Intelligent wheelchair anti-tipping control method and intelligent wheelchair

Abstract

The application provides a smart wheelchair anti-tipping control method and a smart wheelchair. The method comprises: collecting multi-source sensor data related to the stability of the wheelchair center of gravity, wherein the multi-source sensor data comprises environment data, wheelchair posture data and user action data; performing space-time synchronization and feature extraction on the multi-source sensor data to obtain a plurality of single-mode feature data; using a cross-modal attention dynamic weight fusion method, performing feature fusion and decoding on the plurality of single-mode feature data to obtain an anti-tipping perception result of the wheelchair; and based on the anti-tipping perception result, performing a multi-level anti-tipping action. The method of the application realizes accurate anti-tipping perception by collecting multi-source sensor data related to the stability of the wheelchair center of gravity, performs targeted actions according to different situations of the center of gravity deviation, effectively reduces the tipping risk, and improves the safety and stability of the wheelchair use.

Description

Technical Field

[0001] This application relates to the field of intelligent wheelchair technology, and in particular to an intelligent wheelchair anti-tipping control method and an intelligent wheelchair. Background Technology

[0002] With the aging population and increasing travel needs of people with disabilities, smart wheelchairs, as assistive mobility devices, have been widely used in various scenarios such as homes, nursing homes, and rehabilitation institutions. Their safety and stability are directly related to the personal safety of users. In recent years, smart wheelchairs with seat height adjustment, rotation, and fitness functions have gradually become more popular, meeting the needs of users in various scenarios such as posture adjustment and rehabilitation training. However, during the use of such wheelchairs, situations such as seat height adjustment and rotation, user exertion during fitness exercises, and driving on slopes or slippery surfaces can cause sudden changes in the center of gravity, which can easily lead to tipping or rollover accidents, threatening user safety.

[0003] Current anti-tipping technologies for smart wheelchairs mostly employ sensor-based active anti-tipping solutions. These solutions involve installing various sensors on the wheelchair to collect posture or center-of-gravity pressure data. When a center-of-gravity shift exceeds a safety threshold, braking or posture adjustment actions are triggered to actively prevent tipping. However, this active anti-tipping solution lacks effective synchronization and fusion processing of the collected multi-source data. It cannot dynamically adjust based on the validity of data in different scenarios, resulting in fused feature data that fails to accurately reflect the actual stability of the wheelchair's center of gravity and makes it difficult to accurately predict center-of-gravity shift trends. Furthermore, the anti-tipping control actions are mostly single emergency braking, lacking hierarchical control logic and unable to execute targeted anti-tipping actions based on the degree and trend of center-of-gravity shift. Summary of the Invention

[0004] To address the aforementioned technical problems, this application provides an intelligent wheelchair anti-tipping control method and an intelligent wheelchair.

[0005] A method for preventing tipping over an intelligent wheelchair is designed. The method includes: collecting multi-source sensor data related to the stability of the wheelchair's center of gravity, including environmental data, wheelchair posture data, and user action data; performing spatiotemporal synchronization and feature extraction on the multi-source sensor data to obtain multiple single-mode feature data; employing a cross-modal attention dynamic weight fusion method to fuse and decode the multiple single-mode feature data to obtain the wheelchair's anti-tipping perception result; and executing multi-level anti-tipping actions based on the anti-tipping perception result.

[0006] This solution comprehensively covers factors influencing wheelchair center of gravity by collecting three types of data directly related to center of gravity stability: environment, wheelchair posture, and user actions. This avoids the bias in center of gravity judgment caused by incomplete collection of single-type data. Spatiotemporal synchronization and feature extraction solve the problems of asynchronous time and inconsistent spatial coordinate systems among multi-source data, ensuring the accuracy of single-mode feature data. A cross-modal attention-based dynamic weight fusion method overcomes the limitations of existing fixed-weight fusion, dynamically adjusting weights based on the effectiveness of each modality's data in different scenarios. This makes the decoded anti-tipping perception results more closely match the actual center of gravity state of the wheelchair, accurately capturing center of gravity shift trends. Multi-level anti-tipping actions are executed based on the perception results, replacing the existing single emergency braking control method. Targeted actions can be performed according to different center of gravity shift situations, effectively reducing the risk of tipping and improving the safety and stability of wheelchair use.

[0007] Furthermore, the spatiotemporal synchronization and feature extraction of the multi-source sensor data to obtain multiple single-mode feature data includes: aligning the collected multi-source sensor data with timestamps and mapping all data to a predetermined coordinate system of the wheelchair chassis; filtering, denoising, and normalizing the spatially aligned data to obtain environmental preprocessing data, wheelchair posture preprocessing data, and user action preprocessing data; and extracting features from the environmental preprocessing data, wheelchair posture preprocessing data, and user action preprocessing data to obtain the multiple single-mode feature data.

