrinsic data. The preprocessed image sequence is then output. Specifically, the conversion of the image data into the preset tensor format involves: The three-channel pixel matrix of each corrected image frame is converted to floating-point type. The pixel value is divided by 255 to obtain the normalized pixel value from zero to one. The data is rearranged from the storage order of height multiplied by width multiplied by three channels to the storage order of three channels multiplied by height multiplied by width to form a three-dimensional tensor. For multi-camera data, each three-dimensional tensor is stacked along the newly added dimension according to the camera number to obtain a four-dimensional tensor of camera number multiplied by three channels multiplied by height multiplied by width as the preset tensor format. The normalized pixel value is truncated. The truncation threshold is set to replace values ​​outside of zero to one with zero or one. The camera number, intrinsic parameter data and extrinsic parameter data index information are added to each tensor. The preprocessed image sequence is output in chronological order.

[0022] In this embodiment, the improved BEVFormer perception module includes: An improved BEVFormer model is constructed, which includes corridor anchor point query, safety boundary-driven sampling band, and temporal occupancy prediction attention structure. Camera features are extracted from the preprocessed image sequence to obtain an image feature sequence arranged by camera channel. A set of bird's-eye view grid query vectors is established in the bird's-eye view coordinate system, wherein: An improved BEVFormer model is constructed as follows: In the basic BEVFormer encoder framework, a bird's-eye view query end and a cross-view attention end are set up. Corridor anchor point query, safety boundary driven sampling band, and temporal occupancy prediction attention are embedded in parallel as three types of structures. Corridor anchor point query adds a candidate corridor query vector branch to the bird's-eye view query end. The candidate corridor query vector is initialized by the target direction and the candidate corridor parameters generated by the control command of the previous cycle. The query vector is obtained through a learnable mapping. The safety boundary driven sampling band replaces the sampling point generation method in the cross-view attention end. The sampling points are generated by the boundary band formed by the safety margin and the forward fan-shaped region and are used to determine the sampling position of each camera. The temporal occupancy prediction attention adds a temporal fusion attention layer to the bird's-eye view encoding end. The output is the bird's-eye view feature at the current moment and the predicted bird's-eye view occupancy feature for the future. Camera feature extraction is performed on the preprocessed image sequence. Specifically, feature extraction is performed on the preprocessed image tensor of each camera at each time step. The feature extraction consists of a convolutional backbone and a feature pyramid, outputting a two-dimensional feature map with a fixed number of channels. The feature maps of each camera at the same time step are arranged according to the camera number to form a multi-camera feature group. The image feature sequence is formed in chronological order. The feature map is numerically truncated with a truncation threshold set to -10 to +10. Values ​​less than -10 are set to -10, and values ​​greater than +10 are set to +10. A set of bird's-eye view grid query vectors is established in the bird's-eye view coordinate system. Specifically, a bird's-eye view coordinate system is established with the geometric center of the robot chassis as the origin. The bird's-eye view coordinate system is a two-dimensional plane coordinate system. The forward direction is defined as the positive longitudinal direction, and the left direction is defined as the positive lateral direction. The coordinate unit is meters. The current plane is regarded as the ground projection plane. In the bird's-eye view coordinate system, the bird's-eye view coverage range is set to 20 meters forward, 5 meters backward, 10 meters left and 10 meters right. The grid resolution is set to 0.2 meters. The bird's-eye view plane is discretized into regular grids according to the coverage range and resolution. A learnable query vector is assigned to the center point of each grid cell and a position code is superimposed. The position code is obtained by linear mapping of the lateral and longitudinal coordinates of the grid center point and then spliced ​​together. The query vectors of all grid cells are flattened in row and column order to form a set of bird's-eye view grid query vectors. The query vector dimension is set to 256. Based on corridor anchor point queries, a set of corridor query vectors is generated for candidate travel corridors. These candidate corridors are represented parametrically, determined by parameters such as corridor centerline, width, orientation, and curvature. The set of corridor query vectors and the bird's-eye view query set are used together as input for bird's-eye view spatial encoding, and corridor-level feature aggregation is performed. Specifically: The process of generating a set of candidate travel corridor query vectors is as follows: nine candidate travel corridors are generated in the bird's-eye view coordinate system starting from the robot's current pose. The corridor orientation is taken from the target direction at 40-degree intervals on the left and right. The corridor centerline extends forward by eight meters according to the corresponding orientation. The corridor width is set to the robot's external width plus twice the minimum safety distance. The corridor curvature parameter is the difference between the corridor orientation and the current heading divided by eight meters. The corridor centerline endpoint coordinates, corridor width, corridor orientation, and corridor curvature parameter are concatenated into a corridor parameter vector. This vector is then projected through linear mapping to form a corridor query vector of dimension 256, thus forming a set of corridor query vectors. The corridor-level feature aggregation is performed as follows: the corridor query vector and the bird's-eye view grid query vector are input into the bird's-eye view spatial encoder. Spatial cross attention is used, with the corridor query vector as