Shunting operation environment intelligent sensing method based on vision and radar

Abstract

The invention discloses a shunting operation environment intelligent sensing method based on vision and radar. The method comprises the following steps: generating a fusion point cloud by using binocular vision and radar; constructing a track topology reference point cloud according to line design data, and forming a safe reachable domain in combination with train power parameters; extracting train contour points and track adjacent points from the fusion point cloud, and sampling prediction track points according to a safe reachable domain to form a space-time four-point base set; searching a point set meeting the relationship between the distance ratio and the orbital sequence based on the improved Super4PCS, and solving rigid body transformation candidates; acting the rigid body transformation candidates on the fused point cloud and calculating a geometric residual error and a safety margin to obtain a collaborative score; selecting rigid body transformation with the highest score as a train registration pose and generating a residual risk adaptive weight field; and splicing the registration pose sequences into a train movement track, and generating an environment perception result and a safety protection instruction in combination with a safety reachable domain. According to the invention, the stability of shunting environment registration and safety perception is improved.

Description

Technical Field

[0001] This invention relates to the fields of machine vision and rail transit safety perception technology, and in particular to an intelligent perception method for shunting operation environment based on vision and radar. Background Technology

[0002] Shunting operations are a critical link in the railway transportation system with extremely high frequency and concentrated risks. They involve continuous actions such as train movement, coupling and uncoupling, switch switching, and crossing station lines, and are characterized by a highly dynamic environment and densely distributed personnel and equipment. Traditional shunting operations mainly rely on manual observation of the surrounding environment, handheld terminal commands, and fixed camera monitoring for supplementary judgment. However, these methods are limited by changes in lighting, obstructions, insufficient observation range, and human eye fatigue, resulting in unstable environmental perception. Regarding visual perception and ranging technology at the locomotive end, existing research mostly uses monocular cameras, binocular cameras, or lidar for obstacle detection. However, these typically use only one sensor and lack multi-source information fusion mechanisms, making them prone to instability in rain, snow, nighttime, smoke, or backlighting conditions. Furthermore, existing point cloud processing methods in railway operation scenarios primarily focus on static target recognition, making it difficult to obtain continuous and reliable three-dimensional environmental structures during locomotive movement, thus hindering the provision of stable input for shunting automation.

[0003] In point cloud registration and train pose estimation, existing studies often employ ICP-type algorithms, NDT-type algorithms, or geometric matching algorithms based on 4PCS for registration. However, these methods generally assume a static environment and regular structure. When faced with the complex topology of railway spaces such as tracks, lines, and turnouts, matching is prone to mismatches, local optima, or sensitivity to noise. Especially in shunting operations, the locomotive position change range is small, the field of view is limited, and the local overlap area of ​​the point cloud is insufficient. Traditional ICP methods are highly sensitive to the initial pose and cannot reliably obtain correct registration results. Although the standard Super4PCS improves global matching capabilities, it does not consider track geometric topology relationships or train dynamic constraints, thus selecting a large number of unreasonable four-point combinations, resulting in low efficiency and low reliability of rigid body transformation solutions. At the same time, existing studies rely heavily on a single geometric error index in evaluating registration results, lacking the safety margin constraints of track operating boundaries, making it impossible to provide more reliable pose judgments in safety-related scenarios.

[0004] In track environment safety analysis, existing technologies mostly use fixed cameras, ground sensors, or section occupancy detection devices to generate track occupancy status, which cannot cover the dynamic perspective of the locomotive. In locomotive-side environmental perception, some literature proposes using deep learning for obstacle detection and track segmentation, but the results are difficult to correlate with train motion and reachable areas, and cannot adjust evaluation weights in real time according to registration errors. Furthermore, existing methods do not incorporate the spatially varying characteristics of perception errors into sampling strategies, failing to achieve focused detection of key track areas and lacking a risk adaptive mechanism that can be updated over time. In shunting operations, track alignment, turnout distribution, and safety boundaries all change with location; existing methods lack the ability to jointly judge using track topology and train dynamic constraints, making it difficult to generate stable track occupancy status, offset assessments, and safety protection commands in complex scenarios.

[0005] Therefore, how to provide an intelligent perception method for shunting operation environment based on vision and radar is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose an intelligent perception method for shunting operation environment based on vision and radar. This invention utilizes binocular vision and radar fusion imaging, combines track topology reference point cloud with train dynamics safe reachability domain, and constructs a spatiotemporal four-point basis and residual risk adaptive weight field under the improved Super4PCS registration framework to achieve integrated processing of train pose estimation, track occupancy determination and obstacle risk analysis. It has the advantages of high environmental perception accuracy, quantifiable safety margin and strong adaptability to complex track and turnout environments in shunting scenarios.

[0007] A method for intelligent perception of shunting operation environment based on vision and radar according to an embodiment of the present invention includes the following steps: The train's front-end binocular vision sensor and radar sensor collect images and distance data, which are then calibrated and transformed to generate a fused point cloud. Generate a track topology reference point cloud based on the line design data, calculate the reachable pose according to the train dynamic parameters, and construct a safe reachable domain. Train outline points and track neighboring points are extracted from the fused point cloud, and predicted trajectory points are selected from the track topology reference point cloud according to the safe reach domain, and combined into a spatiotemporal four-point basis set. Based on the spatiotemporal four-point basis set, the improved Super4PCS is called to search for the spatiotemporal four-point basis set that satisfies the distance ratio and track order relationship in the track topology reference point cloud, and solve for the rigid body transformation candidate that satisfies the safe reachability domain constraint. The rigid body transformation candidate is applied to the fused point cloud to calculate the geometric residual and safety margin, and a collaborative scoring function is constructed based on the geometric residual and safety margin. The rigid body transformation with the highest collaborative score is selected as the train registration pose. The residual field is updated based on the geometric residual, and an adaptive weight field for residual risk is generated to adjust the spatiotemporal four-point basis set sampling and collaborative score weights in the next time window. The train registration pose sequence is stitched together to form the train motion trajectory. Combined with the safe reachable domain, track topology reference point cloud and residual risk adaptive weight field, environmental perception results and safety protection instructions are generated.

[0008] Optionally, the improved Super4PCS: Using points in the fused point cloud that belong to the train outline region and the track adjacent region, and predicted trajectory points within the safe reach domain as input, a spatiotemporal four-point basis set with time index is constructed. Search for spatiotemporal four-point sets that satisfy the spatiotemporal conformal constraints in the orbital topology reference point cloud according to the consistency of the Euclidean distance ratio of the spatiotemporal four-point basis sets and the monotonic order relationship along the orbital centerline, and solve for rigid body transformation candidates for each spatiotemporal four-point set. Candidate rigid body transformations with translation vectors and attitude angles exceeding the safe reachability domain are filtered out. The remaining candidates are then applied to the fused point cloud. The geometric residuals within the prediction time window and the minimum safety margin relative to the boundary of the safe reachability domain are statistically analyzed. A collaborative scoring function is constructed with the goal of reducing geometric residuals and increasing safety margin. Candidate rigid body transformations are scored and ranked, and the rigid body transformation with the highest collaborative score is selected as the train registration pose.

