A kidnapping detection and recovery method for wheeled robot

Abstract

The application discloses a kind of abductions detection and recovery method for wheeled robot, comprising: the real-time state parameters of wheeled robot are obtained using inertial measurement unit and wheel speed meter, feature extraction is carried out to it and whether robot is in the state of abduction is detected through classification model;If abduction state is detected, surrounding environment parameters are collected using laser radar and RGB-D camera, and whether robot is in the state of blockade is judged based on point cloud fusion and obstacle ratio;According to the blockade state, the adaptive positioning recovery method is selected, and whether the robot movement is limited and alarm is judged according to the recovery effect.

Description

Technical Field

[0001] This invention relates to the field of abduction detection technology for wheeled androids, and more particularly to a method for abduction detection and recovery for wheeled androids. Background Technology

[0002] Currently, wheeled automata robots have achieved rapid development in warehousing and logistics, intelligent manufacturing, and security inspection, with their actual deployment numbers increasing daily. Accurate perception of position and attitude is crucial for robots when performing autonomous navigation and handling tasks. A typical anomaly is the "kidnapping problem": during operation, the robot is moved to different unknown locations by human intervention or accident, causing a significant deviation between its internal positioning estimate and its actual physical location. Failure to detect and restore the kidnapping state in a timely and accurate manner will directly affect the robot's operational safety and even pose a potential threat to the surrounding environment and personnel.

[0003] Current mainstream methods for robot abduction detection and recovery typically rely on sensors such as LiDAR and vision sensors on mobile robots. There are two main technical approaches: one is map-matching-based abduction detection and recovery, which utilizes SLAM (Simultaneous Localization and Mapping) technology to match environmental features collected by LiDAR or vision cameras with a pre-built map. After an abduction event occurs, if the map-matching algorithm detects a significant discrepancy between the perceived environmental data and the robot's own position information (such as a surge in observation residuals), it is considered an abduction. The other approach is based on sensor data changes. Some studies compare the robot's own motion state with external observations (such as wheel speedometers and IMUs) to analyze attitude and displacement changes under sensor fusion. If the wheel speedometer data shows the vehicle is not driven, has abnormal inertial acceleration, but the perceived position changes drastically, it is considered a possible abduction event. Detection relies on kinematic models and multi-source sensor data verification.

[0004] However, all of the above technical solutions have certain problems:

[0005] First, map matching methods heavily rely on the quality of LiDAR and visual features. In dynamic environments, such as densely populated areas, moving goods, or scenes with drastic changes in lighting, map features can change, leading to misjudgments by matching algorithms: dynamic occlusion or loss of feature points can misclassify a normal scene as a kidnapping event, or conversely, it may fail to detect the true kidnapping situation in a timely manner.

[0006] Secondly, current methods often use a single threshold of positioning error or observation residual to determine kidnapping events, neglecting the data correlation of robot kinematics itself. For example, data from wheel speedometers and IMUs are not fully integrated, and relying solely on map errors may mistakenly identify temporary occlusions or anomalies as kidnapping, or make inaccurate judgments when actual kidnapping actions (such as being carried) occur because abnormal driving forces or accelerations are not detected.

[0007] Third, after abduction detection is triggered, existing recovery operations typically employ global particle filtering or large-scale scene feature rematching, which consumes significant computational resources and time, making them unsuitable for real-time response. This is especially problematic in dynamic, high-frequency abduction and recovery scenarios, easily causing system latency and hindering the robot's rapid return to work.

[0008] Fourth, when a robot encounters an obstacle (such as being trapped by surrounding obstacles), existing methods are insufficient in recognizing such local spatial anomalies. They can only make simple judgments based on their own positioning errors, leading to an increased false alarm rate. Furthermore, the recovery positioning process is inefficient and lacks adaptive window adjustment and state tracking mechanisms for special scenarios.