the query vector and the multi-camera feature vectors extracted at cross-view sampling points as the key vector and value vector. The dot product similarity between the query vector and each key vector is calculated for each corridor query vector. Normalization is performed to obtain the attention weight. The corresponding value vectors are weighted and summed according to the attention weight to obtain the corridor aggregated feature vector, forming the corridor-level feature aggregation result. Cross-view sampling positions are determined using a safety boundary-driven sampling band. This band is generated by a safety margin and distributed along the robot's safety boundary contour and the forward fan-shaped region in the bird's-eye view. The cross-view sampling positions are projected onto the pixel coordinate system of each camera's image feature sequence to obtain the set of sampling points for each camera. The safety margin is defined as the equivalent outward extension distance for generating the safety boundary. The safety margin is equal to the sum of the robot's outer radius and the minimum safety distance, plus the distance margin corresponding to the pose uncertainty. The robot's outer radius is half the length of the diagonal of the robot's outer envelope. The minimum safety distance is set to 0.3 meters. The distance margin corresponding to the pose uncertainty is the larger of the lateral standard deviation and the longitudinal standard deviation of the positioning, multiplied by three. The robot's safety boundary contour and forward fan-shaped region are distributed in the bird's-eye view space. Specifically, the robot's outer contour is expanded outward with the robot's geometric center as the center and the safety margin as the outward expansion distance to obtain the safety boundary contour. The contour is then expanded outward by 0.2 meters to form a contour band. In the bird's-eye view coordinate system, a forward fan-shaped region is constructed with the robot's forward direction as the center direction of the fan. The fan angle range is set to 60 degrees on both the left and right sides, and the maximum radius of the fan is set to 10 meters. Within the contour band, the first sampling position set is generated by sampling along the contour at 0.2-meter intervals with equal arc length. Within the forward fan-shaped region, the second sampling position set is generated at angular intervals of 5 degrees and radial intervals of 0.5 meters. The first sampling position set and the second sampling position set are merged to form a cross-view sampling position set. Key vectors and value vectors are extracted from the image feature sequence using the sampling point sets of each camera. Spatial cross-attention fusion is then performed using the bird's-eye view query set and the corridor query vector set to obtain the bird's-eye view feature map at the current time, where: Key vectors and value vectors are extracted from the image feature sequence. Specifically, for each sampling point in the camera sampling point set, the pixel coordinates of the sampling points are converted into the corresponding camera feature map coordinates. The conversion rule is to divide the horizontal pixel coordinates of the sampling points by the feature map downsampling factor to obtain the horizontal feature coordinates, and divide the vertical pixel coordinates of the sampling points by the feature map downsampling factor to obtain the vertical feature coordinates. The downsampling factor is set to eight. Bilinear interpolation is used to sample the camera feature map at the horizontal and vertical feature coordinates to obtain the sampled feature vectors. The sampled feature vectors are then processed by the first linear mapping to obtain the key vectors and by the second linear mapping to obtain the value vectors. The first linear mapping consists of a single-layer fully connected transformation, and the second linear mapping consists of a series of two-layer fully connected transformations. The spatial cross-attention fusion is performed as follows: the bird's-eye view query set and the corridor query vector set are concatenated to form a query vector set. The similarity between each query vector in the query vector set and the key vector set of the corresponding camera is calculated. The similarity is defined as the dot product of the query vector and the key vector divided by the square root of 256. The similarity of each query vector within the same camera is normalized to obtain the attention weight. The attention weights of multiple cameras are concatenated according to the camera dimension and normalized again. The corresponding value vectors are weighted and summed according to the attention weight to obtain the fusion vector. A temporal occupancy prediction attention method is used to perform temporal fusion of the current-time bird's-eye view feature map and the historical-time bird's-eye view feature maps. Based on the temporal fusion result, an occupancy prediction output is generated, which outputs a bird's-eye view occupancy raster sequence and a bird's-eye view risk raster sequence arranged by time step. Specifically, the occupancy prediction output generated based on the temporal fusion result is as follows: Using the temporally fused bird's-eye view feature map as the prediction input, a current-time prediction branch and a future prediction branch are generated for each grid cell on the bird's-eye view grid. The future prediction branch contains five time steps. For each time step, the corresponding grid value is output through the occupancy prediction head and the risk prediction head. The occupancy prediction head outputs the probability that each grid cell is impassable. The calculation method is to input the grid cell feature vector into two layers of convolution to obtain a one-dimensional occupancy log value, which is then logically converted into a probability of zero to one. The risk prediction head outputs the risk value of each grid cell. The grid cell feature vector is input into two layers of convolutional network to obtain a risk log value, which is then linearly scaled and mapped to a risk value of zero to one. Grid cells with an occupancy probability greater than 0.5 are marked as impassable, and the rest are marked as passable. The occupancy grid and risk grid of each time step are arranged in chronological order to form the bird's-eye view occupancy grid sequence and the bird's-eye view risk grid sequence.