[0009] Optionally, the generation of the fused point cloud specifically includes: A binocular vision sensor and a radar sensor are installed at the front of the train, facing the direction of travel. The field of view covers the track area and the adjacent area in front of the track. A time synchronization method is set so that the binocular vision sensor and the radar sensor can collect image data and distance data in the direction of travel under a unified time reference, and time tags are registered for the collected data. Perform geometric calibration and distortion correction on the binocular vision sensor, process the image data according to the calibration results, convert the image data with the same time label into a binocular point cloud representing the spatial structure of the scene in front, and register a time label consistent with the image data for the binocular point cloud; The radar sensor is calibrated, the distance data is processed according to the calibration parameters, the distance data with the same time label is converted into a radar point cloud representing the spatial structure of the scene in front, and the radar point cloud is registered with a time label consistent with the binocular point cloud. Establish coordinate transformation relationships between the binocular vision sensor coordinate system, the radar sensor coordinate system, and the orbital coordinate system. Transform the binocular point cloud and the radar point cloud to the orbital coordinate system. At each sampling time, according to the time label, merge the points in the binocular point cloud and the points in the radar point cloud in the orbital coordinate system into a frame of fused point cloud, forming a fused point cloud sequence in the orbital coordinate system.

[0010] Optionally, the construction of the secure reachable domain specifically includes: Read the track design data and import it into the train's onboard control unit. The track design data includes the track centerline coordinates, track number, turnout position parameters, platform edge position parameters, and clearance boundary parameters. Based on the track design data, align and transform the track centerline in the track coordinate system. Discrete sampling is performed along the mileage direction according to the centerline coordinates of the track in the track coordinate system. Discrete sampling points are generated at the centerline positions of each track according to the track number. Discrete sampling points are generated in the frog area and switch area according to the turnout position parameters. Discrete sampling points are generated at the platform edge position and the clearance boundary position according to the platform edge position parameters and the clearance boundary parameters. The mileage value, lateral offset and topology type mark are registered for each discrete sampling point. All discrete sampling points are combined into a track topology reference point cloud. Train dynamic parameters are obtained, including train mass, permissible speed limit, traction capacity parameters, and braking capacity parameters. The train operation process is divided into discrete time steps within a preset prediction time window. Based on the train dynamic parameters and each discrete time step, the reachable position coordinates and attitude parameters of the train in the track coordinate system are calculated. The position coordinates and attitude parameters corresponding to each discrete time step are registered as reachable poses and combined to form a set of reachable poses. In the orbital coordinate system, the position range and attitude range are determined based on the set of reachable poses. The position range is represented as the spatial region that encloses the orbital centerline, and the attitude range is represented as the range of attitude values ​​allowed for the corresponding position coordinates. Time labels corresponding to the discrete time steps within the prediction time window are registered for the position range and attitude range. The position range, attitude range, and time labels are combined into a safe reachable domain.

[0011] Optionally, the construction of the spatiotemporal four-point basis set specifically includes: In the track coordinate system, points belonging to the outer surface of the train are selected from the fused point cloud based on the train's external dimensions and installation location. The outer envelope of the train is established, and points that meet the envelope conditions are marked as train outline points. Based on the track centerline, points whose lateral distance falls within the threshold range are selected from the fused point cloud according to a pre-set lateral distance threshold and marked as track neighbor points. The time labels registered in the fused point cloud are inherited, and train outline points and track neighbor points with the same time label are grouped into the same data set to form the current point set with the corresponding time label. Each current point set records the spatial coordinates and time labels of the train outline points and track neighbor points in the track coordinate system. Based on the orbit centerline and the safe reachability region in the orbit coordinate system, predictive trajectory points are sampled from the orbit topology reference point cloud along the orbit centerline at discrete time steps within the prediction time window. The predicted trajectory points satisfy the constraint condition of falling within the safe reachability region. A time label and topology type mark are registered for each predicted trajectory point. All predicted trajectory points are grouped according to the time label to form a set of predicted trajectory points. Each predicted trajectory point in the set of predicted trajectory points records the spatial coordinates, time label and topology type mark in the orbit coordinate system. Under each time label, the train outline points and track neighboring points in the current point set are sorted in ascending order of mileage value, where mileage value is the distance parameter along the track centerline. From the sorting results, the train outline point with the smallest mileage value and the track neighboring point with the largest mileage value are selected as the current two points. Under the same time label, the predicted trajectory points in the predicted trajectory point set are sorted in ascending order of mileage value. From the sorting results, the predicted trajectory point with the smallest mileage value and the predicted trajectory point with the largest mileage value are selected as the two predicted trajectory points. The current two points and the two predicted trajectory points are combined into a set of spatiotemporal four-point bases, and the spatiotemporal four-point bases corresponding to all time labels are combined into a set of spatiotemporal four-point bases.

[0012] Optionally, the generation of the rigid body transformation candidates specifically includes: In the improved Super4PCS, the coordinates of the four points of each set of spatiotemporal four-point basis in the orbital coordinate system are read from the spatiotemporal four-point basis set. The Euclidean distance between any two of the four points is calculated to obtain the pairwise distance between the four points. The pairwise distances are combined into a set of side lengths. The ratio of any two side lengths in the set of side lengths is recorded as the side length ratio value. The side lengths and side length ratio values ​​are arranged in a preset order, and time tags and topology type tags are registered to form a set of spatiotemporal four-point basis features. In the orbital topology reference point cloud, candidate point sets are filtered according to time labels and topology type labels. Under each time label, four points are selected from the candidate point sets to form a reference four-point combination. The pairwise distance and side length ratio of the four points in the reference four-point combination are calculated. The side length ratio of the reference four-point combination is compared item by item with the side length ratio of the same time label and the same topology type label in the spatiotemporal four-point basic feature set. When the difference of all side length ratios falls within the preset error threshold and the mileage values ​​of the four points in the direction of the orbital centerline are monotonically related, the reference four-point combination is registered as a reference spatiotemporal four-point in the spatiotemporal four-point set. For each set of four-point base points and corresponding reference four-point base points, read the spatial coordinates of the four points in the orbital coordinate system. Calculate the centroid coordinates of the four points in the base point and the centroid coordinates of the four points in the reference four-point base. Combine the centroid coordinates of each set of four points with the coordinate differences of each point to form a decentroided coordinate set. Construct a covariance relationship matrix between the point pairs based on the two decentroided coordinate sets. Obtain the rotation matrix used to align the two decentroided coordinate sets based on the covariance relationship matrix. Apply the rotation matrix to the decentroided coordinate set of the base point. Determine the rotation matrix by comparing the difference between the rotated coordinates and the decentroided coordinates of the reference four-point base. Calculate the translation vector based on the rotation matrix and the two sets of centroid coordinates. Combine the rotation matrix and translation vector into a rigid body transformation and register a time label and topology type mark for the rigid body transformation. A set of train reference points is selected in the orbital coordinate system. Each rigid body transformation is applied to the set of train reference points to obtain the spatial position and attitude of the set of train reference points after the rigid body transformation. The position and attitude of each point in the set of train reference points are compared according to the position range and attitude range in the safe reachable domain. When the position and attitude of all train reference points fall within the coverage of the safe reachable domain, the corresponding rigid body transformation is registered as a rigid body transformation candidate and added to the rigid body transformation candidate set.