[0009] The purpose of this invention is to design a kidnapping detection and recovery method for wheeled androids that addresses the problems of the existing technology. Summary of the Invention

[0010] In view of this, the purpose of this invention is to propose a kidnapping detection and recovery method for wheeled androids, which can solve the above-mentioned problems.

[0011] This invention provides a method for abduction detection and recovery of wheeled androids, comprising:

[0012] The real-time state parameters of the wheeled android are obtained by using an inertial measurement unit and wheel speedometer, and features are extracted from them. The robot is then classified using a classification model to detect whether it is in a kidnapping state.

[0013] If a kidnapping situation is detected, the robot will use LiDAR and RGB-D camera to collect environmental parameters and determine whether the robot is in a blocking situation based on point cloud fusion and obstacle ratio.

[0014] Select an adaptive positioning recovery method based on the containment status, and determine whether to restrict robot movement and issue an alarm based on the recovery effect.

[0015] The beneficial effects of this invention are:

[0016] First, it integrates inertial measurement units and wheel speedometers for data alignment, noise reduction, and multi-dimensional feature construction, and uses a classification model for intelligent discrimination. This solves the problem that traditional robots with single sensors struggle to identify abnormalities in a timely and accurate manner when encountering external forces such as kidnapping, lifting, or dragging; it also overcomes the interference of common noises such as slippage and vibration on anomaly detection, preventing situations where the robot is mistakenly believed to be safe after being moved.

[0017] Secondly, by using spatial layering statistics of laser / point cloud data and identifying obstacle structure types (continuous and dispersed), combined with environmental adaptive thresholds and criteria, the system dynamically tracks changes in obstacle enclosure. This effectively solves the problem that robots cannot accurately distinguish between encirclement and ordinary obstacle stacks in real 3D environments; it overcomes the problems of poor environmental adaptability and misjudgment of dispersed / temporary obstacles caused by uniform thresholds, and avoids excessive triggering of alarms / shutdowns when there is no actual threat.

[0018] Third, based on the state determination results (non-blockade / cleared), a multi-level closed-loop control recovery strategy is adopted, employing a small-window rapid recovery and a large-window global relocation in stages. This strategy integrates IMU / wheel velocity measurement, point cloud feature map matching, and confidence level determination. It solves the problems of difficult accurate positioning and recovery failure caused by short-term loss of positioning or extreme displacement (such as dragging, moving, or escaping after being blocked), and avoids continued loss of control or false recovery due to blind attempts. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings required in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of the method in this embodiment. Detailed Implementation

[0021] To facilitate understanding by those skilled in the art, the structure of the present invention will now be described in further detail with reference to the accompanying drawings. It should be understood that, unless otherwise specified, the order of the steps mentioned in this embodiment can be adjusted according to actual needs, and they can even be executed simultaneously or partially simultaneously.

[0022] like Figure 1 As shown, this embodiment of the invention provides a method for abduction detection and recovery of wheeled androids, including:

[0023] S1 uses an inertial measurement unit and wheel speedometer to obtain real-time state parameters of the wheeled android, extracts features from them, and uses a classification model to detect whether the robot is in a kidnapping state.

[0024] S101 uses an inertial measurement unit to obtain the three-axis acceleration and three-axis angular velocity of the wheeled robot, and uses a wheel speed meter to obtain the rotational speed of the left and right wheels of the wheeled robot.

[0025] S102 performs time alignment and noise reduction on the triaxial acceleration, triaxial angular velocity, and left and right wheel speeds, and then extracts features to construct a feature vector.

[0026] S1021 divides the triaxial acceleration, triaxial angular velocity, and left and right wheel speeds into sliding windows of preset duration;

[0027] S1022 calculates the mean, variance, and extreme value difference of the three-axis acceleration, three-axis angular velocity, and left and right wheel velocities within each window, and splices them to form a 24-dimensional feature vector.

[0028] S103 inputs the feature vector into the kidnapping state discrimination model to obtain the kidnapping state result.