[0023] In this embodiment, the security expansion processing module includes: Read the current-time grid cells from the bird's-eye view occupied grid sequence and risk grid sequence, convert the occupied status and risk level of each cell in the grid into passable and impassable markers, and generate a grid set of impassable areas; The expansion radius is determined based on the robot's safety margin. The expansion radius is equal to the sum of the robot's outer radius and the minimum safe distance, plus the distance margin corresponding to the pose uncertainty. In the bird's-eye view grid coordinate system, the grid set of the impassable area is subjected to morphological expansion processing to obtain the expanded impassable area. A locally traversable region is constructed using grid cells within the bird's-eye view grid coverage area, excluding the expanded impassable areas. The grid representation of the locally traversable region and its corresponding spatial boundary information are then output.

[0024] In this embodiment, the polar coordinate mapping module includes: A polar coordinate grid is established with the bird's-eye view coordinates corresponding to the robot's current pose as the origin of the polar coordinate grid. The polar coordinate grid consists of an angle dimension and a distance dimension. The angle dimension is divided into angle sectors, and the distance dimension is divided into distance loops. The bird's-eye view grid cells in the local traversable domain are converted into polar coordinate cells. For each bird's-eye view grid cell, the distance and azimuth values ​​from the polar coordinate origin to the center point of the bird's-eye view grid cell are calculated. The traversable markers of the bird's-eye view grid cells are mapped to polar coordinate cells containing the current distance and azimuth values ​​to generate a polar coordinate traversable domain representation. The target direction is calculated based on the coordinates of the target point in the bird's-eye view coordinate system and the coordinates of the robot's current pose in the bird's-eye view coordinate system. The target direction is the azimuth angle from the robot's current position to the target point. The target direction is mapped to the angle dimension of the polar coordinate grid to determine the target angle sector where the target direction is located.