[0013] Optionally, the construction of the collaborative scoring function specifically includes: In the orbital coordinate system, each rigid body transformation is read from the rigid body transformation candidate set. Each rigid body transformation is applied to the fusion point cloud of the corresponding time label according to the time label, and a transformed fusion point cloud frame corresponding to the rigid body transformation is generated. The rigid body transformation number and time label are registered for each frame of transformed fusion point cloud. For each rigid body transformation in the candidate set of rigid body transformations, the points in the corresponding transformed fused point cloud are traversed in the orbit coordinate system. The nearest point of each point is found in the orbit topology reference point cloud through nearest neighbor search. The distance from each point to the nearest point is calculated as the point-level geometric residual. The point-level geometric residuals are assigned to the corresponding time windows according to the time label. The distance mean, distance variance and preset quantile are calculated for the point-level geometric residuals in each time window to generate geometric residual statistics associated with the rigid body transformation. A set of train reference points is established in the track coordinate system. The points in the set of train reference points cover the front position, rear position, and key positions of the train body. For each rigid body transformation in the candidate set of rigid body transformations, the rigid body transformation is applied to the set of train reference points to obtain the spatial position of the set of train reference points after the rigid body transformation. For each point in the set of train reference points, the shortest distance to the boundary of the safe reachable domain is calculated according to the definition of the boundary of the safe reachable domain. The minimum value is selected from all the distance values ​​of the train reference points and recorded as the safety margin corresponding to the rigid body transformation. A collaborative scoring function is constructed, with geometric residual statistics and safety margin as inputs. For each rigid body transformation in the candidate set, a penalty term is set based on the distance mean, distance variance, and preset quantile in the geometric residual statistics, and a reward term is set based on the safety margin. The collaborative score value is obtained by combining the penalty and reward terms. The collaborative score value increases as the geometric residual statistics decrease and as the safety margin increases. A collaborative score value is registered for each rigid body transformation in the candidate set.

[0014] Optionally, the generation of the residual risk adaptive weight field specifically includes: The train onboard control unit reads the coordination score of each rigid body transformation from the candidate set of rigid body transformations, sorts the coordination score values, determines the rigid body transformation with the largest coordination score, registers the train pose corresponding to the selected rigid body transformation as the train registration pose of the current time window in the track coordinate system, and establishes a correspondence between the train registration pose of the current time window and the time window identifier. The train onboard control unit applies the train registration pose of the current time window to the fused point cloud of the current time window in the track coordinate system to obtain the fused point cloud after the transformation of the current time window. In the track coordinate system, a nearest neighbor search is performed between each point in the fused point cloud after the transformation of the current time window and the track topology reference point cloud to obtain the nearest neighbor with the smallest distance and calculate the distance between the two points. The distance is recorded as a point-level geometric residual. The track coordinate system is divided into spatial grids, and each point-level geometric residual is assigned to the corresponding spatial grid according to the spatial grid in which the point is located. The average value and preset quantile value are calculated for the point-level geometric residual in each spatial grid to form a residual field indexed by the spatial grid. The train's onboard control unit combines the residual field with the preset risk level configuration, assigns higher weights to spatial grids with larger residual values ​​and lower weights to spatial grids with smaller residual values, constructs an adaptive weight field for residual risk in the track coordinate system, registers weight values ​​for each spatial grid, and establishes a correspondence between the adaptive weight field for residual risk and the time window identifier. In the next time window, the train's onboard control unit adjusts the sampling probability of the spatiotemporal four-point basis in the spatiotemporal four-point basis set according to the weight values ​​of the spatial grid in the residual risk adaptive weight field. It increases the sampling probability of the spatiotemporal four-point basis in the spatial grid with higher weight values. In the collaborative scoring function, it sets the weight coefficients of the geometric residual related parameters according to the weight values ​​corresponding to the spatial grid. It generates a spatiotemporal four-point basis set sampling scheme and collaborative scoring parameter configuration for the next time window, and provides the sampling scheme and parameter configuration to the improved Super4PCS calculation process for the next time window.

[0015] Optionally, the generation of the environmental perception results and security protection instructions specifically includes: In the orbital coordinate system, the train registration poses are read in the order of time tags. The train registration poses corresponding to each time tag are connected in chronological order to form a train registration pose sequence. Time tags and mileage indexes are registered for each train registration pose in the train registration pose sequence. In the track coordinate system, the spatial position of each train registration pose in the train registration pose sequence is projected onto the track centerline direction to obtain the mileage direction position sequence. At the same time, the offset of each train registration pose in the lateral direction relative to the track centerline is calculated. The mileage direction position sequence and the lateral offset are combined to form the train motion trajectory, where each time label in the train motion trajectory corresponds to the mileage direction position and the lateral offset. In the track coordinate system, the train trajectory is superimposed on the safe reachability domain. The current position coordinates of the train are read for each time tag. The shortest distance between the current position of the train and the boundary of the safe reachability domain is calculated according to the boundary position defined by the safe reachability domain. The shortest distance is recorded as the trajectory safety margin of the corresponding time tag and arranged in the order of the time tags to form a trajectory safety margin sequence. In the track coordinate system, the train trajectory, track safety margin sequence, track topology reference point cloud, and residual risk adaptive weight field are input into the environment analysis module. The track is divided into sections along the track centerline in the mileage direction. Within each track section, the track occupancy status is determined by combining the track topology reference point cloud. The degree of lateral offset of the train is calculated by combining the lateral offset in the train trajectory. The obstacle risk level is determined by combining the spatial grid weight values ​​in the residual risk adaptive weight field with the distance values ​​in the track safety margin sequence. The track occupancy status, lateral offset, and obstacle risk level are output according to the time label to form the environmental perception result. The train's onboard control unit generates safety protection commands according to preset control rules based on environmental perception results and trajectory safety margin sequence. Within the time tag interval where any one of the three conditions is met—track occupancy status, obstacle risk level exceeding a preset threshold, or trajectory safety margin below a preset threshold—speed control commands, braking control commands, and alarm commands are generated according to preset priorities, and the safety protection commands are sent to the train control system and shunting operation control system for execution.