[0029] In this step, the inertial measurement unit (IMU) can obtain three-axis acceleration and three-axis angular velocity in real time, comprehensively reflecting the robot's kinematics, attitude changes, and force characteristics. It is extremely sensitive to external disturbances (such as lifting, rotation, and moving). During abduction, the IMU can detect abnormal motion trajectories (e.g., sudden loss of wheel speed while acceleration / angular velocity changes). Even if the wheels are suspended, slipping, or stuck, the IMU can still collect the actual motion changes, demonstrating good independence. The wheel speedometer reflects the actual motor driving force and speed of the left and right wheels. If the wheels are not rotating, it indicates no active forward movement, but if the robot body is moving, it is highly suspicious.

[0030] Using only an IMU (Instrument Detector) is prone to false alarms due to mechanical vibration and electromagnetic interference, while using only a wheel speedometer is prone to missed alarms due to slippery ground or sensor malfunction. Fusion technology allows for mutual verification, improving the tolerance for false and missed detections. The most direct physical characteristic of kidnapping is that the wheels remain stationary while the robot body moves. This requires comparing data from both the IMU (detecting the robot body) and the wheel speedometer (detecting the actuators); both sets of data are indispensable. Kidnapping events often manifest as sudden deceleration (or zero speed), drastic posture changes, and no obvious correlation in acceleration. A single feature may lead to misjudgment; multi-level statistical combinations can more accurately capture short-term and structural motion anomalies.

[0031] If S2 detects a kidnapping situation, it uses LiDAR and RGB-D camera to collect surrounding environmental parameters and determines whether the robot is in a blocking situation based on point cloud fusion and obstacle ratio.

[0032] S201 uses LiDAR to acquire radar point cloud data around the wheeled robot and uses an RGB-D camera to acquire color and depth images around the wheeled robot to generate a three-dimensional visual point cloud.

[0033] S202 fuses radar point cloud data and 3D visual point cloud to form a unified spatial fused point cloud, and performs voxel filtering downsampling on the fused point cloud;

[0034] S203 performs three-dimensional spatial layering of containment discrimination on the wheeled android based on the fused point cloud, and obtains the containment discrimination result;

[0035] S2031 sets up a nearest neighbor detection area with the wheeled robot as the center, and divides the point cloud data in the area into bottom layer, middle layer and top layer according to height. The proportion of obstacle points in each layer to the total number of point cloud points is calculated to obtain the layer obstacle proportion vector.

[0036] In this step, most obstacles encountered by the robot are not confined to a single plane. For example, there are low objects on the ground, medium-to-high obstacles such as tabletops or windowsills, and high-level obstacles such as hanging objects or beams. In real-world environments, obstacles are located at different heights and have drastically different effects on the robot's movement. Low obstacles (such as wires, thresholds, and mats) do not impede robot movement but can affect specific sensors (e.g., only near-ground lasers, but wheels can run over them). Mid-level obstacles (such as walls, furniture, and a person's lower legs) are the main obstacles hindering movement and require focused identification. High-level obstacles (such as beams, curtains, and hanging decorations) do not affect the mobile chassis but are crucial for stereo perception or the location of high-positioned LiDAR. Therefore, dividing the point cloud data within the area into three layers based on height—bottom layer (0-0.5m), middle layer (0.5-1.5m), and top layer (above 1.5m)—solves the problem of misjudgment and false alarms that easily occur in two-dimensional obstacle statistics when the robot faces undulating and complex three-dimensional environments. Layered identification can accurately determine whether an obstacle truly affects movement safety, providing a basis for subsequent threshold selection.

[0037] S2032 identifies obstacle types based on the layered obstacle proportion vector combined with spatial distribution characteristics. The obstacle types include: continuous structures and dispersed structures.