[0025] In this embodiment, the vector field histogram planning module includes: Read the polar coordinate traversable domain representation and scan it along the angular dimension. Within each angular sector, determine the minimum impassable distance from the polar coordinate origin outwards. Convert the angular range corresponding to the minimum impassable distance into blocking angular intervals, generating a set of blocking angular intervals. Specifically, the conversion of the angular range corresponding to the minimum impassable distance into blocking angular intervals is as follows: Within each angular sector, traverse the polar coordinate cells from near to far along the distance dimension, find the first polar coordinate cell marked as impassable, and record the distance from the cell center to the origin of the polar coordinates as the minimum impassable distance. Using the central angle of the angular sector as the reference angle, define the robot's equivalent half-width as half the robot's external width plus the minimum safe distance and set it to 0.6 meters. When the minimum impassable distance is less than 0.2 meters, the entire angular sector is determined to be a blocked interval; otherwise, calculate the angle expansion amount. The angle expansion amount is equal to the input of the arcsine function, which is the robot's equivalent half-width divided by the minimum impassable distance. The left boundary angle of the blocked angular interval is equal to the reference angle minus the angle expansion amount, and the right boundary angle is equal to the reference angle plus the angle expansion amount. Merge the blocked angular intervals obtained from adjacent angular sectors according to the angular overlap relationship to form a set of blocked angular intervals. An angle interval graph is used to perform interval merging and interval adjacency establishment on the set of blocked angle intervals. The set of blocked angle intervals is deduplicated and merged in the angle dimension to form the connection relationship of blocked intervals. The set of passable angle intervals and the connection relationship of passable angle intervals are generated on the complement of the connection relationship of blocked intervals. A polar coordinate distance field is constructed within the polar coordinate traversable domain defined by the set of traversable angle intervals. The distance field value of each polar coordinate cell in the polar coordinate distance field is defined as the shortest distance from the polar coordinate cell to the boundary of the impassable region. The shortest distance is the minimum value of the Euclidean distances from the center point of the polar coordinate cell to any point on the boundary of the impassable region. Specifically, the construction of the polar coordinate distance field within the polar coordinate traversable domain defined by the set of traversable angle intervals is as follows: Within the angular range corresponding to the set of traversable angular intervals, retain traversable polar coordinate units in the polar coordinate traversable domain representation, and mark all polar coordinate units within the remaining angular range as traversable. Extract the traversable region boundary within the retained region. The traversable region boundary is defined as the set of center points of traversable polar coordinate units when any traversable polar coordinate unit is adjacent to a traversable polar coordinate unit in the angular dimension or the distance dimension. For each traversable polar coordinate unit, calculate the Euclidean distance from the center point of the traversable polar coordinate unit to the set of boundary center points. Take the minimum value among all Euclidean distances as the distance field value of the traversable polar coordinate unit, and set the distance field value of the traversable polar coordinate units to zero. To reduce the computational load, only search for the boundary center point within the five-by-five neighborhood centered on the traversable polar coordinate unit. If there is no boundary center point in the neighborhood, expand the search neighborhood to nine-by-nine. The channel center ridge sequence is determined by extracting the ridges from the polar coordinate range field. This sequence consists of polar coordinate units that satisfy a local maximum condition: the range field value of each polar coordinate unit is greater than the range field value of its adjacent polar coordinate unit in the angular dimension. Mapping the channel center ridge sequence to the angular dimension yields the corresponding channel center direction sequence, where: The channel center ridge sequence is determined by extracting ridges from the polar coordinate distance field. Specifically, the polar coordinate distance field is traversed only for polar coordinate units with distance field values ​​greater than zero. For each polar coordinate unit, the distance field values ​​of the preceding and following angle units in the angular dimension, as well as the inner and outer distance units in the distance dimension, are read. When the distance field value of a polar coordinate unit is strictly greater than the distance field values ​​of four adjacent units, the polar coordinate unit is marked as a ridge candidate unit. The ridge candidate units are sorted from near to far in the distance dimension. Adjacent ridge candidate units are connected according to the adjacency relationship in the angular dimension to form ridge segments. The connection rule is that the difference between the angle indexes of adjacent ridge candidate units does not exceed one and the difference between the distance indexes does not exceed one. Ridge segments with a length of less than five polar coordinate units are removed. The remaining ridge segments are arranged in order to form the channel center ridge sequence. The channel center ridge sequence is mapped to the angle dimension to obtain the channel center direction sequence corresponding to the channel center ridge sequence. Specifically, for each polar coordinate unit in the channel center ridge sequence, the angle sector index and the distance ring index are read. The center angle of the angle sector corresponding to the angle sector index is taken as the channel center direction angle of the polar coordinate unit. When multiple ridge polar coordinate units appear in the same angle sector, the channel center direction angle corresponding to the polar coordinate unit with the smallest distance ring index is selected as the channel center direction angle of the angle sector. The channel center direction angles arranged from near to far in the distance dimension constitute the channel center direction sequence. The direction angles with a difference of more than 20 degrees between adjacent direction angles in the direction angle sequence are regarded as breakpoints. The sequence is truncated into multiple direction subsequences at the breakpoints. The breakpoint threshold is set to 20 degrees.