[0008] In this scheme, time synchronization of data from different types of sensors is achieved through timestamp alignment, avoiding delays in center of gravity status judgment caused by differences in data acquisition time. All data are uniformly mapped to the predetermined coordinate system of the wheelchair chassis, resolving the issue of inconsistent spatial references between different sensor data, ensuring spatial correlation and comparability of the data, and laying the foundation for subsequent feature extraction and fusion. Filtering and denoising effectively remove invalid data generated during sensor acquisition due to environmental interference and equipment errors, reducing the impact of noise on feature extraction and improving data purity. Normalization eliminates dimensional differences between different types of data, preventing a single type of data from dominating the feature fusion result due to excessively large numerical ranges. Feature extraction is performed on the three types of preprocessed data separately, accurately capturing the core features related to center of gravity stability in each type of data, providing high-quality single-mode feature support for subsequent cross-modal fusion, and improving the accuracy of anti-tipping sensing results.

[0009] Furthermore, the step of extracting features from the environmental preprocessing data, wheelchair posture preprocessing data, and user action preprocessing data to obtain the multiple single-mode feature data includes: using a neural network to extract features from the environmental preprocessing data, wheelchair posture preprocessing data, and user action preprocessing data to obtain the corresponding single-mode feature data.

[0010] In this solution, neural networks are used to automatically extract deep spatiotemporal features of each modality data, which can more effectively capture the nonlinear and temporal dependencies related to center of gravity stability in road conditions, wheelchair posture, and user actions, thereby further improving perception accuracy.

[0011] Furthermore, the method of employing cross-modal attention dynamic weight fusion to perform feature fusion and decoding on the multiple single-modal feature data to obtain the anti-tipping perception result of the wheelchair includes: based on a cross-modal attention fusion network, learning the correlation between the multiple single-modal feature data through multiple attention heads to determine the fusion weight of each single-modal feature data; performing weighted fusion on the multiple single-modal feature data based on the fusion weight to generate a fusion feature vector; and decoding the fusion feature vector to obtain the anti-tipping perception result; the anti-tipping perception result includes the current center of gravity state, the future center of gravity shift trend, and the dynamic safety boundary.

[0012] In this solution, multiple attention heads of a cross-modal attention fusion network can accurately learn the intrinsic relationships between various single-modal feature data, enabling dynamic adaptive adjustment of fusion weights. This solves the problem in existing technologies where fixed weights cannot adapt to changes in data validity across different scenarios. For example, when driving on a ramp, the weights of environmental data (ramp angle) and wheelchair posture data are automatically increased, while the weights of user action data are appropriately decreased, ensuring that the fusion result matches the actual scenario requirements. Based on dynamic weights, weighted fusion generates a fusion feature vector that integrates the advantages of various single-modal features, accurately reflecting the wheelchair's center of gravity stability. The decoded anti-tipping perception result not only includes the current center of gravity state but also predicts future center of gravity shift trends and dynamic safety boundaries, providing lead time for subsequent multi-level anti-tipping actions. The output of dynamic safety boundaries can dynamically adjust safety thresholds according to real-time scenarios, adapting to safety requirements under different driving speeds and user actions, thus improving the flexibility and adaptability of anti-tipping control.

[0013] Furthermore, the multiple attention heads include a set of specialized attention heads and a set of global attention heads; the specialized attention heads are used to learn the correlation between the multiple single-modality feature data; the global attention heads are used to complete the information through the other modality features when a certain modality feature data is missing or the confidence level is lower than a threshold, and automatically reduce the fusion weight of the missing or low-confidence modality feature.

[0014] In this solution, the specialized attention head focuses on learning the correlation between various single-modal feature data, ensuring the accuracy of fusion weights under normal circumstances and improving the accuracy of anti-tipping sensing results. The global attention head solves the problem in existing technologies where fusion results are distorted and anti-tipping control fails when a certain modality's data is missing or invalid. When a sensor malfunctions, resulting in missing corresponding modality data, or when data confidence is too low (e.g., sensor interference), the global attention head can supplement the information using other valid modality data, while automatically reducing the weight of the failed modality to avoid invalid data affecting the fusion results, ensuring the continuity and reliability of anti-tipping sensing and control actions. The dual attention head design gives the wheelchair anti-tipping system stronger anti-interference capabilities and fault tolerance, adapting to complex usage environments and further ensuring user safety.

[0015] Furthermore, the multi-level anti-tipping action based on the anti-tipping perception result includes: before the actual shift of the center of gravity, performing a pre-compensation action for the center of gravity based on the predicted future shift trend of the center of gravity; when the shift of the center of gravity exceeds a preset safety threshold but does not reach a preset safety critical value, adjusting the wheelchair posture to pull the center of gravity back to the safe range based on the real-time target stable center of gravity; and when the shift of the center of gravity reaches the preset safety critical value, performing wheelchair degree of freedom locking and / or emergency braking.