[0038] In this step, the spatial distribution of obstacles is analyzed to distinguish between continuous structures (such as walls and shelves, which have continuous interception capabilities) and dispersed structures (such as isolated objects and pedestrians). Continuous structures generally form real blockades, while dispersed obstacles can usually be bypassed. This step avoids the robot from falsely stopping or reporting blockades due to occasional, localized, or randomly appearing obstacles (such as temporarily moving objects). It improves the robot's ability to distinguish between substantial blockades and temporary, bypassable obstacles, significantly reducing the false alarm rate.

[0039] S2033 adaptively sets the obstacle ratio determination threshold for each spatial layer based on the current environment type of the wheeled android.

[0040] This step introduces environmental type recognition (such as corridors, open areas, shelving areas, etc.) to automatically adjust the warning thresholds for the proportion of obstacles on each level. This addresses the issues of false containment and delayed false alarms caused by differences in environmental width under a uniform threshold.

[0041] S2034 If the proportion of obstacles in any layer continuously exceeds the adaptive obstacle proportion determination threshold, and the obstacle type is a continuous structure, and the proportion continues to exceed the preset time threshold, then it is determined to be a blockade state; otherwise, it is a non-blockade state.

[0042] In this step, a triple-combination condition is used, combining temporal statistics (e.g., N consecutive frames or t cumulative seconds) with type discrimination and proportion threshold, to suppress false alarms caused by instantaneous anomalies or transient dynamic objects. This effectively eliminates false alarms caused by short-term interference, focusing on persistent, substantial physical containment. It also enhances the system's stability in environments with dynamic pedestrian flow and numerous moving obstacles.

[0043] S3 selects an adaptive positioning recovery method based on the containment status, and determines whether to restrict robot movement and issue an alarm based on the recovery effect.

[0044] If the wheeled robot is currently in a non-blocking state, the small window positioning and recovery strategy is adopted.

[0045] In this step, only when the robot loses its location or malfunctions in a non-blocking state will it first attempt to quickly restore its location within a small local window area. Only when the local information is insufficient to meet the restoration accuracy will it switch to a global large window strategy, balancing efficiency and robustness.

[0046] S3011 predicts the current position of the wheeled robot using an inertial measurement unit and wheel speedometer, and delineates a small window area with the predicted position as the center and according to a set first radius.

[0047] S3012 extracts the environmental features of the fused point cloud, matches them with the pre-established map environmental features of the small window area, and determines whether the current matching confidence is higher than the confidence threshold.

[0048] S3013 If the current matching confidence is higher than the first confidence threshold, then the current position and orientation of the wheeled robot are corrected and updated according to the optimal registration solution of the fused point cloud features and small window area map features, and the localization recovery is completed.

[0049] S3014 If the current matching confidence level is lower than the first confidence level threshold, switch to the large window global relocation strategy.

[0050] In this step, the short-term drift range is usually limited, and the accurate position can be found with a high probability within a small local window, avoiding the high computational cost and low efficiency of the entire map. Local feature matching is efficient with few mismatches, and the environmental features within the small window are relatively stable, resulting in more accurate registration. If the matching confidence is higher than a set threshold, it indicates that the registration in this area is reliable, and the point cloud-map registration result can be used as a new position estimate for the robot to correct the localization state and update the pose. Only when the confidence meets the threshold is the position adopted to minimize the risk of incorrect recovery and erroneous actions. If the local matching confidence in the small window is low, it indicates that the prediction error has accumulated significantly or the environment has changed greatly, and the local recovery is unreliable. At this time, we actively switch to global relocalization and search for the optimal registration pose again on the entire map.

[0051] S302 If the current wheeled robot is in a blocked state, a large window repositioning recovery strategy is adopted;

[0052] The S3021 records the last known pose of the wheeled android before it enters the containment state, and continuously predicts the pose change using the inertial measurement unit and wheel speedometer during the containment and kidnapping state, updating the cumulative motion trajectory data in real time.

[0053] S3022 After the containment and kidnapping states are lifted, the predicted pose of the current position is calculated based on the last known pose before containment, combined with the predicted pose changes during the period, and the large window global relocation area is delineated with the predicted pose as the center according to the set second radius.