[0026] In this embodiment, the route generation control module includes: Receive the target direction, the set of passable angle intervals, the connectivity of the passable angle intervals, and the channel center ridge sequence. Determine the target candidate interval containing the target direction from the set of passable angle intervals. When the target direction does not fall into any passable angle interval, determine the candidate interval with the smallest angular deviation from the target direction based on the connectivity of the passable angle intervals. A local travel route is generated based on the channel center ridge sequence. The local travel route is composed of the point sequence of the channel center ridge sequence in the bird's-eye view coordinate system. The continuous path curve is obtained by curve fitting of the point sequence. The curvature of the continuous path curve is determined by the rate of change of the tangent direction of the path curve at the corresponding position. The control commands are calculated based on the local travel path. These commands include linear velocity and angular velocity. The linear velocity is the smaller of a preset maximum linear velocity and a safe upper limit of linear velocity calculated based on the nearest obstacle distance on the local travel path. The angular velocity is the product of the linear velocity and the curvature of the local travel path at the robot's current position. Specifically: Starting from the robot's current position, samples of equal arc length are taken from the bird's-eye view grid along the local travel path. The sampling interval is set to 0.2 meters. For each sampling point, the Euclidean distance to the boundary of the impassable area in the bird's-eye view traversable domain is calculated, and the minimum value is taken as the nearest obstacle distance. The safe linear velocity upper limit is calculated based on the nearest obstacle distance. The calculation method is that the safe linear velocity upper limit is equal to the square root of the product of twice the maximum deceleration and the nearest obstacle distance. The maximum deceleration is set to 2.0 meters per second squared. The preset maximum linear velocity is set to 1.2 meters per second. The linear velocity is taken as the smaller value between the preset maximum linear velocity and the safe linear velocity upper limit. When the nearest obstacle distance is less than 0.3 meters, the linear velocity is directly set to zero.

[0027] In this embodiment, the validity verification execution module includes: Receive the bird's-eye view occupied grid sequence, risk grid sequence and local travel route, discretize the local travel route into a path sampling point sequence in the bird's-eye view coordinate system, and configure the path sampling point sequence with a time index that increases according to the control period to cover the current time to the future time range; For each path sampling point in the path sampling point sequence, based on the time index, select the corresponding time step of the occupancy grid and risk grid from the bird's-eye view occupancy grid sequence and the risk grid sequence, determine the occupancy status of the path sampling point in the corresponding occupancy grid, read the risk value of the path sampling point in the corresponding risk grid, and generate the path occupancy determination sequence and the path risk value sequence. When there are no impassable markers in the path occupancy determination sequence and the maximum risk value in the path risk value sequence is less than the preset risk threshold, the local travel route is determined to be valid, and a control command is output to drive the robot to move. When the conditions are not met, the local travel route is determined to be invalid, and a preset safety control command is output to regenerate the local travel route, where the preset risk threshold is set to 0.65.

[0028] Example 1: To verify the feasibility of the present invention in practice, it was applied to an autonomous navigation task for a mobile robot in a mixed indoor and outdoor environment. The environment included narrow passages, glass and reflective walls, low-textured floors, shelf corners and temporary stacks of goods, and dynamic obstacles such as pedestrians, handcarts and forklifts. In the current scenario, the traditional solution of directly converting visual detection results into grids and then planning is prone to discontinuity of the passable domain due to false detections, missed detections and detection box jitter, resulting in swaying of the route, detours along the edges, repeated attempts at corners, and even partial jamming. When encountering dynamic obstacles, planning updates are often delayed and obstacle avoidance actions are not smooth, leading to frequent sudden stops or an increased risk of close encounters.

[0029] After acquiring images through multiple cameras and completing synchronization and preprocessing, the robot enters the improved BEVFormer perception module, which outputs the current and five-step bird's-eye view occupancy grid sequence and risk grid sequence. Based on the safety margin, the inaccessible area is expanded to obtain the local accessible domain, which is then mapped to a polar coordinate grid to form a polar coordinate accessible domain representation and target direction. In the local planning stage, the vector field histogram method is used to construct the set of accessible angle intervals and the interval connectivity relationship from the angle interval map. The channel center ridge sequence is extracted from the polar coordinate distance field ridge line to generate a centered and continuous local travel route and calculate the linear velocity and angular velocity control commands. Before and after execution, the effectiveness of the local travel route is determined by the bird's-eye view occupancy and risk sequences. When an inaccessible marker or risk boundary violation occurs in the future time domain, a safety control command is output and an update recalculation is triggered, thereby maintaining a real-time closed loop of planning, obstacle avoidance and control when dynamic obstacle entry and perception uncertainty increase.