[0016] The beneficial effects of this invention are: This invention fuses data from binocular vision sensors and radar sensors in a unified track coordinate system, and introduces a track topology reference point cloud constructed based on track design data and a safe reachable domain calculated from train dynamic parameters. Under the improved Super4PCS framework, train pose registration and environmental modeling are completed, forming a unified perception system that coordinates "perceptual data - track topology - dynamic constraints". Compared with schemes that rely solely on image detection or local point cloud matching, this invention can still maintain stable three-dimensional environmental representation and train pose estimation capabilities in complex track and turnout areas, providing a structured and traceable spatial basis for subsequent track occupancy determination and safety decisions.

[0017] This invention constructs a spatiotemporal four-point basis set with time index, jointly encodes train contour points, track neighbor points, and predicted trajectory points in the fused point cloud and track topology reference point cloud, and introduces distance ratio constraints, track order constraints, and safe reachability domain constraints into the improved Super4PCS. Combined with a collaborative scoring function composed of geometric residual statistics and safety margin, this achieves a train pose solution that balances geometric registration accuracy and operational safety margin. Furthermore, this invention constructs a residual field and an adaptive weight field for residual risk based on registration error, and inversely adjusts the spatiotemporal four-point basis sampling strategy and collaborative scoring weights for the next time window. This enables the registration process to adaptively focus on high-risk areas, suppress long-term accumulated errors, and improve the continuity and robustness of environmental perception results during shunting operations.

[0018] Based on the completion of the train registration pose time series construction, this invention combines the safe reachability domain, track topology reference point cloud, and residual risk adaptive weight field to uniformly output track occupancy status, trajectory offset, obstacle risk level, as well as speed control, braking control, and alarm commands. This realizes an integrated closed loop of environmental perception and safety protection for shunting operations. Compared with the traditional method that relies on ground signal equipment and manual observation, this invention can provide control commands with spatial geometric basis and safety margin quantitative support in real time at the locomotive end, which helps to reduce the risk of human misjudgment and improve the automation level and safety protection capability of shunting operations. Attached Figure Description

[0019] 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:

[0020] Figure 1 This is a flowchart of an intelligent perception method for shunting operation environment based on vision and radar proposed in this invention; Figure 2This is a schematic diagram of the improved Super4PCS spatiotemporal four-point base construction of a vision and radar-based intelligent perception method for shunting operation environment proposed in this invention. Detailed Implementation

[0021] 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.

[0022] refer to Figure 1-2 A method for intelligent perception of shunting operation environment based on vision and radar includes the following steps: The train's front-end binocular vision sensor and radar sensor collect images and distance data, which are then calibrated and transformed to generate a fused point cloud. Generate a track topology reference point cloud based on the line design data, calculate the reachable pose according to the train dynamic parameters, and construct a safe reachable domain. Train outline points and track neighboring points are extracted from the fused point cloud, and predicted trajectory points are selected from the track topology reference point cloud according to the safe reach domain, and combined into a spatiotemporal four-point basis set. Based on the spatiotemporal four-point basis set, the improved Super4PCS is called to search for the spatiotemporal four-point basis set that satisfies the distance ratio and track order relationship in the track topology reference point cloud, and solve for the rigid body transformation candidate that satisfies the safe reachability domain constraint. The rigid body transformation candidate is applied to the fused point cloud to calculate the geometric residual and safety margin, and a collaborative scoring function is constructed based on the geometric residual and safety margin. The rigid body transformation with the highest collaborative score is selected as the train registration pose. The residual field is updated based on the geometric residual, and an adaptive weight field for residual risk is generated to adjust the spatiotemporal four-point basis set sampling and collaborative score weights in the next time window. The train registration pose sequence is stitched together to form the train motion trajectory. Combined with the safe reachable domain, track topology reference point cloud and residual risk adaptive weight field, environmental perception results and safety protection instructions are generated.

[0023] In this embodiment, the improved Super4PCS: Using points in the fused point cloud that belong to the train outline region and the track adjacent region, and predicted trajectory points within the safe reach domain as input, a spatiotemporal four-point basis set with time index is constructed. Search for spatiotemporal four-point sets that satisfy the spatiotemporal conformal constraints in the orbital topology reference point cloud according to the consistency of the Euclidean distance ratio of the spatiotemporal four-point basis sets and the monotonic order relationship along the orbital centerline, and solve for rigid body transformation candidates for each spatiotemporal four-point set. Candidate rigid body transformations with translation vectors and attitude angles exceeding the safe reachability domain are filtered out. The remaining candidates are then applied to the fused point cloud. The geometric residuals within the prediction time window and the minimum safety margin relative to the boundary of the safe reachability domain are statistically analyzed. A collaborative scoring function is constructed with the goal of reducing geometric residuals and increasing safety margin. Candidate rigid body transformations are scored and ranked, and the rigid body transformation with the highest collaborative score is selected as the train registration pose.

[0024] This invention constructs a spatiotemporal four-point basis with time index by selecting train contour points, track neighboring points, and predicted trajectory points within the safe reach domain from the fused point cloud. In the track topology reference point cloud, candidate point groups satisfying spatiotemporal conformal relationships are searched under constraints of distance ratio consistency and the order relationship along the track centerline. Rigid body transformations that do not conform to train motion capabilities are then eliminated based on the safe reach domain. This ensures that the registration process is subject to dual constraints from the track geometry and dynamic boundaries from the outset, effectively compressing the search space and significantly reducing mismatches. Subsequently, a collaborative scoring model is constructed using geometric residuals and minimum safety margin, enabling the algorithm to identify the most reliable rigid body transformations faster at the same resolution. This maintains stable registration performance even in environments with dense turnouts, significant curve radius variations, and insufficient local overlap of point clouds, ensuring that train pose estimation has a quantifiable safety margin.

[0025] In this embodiment, the generation of the fused point cloud specifically includes: A binocular vision sensor and a radar sensor are installed at the front of the train, facing the direction of travel. The field of view covers the track area and the adjacent area in front of the track. A time synchronization method is set so that the binocular vision sensor and the radar sensor can collect image data and distance data in the direction of travel under a unified time reference, and time tags are registered for the collected data. Perform geometric calibration and distortion correction on the binocular vision sensor, process the image data according to the calibration results, convert the image data with the same time label into a binocular point cloud representing the spatial structure of the scene in front, and register a time label consistent with the image data for the binocular point cloud; The radar sensor is calibrated, the distance data is processed according to the calibration parameters, the distance data with the same time label is converted into a radar point cloud representing the spatial structure of the scene in front, and the radar point cloud is registered with a time label consistent with the binocular point cloud. Establish coordinate transformation relationships between the binocular vision sensor coordinate system, the radar sensor coordinate system, and the orbital coordinate system. Transform the binocular point cloud and the radar point cloud to the orbital coordinate system. At each sampling time, according to the time label, merge the points in the binocular point cloud and the points in the radar point cloud in the orbital coordinate system into a frame of fused point cloud, forming a fused point cloud sequence in the orbital coordinate system.