[0054] S3023 collects and extracts the environmental features of the fused point cloud at the current moment, matches them with the pre-established map environmental features within the large window area, and determines whether the current matching confidence is higher than the set second confidence threshold.

[0055] S3024 If the matching confidence is higher than the second confidence threshold, then based on the optimal registration solution of the fused point cloud features and the large window area map features, the current position and orientation of the wheeled robot are corrected and updated to complete the global localization recovery.

[0056] If the matching confidence level is lower than the second confidence level threshold, S3025 will trigger a measure to prohibit the movement of the wheeled robot and report the anomaly to the server.

[0057] In this step, the last high-confidence localization result before the blockade / kidnapping occurs serves as the recovery baseline, minimizing uncertainties during the loss of localization. The last known pose of the wheeled android before entering the blockade state includes its current position and orientation. When encountering a blockade / kidnapping, the robot may drift significantly or even be forcibly moved; the small window strategy may fail. Therefore, a larger global search window (the second radius is larger than the first radius, typically covering hundreds of square meters) needs to be defined around the last position based on the predictable range of change. Within this large window, global spatial features of the current point cloud and each sub-region within the map are extracted. Optimal registration is found through high-order feature matching, and only when the confidence level is sufficiently high is it used as the basis for pose recovery. If a feature registration result with sufficient confidence cannot be obtained for all candidate sub-regions within the current large window, it indicates severe loss of localization information or environmental confusion; hasty recovery may lead to a catastrophic navigation accident.

[0058] If both the kidnapping and containment states have ended and the location recovery attempt has been completed, then S303 will perform a relocation matching degree determination and anomaly handling.

[0059] S3031 calculates the matching degree between the current fused point cloud features acquired by the lidar and camera and the map features after relocalization, wherein the matching degree is the ratio of the number of matching feature points to the total number of feature points;

[0060] S3032 determines whether the matching degree is lower than a preset matching degree ratio threshold, and whether the number of frames that the low matching degree state lasts exceeds a preset frame number threshold, and triggers an exception handling mechanism based on the determination result.

[0061] If the matching degree is lower than the matching degree ratio threshold and the number of consecutive frames does not exceed the frame number threshold, it is judged as a brief abnormality, and the movement speed of the wheeled robot is reduced.

[0062] If the matching degree is lower than the matching degree ratio threshold and the number of consecutive frames exceeds the frame number threshold, the positioning recovery attempt is determined to have failed. The wheeled robot stops moving, a local audible and visual alarm is triggered, and the abnormal information is reported to the server.

[0063] In this step, after the robot detects that the kidnapping and containment states have been lifted and completes a localization recovery attempt, it cannot directly assume that its current pose is reliable. It is also necessary to verify the quality of the recovery effect to prevent a new round of loss of control caused by erroneous recovery. Therefore, it enters the relocalization matching degree determination stage and performs online quality detection on the recovered localization results.

[0064] Point cloud data acquired by LiDAR and cameras is matched with corresponding regional features in a local high-precision map. The matching degree is calculated as the number of effectively registered feature points divided by the total number of feature points to be matched. The matching degree reflects the consistency between the robot's current position and the known environment model. It can eliminate false recoveries caused by poor environmental repeatability, delayed map updates, or local dynamic objects.

[0065] It's not enough to just check the matching degree of a single frame; the duration of the anomaly must also be observed. By setting a frame threshold, short-term interference (such as momentary occlusion, reflection, etc.) can be distinguished from persistent anomalies (such as being trapped in a new environment / significant map changes). A serious anomaly is only identified when the matching degree continuously falls below the threshold and the number of consecutive frames reaches the set value; otherwise, it is treated as a brief anomaly. This solves the problems of false stops / false alarms caused by occasional interference and the robot's susceptibility to harmless minor anomalies. Engineering scenarios are complex, with frequent short-term occlusions, reflections, and people passing by. Simply stopping the robot for all of these would significantly impact work efficiency and user experience. Proactive slowdown allows the robot more time to recover its localization, reducing the probability of blind collisions and providing more frames for the localization algorithm to recover. Classifying anomalies by duration ensures safety while significantly reducing false stops.