[0030] To verify the beneficial effects, the same robot platform was continuously run in similar mixed scenarios, fully recording task logs, camera input and preprocessing results, bird's-eye view occupancy and risk output, local navigable regions, polar coordinate navigable region representation, channel center ridge sequence, local travel routes, control commands, and execution feedback information. Simultaneously, safety-related data such as minimum distance to obstacles, emergency stop triggering, route recalculation triggering, and validity judgment results were recorded. The system operated stably under a fixed control cycle, with local planning continuously updated. When encountering dynamic obstacles crossing, passing, or sudden incursions, it could quickly adjust the navigable angle range and reconstruct the channel center ridge, ensuring a continuous output route without significant directional jumps. In typical challenging areas such as narrow channels, shelf corners, and reflective walls, the local travel route stably adhered to the channel center, with significantly reduced lateral swaying. The robot stopped less frequently when passing corners and passing areas, resulting in smoother overall movement. In segments where occlusion, reflection, or missing texture increases perceptual uncertainty, validity determination can promptly trigger safety controls and regenerate local travel routes, avoiding misjudging impassable areas as passable and resulting in driving close to obstacles. It also reduces repeated attempts and local jamming, improving safety, robustness, and engineering usability in complex dynamic scenarios.

[0031] Table 1 Comparison of navigation and obstacle avoidance performance in mixed dynamic scenarios

[0032] As shown in Table 1, in terms of path stability and feasibility indicators, this invention achieves the best results in both root mean square path jitter and planning failure rate. The jitter is 6.8 cm, significantly lower than the 18.9 cm of the vector field histogram method, the 15.7 cm of the dynamic window method, and the 12.4 cm of the artificial potential field method. The planning failure rate is 1.1%, also better than the comparison methods. This indicates that the structured representation of bird's-eye view occupancy and risk, as well as the channel center ridge generation method, can reduce lateral swaying of the route and the generation of infeasible paths, thereby improving stable passage capability in complex environments.

[0033] In terms of dynamic scene safety and response capability, the success rate of dynamic obstacle avoidance of the present invention is 96.5%, which is higher than 84.2% of the vector field histogram method and 87.6% of the dynamic window method. At the same time, the first response latency is 96ms, which is faster than all the comparison methods. The near miss event rate of the present invention is 0.21 times / hour, which is the lowest among all methods. This shows that in dynamic obstacle entry and rapidly changing scenes, the present invention can complete route adjustment more timely and maintain a safety margin.

[0034] In terms of efficiency and stop-and-go behavior, the average passage speed of this invention is 0.86 m / s, the highest among all methods. Meanwhile, the emergency stop trigger rate is 0.38 times / km, lower than the vector field histogram method's 1.42 times / km, the dynamic window method's 1.15 times / km, and the artificial potential field method's 0.92 times / km. Compared with traditional BEV perception and end-to-end navigation, this invention maintains a higher speed while still preserving lower risks of emergency stops and near misses, demonstrating better passage efficiency and obstacle avoidance smoothness.

[0035] 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. A robot path planning and obstacle avoidance system based on AI visual detection, characterized in that, include: The image acquisition and preprocessing module acquires camera image data from the robot, preprocesses the camera image data, and obtains a preprocessed image sequence. Improve the BEVFormer perception module, construct an improved BEVFormer model, perform feature extraction and bird's-eye view spatial encoding on the preprocessed image sequence, perform corridor-level feature aggregation based on corridor anchor point query, introduce safety boundary-driven sampling band to determine cross-view sampling position, use temporal occupancy prediction attention to perform temporal fusion and prediction, and output bird's-eye view occupancy grid sequence and risk grid sequence. The safety expansion processing module expands the impassable area based on the robot's safety margin to obtain a locally passable area; The polar coordinate mapping module maps the local traversable region to a polar coordinate grid with the robot as the origin, constructs a polar coordinate traversable region representation, and determines the target direction. The vector field histogram planning module, based on the vector field histogram method, uses an angle interval map to construct a set of passable angle intervals and interval connectivity by representing the passable domain in polar coordinates. It also constructs a polar coordinate distance field by extracting the ridge line of the polar coordinate distance field, thus obtaining the set of passable angle intervals and the sequence of channel center ridge lines. The route generation and control module generates a local travel route and calculates control commands based on the target direction, the set of passable angle intervals, and the sequence of the central ridge line of the channel. The validity verification execution module performs real-time validity determination of the local travel route based on the bird's-eye view occupied grid sequence and risk grid sequence, and drives the robot to move through control commands.