[0026] In this embodiment, the construction of the secure reachable domain specifically includes: Read the track design data and import it into the train's onboard control unit. The track design data includes the track centerline coordinates, track number, turnout position parameters, platform edge position parameters, and clearance boundary parameters. Based on the track design data, align and transform the track centerline in the track coordinate system. Discrete sampling is performed along the mileage direction according to the centerline coordinates of the track in the track coordinate system. Discrete sampling points are generated at the centerline positions of each track according to the track number. Discrete sampling points are generated in the frog area and switch area according to the turnout position parameters. Discrete sampling points are generated at the platform edge position and the clearance boundary position according to the platform edge position parameters and the clearance boundary parameters. The mileage value, lateral offset and topology type mark are registered for each discrete sampling point. All discrete sampling points are combined into a track topology reference point cloud. Train dynamic parameters are obtained, including train mass, permissible speed limit, traction capacity parameters, and braking capacity parameters. The train operation process is divided into discrete time steps within a preset prediction time window. Based on the train dynamic parameters and each discrete time step, the reachable position coordinates and attitude parameters of the train in the track coordinate system are calculated. The position coordinates and attitude parameters corresponding to each discrete time step are registered as reachable poses and combined to form a set of reachable poses. In the orbital coordinate system, the position range and attitude range are determined based on the set of reachable poses. The position range is represented as the spatial region that encloses the orbital centerline, and the attitude range is represented as the range of attitude values ​​allowed for the corresponding position coordinates. Time labels corresponding to the discrete time steps within the prediction time window are registered for the position range and attitude range. The position range, attitude range, and time labels are combined into a safe reachable domain.

[0027] In this embodiment, the construction of the spatiotemporal four-point basis set specifically includes: In the track coordinate system, points belonging to the outer surface of the train are selected from the fused point cloud based on the train's external dimensions and installation location. The outer envelope of the train is established, and points that meet the envelope conditions are marked as train outline points. Based on the track centerline, points whose lateral distance falls within the threshold range are selected from the fused point cloud according to a pre-set lateral distance threshold and marked as track neighbor points. The time labels registered in the fused point cloud are inherited, and train outline points and track neighbor points with the same time label are grouped into the same data set to form the current point set with the corresponding time label. Each current point set records the spatial coordinates and time labels of the train outline points and track neighbor points in the track coordinate system. Based on the orbit centerline and the safe reachability region in the orbit coordinate system, predictive trajectory points are sampled from the orbit topology reference point cloud along the orbit centerline at discrete time steps within the prediction time window. The predicted trajectory points satisfy the constraint condition of falling within the safe reachability region. A time label and topology type mark are registered for each predicted trajectory point. All predicted trajectory points are grouped according to the time label to form a set of predicted trajectory points. Each predicted trajectory point in the set of predicted trajectory points records the spatial coordinates, time label and topology type mark in the orbit coordinate system. Under each time label, the train outline points and track neighboring points in the current point set are sorted in ascending order of mileage value, where mileage value is the distance parameter along the track centerline. From the sorting results, the train outline point with the smallest mileage value and the track neighboring point with the largest mileage value are selected as the current two points. Under the same time label, the predicted trajectory points in the predicted trajectory point set are sorted in ascending order of mileage value. From the sorting results, the predicted trajectory point with the smallest mileage value and the predicted trajectory point with the largest mileage value are selected as the two predicted trajectory points. The current two points and the two predicted trajectory points are combined into a set of spatiotemporal four-point bases, and the spatiotemporal four-point bases corresponding to all time labels are combined into a set of spatiotemporal four-point bases.

[0028] In this embodiment, the generation of rigid body transformation candidates specifically includes: In the improved Super4PCS, the coordinates of the four points of each set of spatiotemporal four-point basis in the orbital coordinate system are read from the spatiotemporal four-point basis set. The Euclidean distance between any two of the four points is calculated to obtain the pairwise distance between the four points. The pairwise distances are combined into a set of side lengths. The ratio of any two side lengths in the set of side lengths is recorded as the side length ratio value. The side lengths and side length ratio values ​​are arranged in a preset order, and time tags and topology type tags are registered to form a set of spatiotemporal four-point basis features. In the orbital topology reference point cloud, candidate point sets are filtered according to time labels and topology type labels. Under each time label, four points are selected from the candidate point sets to form a reference four-point combination. The pairwise distance and side length ratio of the four points in the reference four-point combination are calculated. The side length ratio of the reference four-point combination is compared item by item with the side length ratio of the same time label and the same topology type label in the spatiotemporal four-point basic feature set. When the difference of all side length ratios falls within the preset error threshold and the mileage values ​​of the four points in the direction of the orbital centerline are monotonically related, the reference four-point combination is registered as a reference spatiotemporal four-point in the spatiotemporal four-point set. For each set of four-point base points and corresponding reference four-point base points, read the spatial coordinates of the four points in the orbital coordinate system. Calculate the centroid coordinates of the four points in the base point and the centroid coordinates of the four points in the reference four-point base. Combine the centroid coordinates of each set of four points with the coordinate differences of each point to form a decentroided coordinate set. Construct a covariance relationship matrix between the point pairs based on the two decentroided coordinate sets. Obtain the rotation matrix used to align the two decentroided coordinate sets based on the covariance relationship matrix. Apply the rotation matrix to the decentroided coordinate set of the base point. Determine the rotation matrix by comparing the difference between the rotated coordinates and the decentroided coordinates of the reference four-point base. Calculate the translation vector based on the rotation matrix and the two sets of centroid coordinates. Combine the rotation matrix and translation vector into a rigid body transformation and register a time label and topology type mark for the rigid body transformation. A set of train reference points is selected in the orbital coordinate system. Each rigid body transformation is applied to the set of train reference points to obtain the spatial position and attitude of the set of train reference points after the rigid body transformation. The position and attitude of each point in the set of train reference points are compared according to the position range and attitude range in the safe reachable domain. When the position and attitude of all train reference points fall within the coverage of the safe reachable domain, the corresponding rigid body transformation is registered as a rigid body transformation candidate and added to the rigid body transformation candidate set.