[0066] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0067] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0068] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0069] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0070] It should be noted that any reference signs placed between parentheses in the claims should not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The invention can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In a unit claim enumerating several means, several of these means may be embodied by the same item of hardware. The words first, second, and third, etc., do not indicate any order. These words can be interpreted as names.

[0071] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0072] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

[0073] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0074] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms should not be construed as necessarily referring to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

Claims

1. A method for abduction detection and recovery of wheeled androids, characterized in that, include: The real-time state parameters of the wheeled android are obtained by using an inertial measurement unit and wheel speedometer, and features are extracted from them. The robot is then classified using a classification model to detect whether it is in a kidnapping state. If a kidnapping situation is detected, the robot will use LiDAR and RGB-D camera to collect environmental parameters and determine whether the robot is in a blocking situation based on point cloud fusion and obstacle ratio. The adaptive positioning recovery method is selected based on the containment status, and the robot's movement is restricted and an alarm is triggered based on the recovery effect, including: If the wheeled robot is currently in a non-blocking state, a small window positioning and recovery strategy is adopted; If the wheeled robot is currently in a blocked state, a large window repositioning and recovery strategy will be adopted. If both the kidnapping and containment states have ended, and the location recovery attempt has been completed, then the relocation matching degree determination and anomaly handling will be performed.

2. The method for abduction detection and recovery of a wheeled android according to claim 1, characterized in that, The process of acquiring real-time state parameters of the wheeled automaton robot using an inertial measurement unit and wheel speedometer, extracting features from these parameters, and detecting whether the robot is in a kidnapping state using a classification model includes: The three-axis acceleration and three-axis angular velocity of the wheeled robot are obtained using an inertial measurement unit, and the rotational speeds of the left and right wheels of the wheeled robot are obtained using a wheel speed meter. After time alignment and noise reduction of the triaxial acceleration, triaxial angular velocity, and left and right wheel speeds, feature vectors are extracted and constructed. The feature vector is input into the kidnapping state discrimination model to obtain the kidnapping state result.

3. The method for abduction detection and recovery of a wheeled android according to claim 2, characterized in that, The step of time-aligning and denoising the triaxial acceleration, triaxial angular velocity, and left and right wheel speeds, followed by feature extraction to construct a feature vector includes: The triaxial acceleration, triaxial angular velocity, and left and right wheel speeds are divided into sliding windows of preset duration. Calculate the mean, variance, and extreme value difference of the three-axis acceleration, three-axis angular velocity, and left and right wheel velocities within each window, and then concatenate them to form a multi-dimensional feature vector.

4. The method for abduction detection and recovery of a wheeled android according to claim 1, characterized in that, If a kidnapping situation is detected, the robot will use LiDAR and RGB-D cameras to collect surrounding environmental parameters, and determine whether the robot is in a containment state based on point cloud fusion and obstacle ratio, including: The radar point cloud data around the wheeled robot is obtained using LiDAR, and the color and depth images around the wheeled robot are obtained using an RGB-D camera to generate a three-dimensional visual point cloud. Radar point cloud data and 3D visual point cloud are fused to form a unified spatial fused point cloud, and voxel filtering downsampling is performed on the fused point cloud. The confinement judgment of the wheeled android is performed in three-dimensional space based on the fused point cloud, and the confinement judgment result is obtained.