2. The robot path planning and obstacle avoidance system based on AI visual detection according to claim 1, characterized in that, The camera image data includes the original image frames from the camera, corresponding timestamp information, camera intrinsic data, camera extrinsic data, and camera image acquisition parameters.

3. The robot path planning and obstacle avoidance system based on AI visual detection according to claim 1, characterized in that, The obtained preprocessed image sequence includes: The system acquires raw image frames from the camera, reads the timestamp information corresponding to each frame, performs frame-level synchronization and pairing of raw image frames from different cameras according to their timestamps, generates a multi-camera synchronized image frame group arranged in chronological order, and removes image frames whose timestamp deviation exceeds the synchronization threshold. Distortion correction and pixel coordinate remapping are performed on each original image frame in the camera synchronization image frame group. The image is cropped and scaled according to the input resolution. The image is color space consistent to obtain a corrected image frame with a unified color channel expression. The pixel values ​​of the calibrated image frames are normalized, the data format is converted, the image data is converted into a preset tensor format, the channel order is arranged, the corresponding camera intrinsic and extrinsic data are associated with each calibrated image frame, and the preprocessed image sequence is output.

4. The robot path planning and obstacle avoidance system based on AI visual detection according to claim 1, characterized in that, The improved BEVFormer perception module includes: An improved BEVFormer model is constructed, which includes corridor anchor point query, safety boundary driven sampling band and temporal occupancy prediction attention structure. Camera features are extracted from the preprocessed image sequence to obtain the image feature sequence arranged by camera channel, and a set of bird’s-eye grid query vectors is established in the bird’s-eye coordinate system. Based on the corridor anchor point query, a set of corridor query vectors for candidate travel corridors is generated. The candidate travel corridors are represented in a parametric way and are determined by the parameters of corridor centerline, corridor width, corridor orientation and corridor curvature. The set of corridor query vectors and the set of bird’s-eye view query are used together as the query input for bird’s-eye view spatial encoding and corridor-level feature aggregation is performed. Cross-view sampling positions are determined by using a safety boundary-driven sampling band. The safety boundary-driven sampling band is generated by the safety margin and distributed along the robot's safety boundary contour and forward fan-shaped region in the bird's-eye view space. The cross-view sampling positions are projected onto the pixel coordinate system of the image feature sequence of each camera to obtain the set of sampling points for each camera. Key vectors and value vectors are extracted from the image feature sequence using the sampling point set of each camera. Spatial cross-attention fusion is performed using the bird’s-eye view query set and the corridor query vector set to obtain the bird’s-eye view feature map at the current moment. Temporal occupancy prediction attention is used to perform temporal fusion of the current bird's-eye view feature map and the historical bird's-eye view feature map. Based on the temporal fusion result, occupancy prediction output is generated, and the output consists of a bird's-eye view occupancy raster sequence and a bird's-eye view risk raster sequence arranged by time step.

5. The robot path planning and obstacle avoidance system based on AI visual detection according to claim 1, characterized in that, The secure expansion processing module includes: Read the current-time grid cells from the bird's-eye view occupied grid sequence and risk grid sequence, convert the occupied status and risk level of each cell in the grid into passable and impassable markers, and generate a grid set of impassable areas; The expansion radius is determined based on the robot's safety margin. The expansion radius is equal to the sum of the robot's outer radius and the minimum safe distance, plus the distance margin corresponding to the pose uncertainty. In the bird's-eye view grid coordinate system, the grid set of the impassable area is subjected to morphological expansion processing to obtain the expanded impassable area. A locally traversable region is constructed using grid cells within the bird's-eye view grid coverage area, excluding the expanded impassable areas. The grid representation of the locally traversable region and its corresponding spatial boundary information are then output.

6. The robot path planning and obstacle avoidance system based on AI visual detection according to claim 1, characterized in that, The polar coordinate mapping module includes: A polar coordinate grid is established with the bird's-eye view coordinates corresponding to the robot's current pose as the origin of the polar coordinate grid. The polar coordinate grid consists of an angle dimension and a distance dimension. The angle dimension is divided into angle sectors, and the distance dimension is divided into distance loops. The bird's-eye view grid cells in the local traversable domain are converted into polar coordinate cells. For each bird's-eye view grid cell, the distance and azimuth values ​​from the polar coordinate origin to the center point of the bird's-eye view grid cell are calculated. The traversable markers of the bird's-eye view grid cells are mapped to polar coordinate cells containing the current distance and azimuth values ​​to generate a polar coordinate traversable domain representation. The target direction is calculated based on the coordinates of the target point in the bird's-eye view coordinate system and the coordinates of the robot's current pose in the bird's-eye view coordinate system. The target direction is the azimuth angle from the robot's current position to the target point. The target direction is mapped to the angle dimension of the polar coordinate grid to determine the target angle sector where the target direction is located.