[0029] In this embodiment, the construction of the collaborative scoring function specifically includes: In the orbital coordinate system, each rigid body transformation is read from the rigid body transformation candidate set. Each rigid body transformation is applied to the fusion point cloud of the corresponding time label according to the time label, and a transformed fusion point cloud frame corresponding to the rigid body transformation is generated. The rigid body transformation number and time label are registered for each frame of transformed fusion point cloud. For each rigid body transformation in the candidate set of rigid body transformations, the points in the corresponding transformed fused point cloud are traversed in the orbit coordinate system. The nearest point of each point is found in the orbit topology reference point cloud through nearest neighbor search. The distance from each point to the nearest point is calculated as the point-level geometric residual. The point-level geometric residuals are assigned to the corresponding time windows according to the time label. The distance mean, distance variance and preset quantile are calculated for the point-level geometric residuals in each time window to generate geometric residual statistics associated with the rigid body transformation. A set of train reference points is established in the track coordinate system. The points in the set of train reference points cover the front position, rear position, and key positions of the train body. For each rigid body transformation in the candidate set of rigid body transformations, the rigid body transformation is applied to the set of train reference points to obtain the spatial position of the set of train reference points after the rigid body transformation. For each point in the set of train reference points, the shortest distance to the boundary of the safe reachable domain is calculated according to the definition of the boundary of the safe reachable domain. The minimum value is selected from all the distance values ​​of the train reference points and recorded as the safety margin corresponding to the rigid body transformation. A collaborative scoring function is constructed, with geometric residual statistics and safety margin as inputs. For each rigid body transformation in the candidate set, a penalty term is set based on the distance mean, distance variance, and preset quantile in the geometric residual statistics, and a reward term is set based on the safety margin. The collaborative score value is obtained by combining the penalty and reward terms. The collaborative score value increases as the geometric residual statistics decrease and as the safety margin increases. A collaborative score value is registered for each rigid body transformation in the candidate set.

[0030] This invention applies each rigid body transformation candidate to the fused point cloud in the track coordinate system to generate a corresponding transformed point cloud frame. Then, it calculates the point-level geometric residual based on the track topology reference points and statistically analyzes the average, variance, and quantiles over time windows. This enables the error assessment to have finer-grained temporal consistency. Simultaneously, it calculates the minimum safety margin based on the train reference point set, combining track geometric constraints and the actual train operating boundary into the same evaluation system. This makes the registration quality no longer dependent on a single distance index. By constructing a collaborative scoring function by integrating residual penalty terms and safety margin reward terms, the algorithm can automatically select the optimal transformation from candidate solutions with smaller errors and larger safety margins. This significantly enhances the stability and reliability of pose assessment and reduces the risk of misjudgment in turnout areas and curve sections.

[0031] In this embodiment, the generation of the adaptive weight field for residual risk specifically includes: The train onboard control unit reads the coordination score of each rigid body transformation from the candidate set of rigid body transformations, sorts the coordination score values, determines the rigid body transformation with the largest coordination score, registers the train pose corresponding to the selected rigid body transformation as the train registration pose of the current time window in the track coordinate system, and establishes a correspondence between the train registration pose of the current time window and the time window identifier. The train onboard control unit applies the train registration pose of the current time window to the fused point cloud of the current time window in the track coordinate system to obtain the fused point cloud after the transformation of the current time window. In the track coordinate system, a nearest neighbor search is performed between each point in the fused point cloud after the transformation of the current time window and the track topology reference point cloud to obtain the nearest neighbor with the smallest distance and calculate the distance between the two points. The distance is recorded as a point-level geometric residual. The track coordinate system is divided into spatial grids, and each point-level geometric residual is assigned to the corresponding spatial grid according to the spatial grid in which the point is located. The average value and preset quantile value are calculated for the point-level geometric residual in each spatial grid to form a residual field indexed by the spatial grid. The train's onboard control unit combines the residual field with the preset risk level configuration, assigns higher weights to spatial grids with larger residual values ​​and lower weights to spatial grids with smaller residual values, constructs an adaptive weight field for residual risk in the track coordinate system, registers weight values ​​for each spatial grid, and establishes a correspondence between the adaptive weight field for residual risk and the time window identifier. In the next time window, the train's onboard control unit adjusts the sampling probability of the spatiotemporal four-point basis in the spatiotemporal four-point basis set according to the weight values ​​of the spatial grid in the residual risk adaptive weight field. It increases the sampling probability of the spatiotemporal four-point basis in the spatial grid with higher weight values. In the collaborative scoring function, it sets the weight coefficients of the geometric residual related parameters according to the weight values ​​corresponding to the spatial grid. It generates a spatiotemporal four-point basis set sampling scheme and collaborative scoring parameter configuration for the next time window, and provides the sampling scheme and parameter configuration to the improved Super4PCS calculation process for the next time window.

[0032] In this embodiment, the generation of the environmental perception results and security protection instructions specifically includes: In the orbital coordinate system, the train registration poses are read in the order of time tags. The train registration poses corresponding to each time tag are connected in chronological order to form a train registration pose sequence. Time tags and mileage indexes are registered for each train registration pose in the train registration pose sequence. In the track coordinate system, the spatial position of each train registration pose in the train registration pose sequence is projected onto the track centerline direction to obtain the mileage direction position sequence. At the same time, the offset of each train registration pose in the lateral direction relative to the track centerline is calculated. The mileage direction position sequence and the lateral offset are combined to form the train motion trajectory, where each time label in the train motion trajectory corresponds to the mileage direction position and the lateral offset. In the track coordinate system, the train trajectory is superimposed on the safe reachability domain. The current position coordinates of the train are read for each time tag. The shortest distance between the current position of the train and the boundary of the safe reachability domain is calculated according to the boundary position defined by the safe reachability domain. The shortest distance is recorded as the trajectory safety margin of the corresponding time tag and arranged in the order of the time tags to form a trajectory safety margin sequence. In the track coordinate system, the train trajectory, track safety margin sequence, track topology reference point cloud, and residual risk adaptive weight field are input into the environment analysis module. The track is divided into sections along the track centerline in the mileage direction. Within each track section, the track occupancy status is determined by combining the track topology reference point cloud. The degree of lateral offset of the train is calculated by combining the lateral offset in the train trajectory. The obstacle risk level is determined by combining the spatial grid weight values ​​in the residual risk adaptive weight field with the distance values ​​in the track safety margin sequence. The track occupancy status, lateral offset, and obstacle risk level are output according to the time label to form the environmental perception result. The train's onboard control unit generates safety protection commands according to preset control rules based on environmental perception results and trajectory safety margin sequence. Within the time tag interval where any one of the three conditions is met—track occupancy status, obstacle risk level exceeding a preset threshold, or trajectory safety margin below a preset threshold—speed control commands, braking control commands, and alarm commands are generated according to preset priorities, and the safety protection commands are sent to the train control system and shunting operation control system for execution.

[0033] Example 1: To verify the feasibility of this invention in a real-world scenario, it was deployed in a marshalling yard with a large number of tracks, dense turnouts, and frequent shunting operations. Continuous tests were conducted on the locomotive's environmental perception performance in curved tracks, intersecting turnouts, and narrow sections. In such scenarios, traditional solutions typically use two-dimensional lidar in conjunction with ICP point cloud registration and rely on ground track circuits to provide section occupancy information. At the same time, the locomotive position is estimated using an odometer. Due to obstructions from trackside equipment, frequent changes in car body attitude, and limited overlapping areas of point clouds, drift, mismatch, and registration failures often occur in turnout areas and curved sections. The spatial relationship between obstacles and track structures is also difficult to accurately determine, requiring shunting drivers to still make manual observations and judgments, thus bringing high safety pressure.