5. A method for abduction detection and recovery of a wheeled android according to claim 4, characterized in that, The method of performing three-dimensional spatial layering and containment discrimination on the wheeled android based on the fused point cloud, and obtaining containment discrimination results, includes: Centered on the wheeled android, a nearest neighbor detection area is set up, and the point cloud data in the area is divided into bottom layer, middle layer and top layer according to height. The proportion of obstacle points in each layer to the total number of point cloud points is calculated to obtain the layer obstacle proportion vector. Based on the hierarchical obstacle proportion vector combined with spatial distribution characteristics, obstacle types are identified, including continuous structures and dispersed structures. The threshold for determining the proportion of obstacles in each spatial layer is adaptively set according to the current environment type in which the wheeled android is located. If the proportion of obstacles in any layer continuously exceeds the adaptive obstacle proportion judgment threshold, and the obstacle type is a continuous structure, and the proportion continues to exceed the preset time threshold, then it is judged as a blockade state; otherwise, it is a non-blockade state.

6. The method for abduction detection and recovery of a wheeled android according to claim 1, characterized in that, If the wheeled robot is currently in a non-blocking state, the small window positioning recovery strategy includes: The current position of the wheeled robot is predicted by the inertial measurement unit and wheel speedometer, and a small window area is delineated with the predicted position as the center according to the set first radius. Extract the environmental features of the fused point cloud and match them with the pre-established map environmental features of the small window area to determine whether the current matching confidence is higher than the confidence threshold. If the current matching confidence is higher than the first confidence threshold, the current position and orientation of the wheeled robot are corrected and updated according to the optimal registration solution of the fused point cloud features and small window area map features, and the localization recovery is completed. If the current matching confidence level is lower than the first confidence level threshold, then switch to the large window global relocation strategy.

7. The method for abduction detection and recovery of a wheeled android according to claim 1, characterized in that, If the wheeled robot is currently in a blocked state, the large-window relocation and recovery strategy includes: Record the last known pose of the wheeled android before it enters the containment state, and continuously predict the pose change using an inertial measurement unit and wheel speedometer during the containment and kidnapping state, and update the cumulative motion trajectory data in real time. Once the containment and kidnapping states are lifted, the predicted pose of the current position is calculated based on the last known pose before containment, combined with the predicted pose changes during the period. The large window global relocation area is then delineated with the predicted pose as the center and according to the set second radius. Collect and extract the environmental features of the fused point cloud at the current moment, match them with the pre-established map environmental features within the large window area, and determine whether the current matching confidence is higher than the set second confidence threshold. If the matching confidence is higher than the second confidence threshold, the current position and orientation of the wheeled robot are corrected and updated based on the optimal registration solution of the fused point cloud features and the large window area map features, thus completing the global localization recovery. If the matching confidence level is lower than the second confidence level threshold, a measure to prohibit the movement of the wheeled robot will be triggered and the anomaly will be reported to the server.

8. The method for abduction detection and recovery of a wheeled android according to claim 1, characterized in that, If both the kidnapping and containment states have ended, and the location recovery attempt has been completed, then the relocation matching degree determination and anomaly handling include: Calculate the matching degree between the current fused point cloud features acquired by the lidar and camera and the map features after relocalization. The matching degree is the ratio of the number of matching feature points to the total number of feature points. Determine whether the matching degree is lower than a preset matching degree ratio threshold, and whether the number of frames in which the low matching degree state continues exceeds a preset frame number threshold. Trigger an exception handling mechanism based on the determination result.

9. A method for abduction detection and recovery of a wheeled android according to claim 8, characterized in that, The step of determining whether the matching degree is lower than a preset matching degree ratio threshold and whether the number of frames in which the low matching degree state persists exceeds a preset frame threshold, and triggering an exception handling mechanism based on the determination result, includes: If the matching degree is lower than the matching degree ratio threshold and the number of consecutive frames does not exceed the frame number threshold, it is judged as a brief abnormality, and the movement speed of the wheeled robot is reduced. If the matching degree is lower than the matching degree ratio threshold and the number of consecutive frames exceeds the frame number threshold, the positioning recovery attempt is determined to have failed. The wheeled robot stops moving, a local audible and visual alarm is triggered, and the abnormal information is reported to the server.

Images