7. The robot path planning and obstacle avoidance system based on AI visual detection according to claim 1, characterized in that, The vector field histogram planning module includes: Read the polar coordinate passable domain representation and scan the polar coordinate passable domain representation along the angular dimension. Determine the minimum impassable distance from the polar coordinate origin in each angular sector. Convert the angular range corresponding to the minimum impassable distance into a blocking angular interval and generate a set of blocking angular intervals. An angle interval graph is used to perform interval merging and interval adjacency establishment on the set of blocked angle intervals. The set of blocked angle intervals is deduplicated and merged in the angle dimension to form the connection relationship of blocked intervals. The set of passable angle intervals and the connection relationship of passable angle intervals are generated on the complement of the connection relationship of blocked intervals. A polar coordinate distance field is constructed within the polar coordinate traversable domain defined by the set of traversable angle intervals. The distance field value of each polar coordinate unit in the polar coordinate distance field is defined as the shortest distance from the polar coordinate unit to the boundary of the impassable region. The shortest distance is the minimum value of the Euclidean distance from the center point of the polar coordinate unit to any point on the boundary of the impassable region. The channel center ridge sequence is determined by extracting the polar coordinate distance field ridge. The channel center ridge sequence consists of polar coordinate units that satisfy the local maximum condition. The local maximum condition is that the distance field value of the polar coordinate unit is greater than the distance field value of the adjacent polar coordinate unit in the angular dimension, and is also greater than the distance field value of the adjacent polar coordinate unit in the distance dimension. The channel center ridge sequence is mapped to the angular dimension to obtain the channel center direction sequence corresponding to the channel center ridge sequence.

8. A robot path planning and obstacle avoidance system based on AI visual detection according to claim 1, characterized in that, The route generation control module includes: Receive the target direction, the set of passable angle intervals, the connectivity of the passable angle intervals, and the channel center ridge sequence. Determine the target candidate interval containing the target direction from the set of passable angle intervals. When the target direction does not fall into any passable angle interval, determine the candidate interval with the smallest angular deviation from the target direction based on the connectivity of the passable angle intervals. A local travel route is generated based on the channel center ridge sequence. The local travel route is composed of the point sequence of the channel center ridge sequence in the bird's-eye view coordinate system. The continuous path curve is obtained by curve fitting of the point sequence. The curvature of the continuous path curve is determined by the rate of change of the tangent direction of the path curve at the corresponding position. The control commands are calculated based on the local travel path. The control commands include linear velocity and angular velocity. The linear velocity is the smaller of the preset maximum linear velocity and the upper limit of the safe linear velocity calculated based on the nearest obstacle distance on the local travel path. The angular velocity is the product of the linear velocity and the curvature of the local travel path at the robot's current position.

9. A robot path planning and obstacle avoidance system based on AI visual detection according to claim 1, characterized in that, The validity verification execution module includes: Receive the bird's-eye view occupied grid sequence, risk grid sequence and local travel route, discretize the local travel route into a path sampling point sequence in the bird's-eye view coordinate system, and configure the path sampling point sequence with a time index that increases according to the control period to cover the current time to the future time range; For each path sampling point in the path sampling point sequence, based on the time index, select the corresponding time step of the occupancy grid and risk grid from the bird's-eye view occupancy grid sequence and the risk grid sequence, determine the occupancy status of the path sampling point in the corresponding occupancy grid, read the risk value of the path sampling point in the corresponding risk grid, and generate the path occupancy determination sequence and the path risk value sequence. When there are no impassable markers in the path occupancy determination sequence and the maximum risk value in the path risk value sequence is less than the preset risk threshold, the local travel route is determined to be valid, and control commands are output to drive the robot to move. When the conditions are not met, the local travel route is determined to be invalid, and preset safety control commands are output to regenerate the local travel route.

Images