[0034] During the testing process of this invention, a binocular vision sensor and a radar sensor are installed at the front end of the locomotive. Through time synchronization and calibration, image and distance data are uniformly converted to the track coordinate system, continuously generating a fused point cloud sequence. Based on this, the track design data of the marshalling yard is imported to construct a track topology reference point cloud including the track centerline, track branches, and turnout positions. Combined with the locomotive's dynamic parameters, the reachable pose range within the predicted time window is calculated, forming a safe reachable domain consistent with the locomotive's actual motion capability. The train contour points and track adjacent points in the fused point cloud are then compared with the track topology reference point cloud. The predicted trajectory points obtained by sampling according to the safe reach domain together constitute a spatiotemporal four-point basis set, which provides input for the improved Super4PCS. Through distance ratio constraints and track order relationship constraints, the improved Super4PCS filters out candidate point combinations that are more structurally reliable in the topological reference point cloud, and removes rigid body transformations that do not conform to the train's safe reach domain. The system then calculates the geometric residuals and safety margins, uses the collaborative scoring function to select the most reasonable rigid body transformation as the locomotive pose, and constructs an adaptive weight field for residual risk, so that the sampling strategy automatically biases towards high-risk areas.

[0035] During shunting, locomotives need to traverse straight tracks, curved sections, and multiple sets of switches, encountering interference from parked materials, work vehicles, and inspection personnel along the way. Test results show that the lateral offset error of this invention is approximately 0.06m in straight sections, approximately 0.09m in curved sections, and approximately 0.13m in the switch area; the longitudinal mileage error is approximately 0.10m. Compared with the traditional radar + ICP solution, the lateral offset error is reduced by 60% to 70%, and registration failure in the switch area is not observed. Furthermore, the safety margin calculation is stable, without misjudging dangerous distances of less than 0.40m as safe distances. In contrast, the traditional solution exhibited three similar misjudgments in the same environment. The system also demonstrates greater stability in obstacle risk classification, with an accuracy rate improved to approximately 95%, as shown in Table 1.

[0036] Table 1. Comparison of Intelligent Sensing Performance in Shunting Environment

[0037] These data demonstrate that the present invention exhibits higher pose estimation accuracy and more stable environmental understanding in complex track structures and multi-interference scenarios. It integrates point cloud, track topology constraints, dynamic safe reach domain, improved Super4PCS, and risk adaptive mechanism to form a closed-loop perception framework, enabling the system to obtain continuous, reliable, and quantifiable environmental information. This provides evidence-based input for speed control and braking decisions in shunting operations, significantly improving the problems of large positioning drift, unstable matching in turnout areas, and inaccurate risk assessment in traditional schemes.

[0038] 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 visual and radar-based intelligent perception method for shunting operation environment, characterized in that, The method comprises the following steps: The front-end binocular vision sensor and radar sensor of the train collect image and distance data, and generate fused point cloud through calibration and coordinate conversion; According to the track design data, the track topology reference point cloud is generated, the reachable pose is calculated according to the train dynamic parameters, and the safe reachable domain is constructed; The train contour points and track adjacent points are extracted from the fused point cloud, the predicted trajectory points are selected from the track topology reference point cloud according to the safe reachable domain, and the space-time four-point base set is combined; Based on the space-time four-point base set, the improved Super4PCS is called to search for the space-time four-point set that meets the distance ratio and track sequence relationship in the track topology reference point cloud, and solve the rigid transformation candidate that meets the safe reachable domain constraint; The rigid transformation candidate is applied to the fused point cloud, the geometric residual and the safety margin are calculated, and the cooperative scoring function is constructed according to the geometric residual and the safety margin; The rigid transformation with the highest cooperative score is selected as the train registration pose, the residual field is updated according to the geometric residual, the residual risk adaptive weight field is generated, and is used to adjust the sampling of the space-time four-point base set and the cooperative scoring weight in the next time window; The train registration pose sequence is spliced into the train motion trajectory, and the environment perception result and the safety protection instruction are generated in combination with the safe reachable domain, the track topology reference point cloud and the residual risk adaptive weight field. 2.The visual and radar based intelligent perception method for shunting operation environment according to claim 1, characterized in that, The improved Super4PCS: The points in the fused point cloud belonging to the train contour region and the track adjacent region and the predicted trajectory points in the safe reachable domain are taken as inputs to construct a space-time four-point base set with time index; In the track topology reference point cloud, the space-time four-point set that meets the space-time conformal constraint is searched according to the consistent Euclidean distance ratio and the monotonic sequence relationship along the track center line, and the rigid transformation candidate is solved for each space-time four-point set; The rigid transformation candidates with translation vectors and attitude angles exceeding the safe reachable domain are filtered out, the remaining rigid transformation candidates are applied to the fused point cloud, the geometric residual and the minimum safety margin relative to the safe reachable domain boundary in the predicted time window are counted, the cooperative scoring function is constructed to reduce the geometric residual and increase the safety margin, the rigid transformation candidates are scored and sorted, and the rigid transformation with the highest cooperative score is selected as the train registration pose. 3.The visual and radar based intelligent perception method for shunting operation environment according to claim 1, characterized in that, The generation of the fused point cloud specifically comprises: The binocular vision sensor and the radar sensor are installed at the front end of the train, a time synchronization mode is set, the binocular vision sensor and the radar sensor collect image data and distance data in the forward direction under the unified time reference, and time tags are registered for the collected data; The binocular vision sensor is geometrically calibrated and distortion corrected, and the image data is converted into binocular point cloud representing the spatial structure of the forward scene according to the calibration result; The radar sensor is calibrated, the distance data is converted into radar point cloud representing the spatial structure of the forward scene according to the calibration parameters, and the time tag correspondence relationship with the binocular point cloud is maintained; The coordinate transformation relationship among the binocular vision sensor coordinate system, the radar sensor coordinate system and the track coordinate system is established, the binocular point cloud and the radar point cloud are converted to the track coordinate system, and the points in the binocular point cloud and the points in the radar point cloud in the track coordinate system are collected into a frame of fusion point cloud according to the time label at each sampling time, so as to form a sequence of fusion point clouds in the track coordinate system. 4.The visual and radar based intelligent perception method for shunting operation environment according to claim 1, wherein, The construction of the safety reachable domain specifically includes: reading circuit design data, the circuit design data including track center line coordinates, track number, turnout position parameters, platform edge position parameters and clearance boundary parameters; generating a track topology reference point cloud in the track coordinate system according to the track center line coordinates, the track number and the turnout position parameters, and registering a mileage value, a lateral offset and a topology type mark for each point in the track topology reference point cloud; obtaining train power parameters, the train power parameters including mass, allowable speed upper limit, traction capability parameters and braking capability parameters, and calculating a set of reachable poses in the track coordinate system according to the train power parameters and a prediction time window; determining a position range and an attitude range in the track coordinate system according to the set of reachable poses, combining the position range and the attitude range into a safety reachable domain, and registering a time label corresponding to the prediction time window in the safety reachable domain. 5.The visual and radar based intelligent perception method for shunting operation environment according to claim 1, wherein, The construction of the spatiotemporal four-point base set specifically includes: extracting points belonging to a train contour surface from the fusion point cloud as train contour points in the track coordinate system according to train external dimensions and installation positions, extracting points located near the track center line from the fusion point cloud as track adjacent points according to a lateral distance threshold between the train contour points and the track center line, and combining the train contour points and the track adjacent points into a current point set according to a time label; sampling prediction trajectory points from the track topology reference point cloud along the track center line within the prediction time window in the track coordinate system according to the track center line and the safety reachable domain, the prediction trajectory points falling within the safety reachable domain, registering a time label for the prediction trajectory points, and forming a set of prediction trajectory points; under each time label, sorting the train contour points and the track adjacent points in the current point set according to the mileage value from small to large, selecting a train contour point with the smallest mileage value and a track adjacent point with the largest mileage value as two current points, sorting the prediction trajectory points with the same time label in the set of prediction trajectory points according to the mileage value from small to large, selecting a prediction trajectory point with the smallest mileage value and a prediction trajectory point with the largest mileage value, combining the two current points and the two prediction trajectory points into a spatiotemporal four-point base, and forming a spatiotemporal four-point base set by all spatiotemporal four-point bases. 6.The visual and radar based intelligent perception method for shunting operation environment according to claim 1, wherein, The generation of the rigid transformation candidate specifically includes: in the improved Super4PCS, calculating the edge length and the edge length ratio between the four points in the track coordinate system for each spatiotemporal four-point base in the spatiotemporal four-point base set, registering a time label and a topology type mark, and forming a spatiotemporal four-point base feature; The candidate point set is selected according to a time label and a topology type label in the track topology reference point cloud, four points in the candidate point set are sequentially selected to form a reference four-point combination, the edge length and the edge length ratio of the four points in the reference four-point combination are calculated, the edge length ratio is compared with the edge length ratio in the corresponding space-time four-point base feature, when the difference falls within a preset error threshold range and the mileage values of the four points in the track center line direction satisfy a monotonic order relationship, the reference four-point combination is registered as a space-time four-point set; For each group of space-time four-point base and corresponding space-time four-point set, rigid transformation parameters including rotation parameters and displacement parameters are calculated according to the spatial coordinates of the four points in the track coordinate system, and the rotation parameters and the displacement parameters are combined into a rigid transformation; The spatial position and the attitude of the train reference point under the action of the rigid transformation are checked according to the safety reachable domain, when the spatial position and the attitude all fall within the safety reachable domain, the corresponding rigid transformation is registered as a rigid transformation candidate, and is summarized as a rigid transformation candidate set.

7. The visual and radar-based intelligent perception method for shunting operation environment according to claim 1, characterized in that, The construction of the collaborative scoring function specifically includes: The rigid transformation candidate set is read in the track coordinate system, each rigid transformation in the rigid transformation candidate set is applied to the fusion point cloud according to the time label, the transformed fusion point cloud corresponding to the rigid transformation is generated, and the time label is kept consistent with the fusion point cloud; For each rigid transformation in the rigid transformation candidate set, the geometric residual is calculated based on the corresponding transformed fusion point cloud and the track topology reference point cloud, the distance from each point in the transformed fusion point cloud to the adjacent point in the track topology reference point cloud is recorded as the point-level geometric residual, and the point-level geometric residual is counted in the range corresponding to the time label to obtain the geometric residual statistics composed of the distance average, the distance variance and the distance quantile value; For each rigid transformation in the rigid transformation candidate set, the spatial position of the train reference point after the action of the rigid transformation is obtained by applying the rigid transformation to the train reference point in the track coordinate system, the distance from the train reference point to the boundary of the safety reachable domain is calculated, and the minimum value in the distance set is defined as the safety margin corresponding to the rigid transformation; The collaborative scoring function is constructed, the geometric residual statistics and the safety margin are taken as inputs, the collaborative score of each rigid transformation in the rigid transformation candidate set is calculated, and the collaborative score is set to increase with the decrease of the geometric residual statistics and to increase with the increase of the safety margin, so that the collaborative score of each rigid transformation in the rigid transformation candidate set is obtained. 8.The visual and radar based intelligent perception method for shunting operation environment according to claim 1, wherein, The generation of the residual risk adaptive weight field specifically includes: According to the collaborative score of each rigid transformation in the rigid transformation candidate set, the rigid transformation with the highest collaborative score is selected, and the train pose corresponding to the rigid transformation with the highest collaborative score is registered as the current time window train registration pose. In the track coordinate system, the current time window train registration pose is applied to the fused point cloud to obtain a transformed fused point cloud, and a point-level geometric residual is calculated between the transformed fused point cloud and the track topology reference point cloud. The distance from each point in the transformed fused point cloud to the adjacent point in the track topology reference point cloud is recorded as the point-level geometric residual. A spatial grid is divided in the track coordinate system, and the average value and the preset quantile value of the point-level geometric residual are calculated in each spatial grid to generate a residual field. The residual field is combined with the preset risk level configuration to assign a higher weight to the area with a larger point-level geometric residual in the spatial grid and a lower weight to the area with a smaller point-level geometric residual in the spatial grid, thereby forming a residual risk adaptive weight field in the track coordinate system, and recording a weight value for each spatial grid. In the next time window, the sampling probability of the spatio-temporal four-point basis in the spatio-temporal four-point basis set is adjusted according to the residual risk adaptive weight field, and the sampling probability of the spatio-temporal four-point basis is increased in the spatial grid with a higher weight value. In the collaborative scoring function, the weight of the geometric residual related parameter is adjusted according to the residual risk adaptive weight field to generate a spatio-temporal four-point basis set sampling scheme and a collaborative scoring parameter configuration for the next time window. 9.The visual and radar based intelligent perception method for shunting operation environment according to claim 1, wherein, The generation of the environment perception result and the safety protection instruction specifically includes: In the track coordinate system, the train registration poses are read in the order of time tags, and the train registration poses corresponding to the time tags are connected to form a train registration pose sequence; The train registration pose sequence is projected onto the track center line direction in the track coordinate system to calculate the position change in the mileage direction and the offset in the transverse direction, thereby forming a train motion trajectory; The train motion trajectory is superimposed with the safety reachable domain, and the trajectory safety margin sequence is calculated according to the distance between the train position in the track coordinate system and the boundary of the safety reachable domain; The train motion trajectory, the trajectory safety margin sequence, the track topology reference point cloud, and the residual risk adaptive weight field are combined to generate an environment perception result, which includes the track occupancy state in the mileage direction, the transverse offset, and the obstacle risk classification; The safety protection instruction is generated according to the environment perception result and the trajectory safety margin sequence, and the safety protection instruction includes a speed control instruction, a brake control instruction, and an alarm instruction.

Images