[0016] In this solution, the pre-emptive compensation action proactively intervenes before the actual shift in the center of gravity occurs, suppressing sudden changes in the center of gravity at the source. This changes the passive mode of existing technologies that only remedies the situation after the tilt exceeds the threshold. Adjusting the wheelchair posture performs closed-loop correction when the center of gravity exceeds the threshold but has not reached its limit, pulling the center of gravity back to the safe range and achieving stable dynamic balance. Locking and emergency braking provide ultimate safety protection when critical values ​​are reached. The three levels of actions are progressive and coordinated, avoiding frequent triggering of limit protection that would affect normal use, while ensuring user safety in extreme situations.

[0017] Furthermore, the forward center of gravity compensation action includes at least one of controlling the seat to tilt slightly backward or forward and adjusting the differential speed of the drive wheels; the adjustment of the wheelchair posture includes at least one of controlling the seat to rotate in the opposite direction and controlling the seat height adjustment.

[0018] Furthermore, the step of executing multi-level anti-tipping actions based on the anti-tipping perception results also includes: triggering an audible and visual alarm when the center of gravity shifts beyond a preset safety threshold.

[0019] In this solution, an audible and visual alarm is triggered when the center of gravity shifts beyond a preset safety threshold. This promptly alerts the user to the risk and prompts them to adjust their actions, such as stopping exertion or changing their posture. It also warns those around the user, further enhancing human-machine collaborative safety. This is a low-cost and highly efficient auxiliary protection measure.

[0020] Furthermore, the prediction time window for the future center of gravity shift trend is adjusted according to the wheelchair's driving speed, and the prediction time window ranges from 0.3s to 1.2s.

[0021] In this solution, the prediction time window for future center of gravity shift trends is adaptively adjusted based on the wheelchair's speed. A longer window is used in high-speed or high-risk scenarios to allow for more time for compensation and response; a shorter window is used in low-speed or low-risk scenarios to reduce computational consumption and unnecessary intervention. This dynamic adjustment mechanism balances safety and system efficiency, avoiding the problems of over-protection or insufficient response caused by a fixed window.

[0022] Based on the same technological concept, an intelligent wheelchair is designed, comprising: a data acquisition module for acquiring multi-source sensor data related to the wheelchair's center of gravity stability, the multi-source sensor data including environmental data, wheelchair posture data, and user action data; a processing module for performing spatiotemporal synchronization and feature extraction on the multi-source sensor data to obtain multiple single-mode feature data; employing a cross-modal attention dynamic weight fusion method to perform feature fusion and decoding on the multiple single-mode feature data to obtain the wheelchair's anti-tipping perception result; and a control execution module for executing multi-level anti-tipping actions.

[0023] Compared with existing technologies, the beneficial effects of this application are as follows: By collecting multi-source data directly related to center of gravity stability—environment, wheelchair posture, and user actions—it achieves comprehensive coverage of factors affecting wheelchair center of gravity, avoiding center of gravity judgment bias caused by incomplete collection of single-type data; through spatiotemporal synchronization and feature extraction, it solves the problems of asynchronous time and inconsistent spatial coordinate systems of multi-source data, ensuring the accuracy of single-mode feature data; by adopting a cross-modal attention dynamic weight fusion method, it breaks through the limitations of existing fixed weight fusion, and can dynamically adjust the weights according to the effectiveness of each modality data in different scenarios, making the fused anti-tipping perception results more consistent with the actual center of gravity state of the wheelchair and accurately capturing the center of gravity shift trend; and by executing multi-level anti-tipping actions based on the perception results, it replaces the existing single emergency braking control method, and can execute targeted actions according to different center of gravity shift situations, effectively reducing the risk of tipping and improving the safety and stability of wheelchair use. Attached Figure Description

[0024] Figure 1 This is a flowchart illustrating the intelligent wheelchair anti-tipping control method described in this application.

[0025] Figure 2 This is a schematic diagram of the process for spatiotemporal synchronization and feature extraction of multi-source sensor data as described in this application.

[0026] Figure 3This is a schematic diagram illustrating the process of feature fusion and decoding of multiple single-modal feature data using the cross-modal attention dynamic weight fusion method described in this application.

[0027] Figure 4 This is a schematic block diagram of the structure of the intelligent wheelchair described in this application. Detailed Implementation

[0028] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0029] With the aging population and increasing travel needs of people with disabilities, smart wheelchairs, as assistive mobility devices, have been widely used in various scenarios such as homes, nursing homes, and rehabilitation institutions. Their safety and stability are directly related to the personal safety of users. In recent years, smart wheelchairs with seat height adjustment, rotation, and fitness functions have gradually become more popular, meeting the needs of users in various scenarios such as posture adjustment and rehabilitation training. However, during the use of such wheelchairs, situations such as seat height adjustment and rotation, user exertion during fitness exercises, and driving on slopes or slippery surfaces can cause sudden changes in the center of gravity, which can easily lead to tipping or rollover accidents, threatening user safety.

[0030] Current anti-tipping technologies for smart wheelchairs mostly employ sensor-based active anti-tipping solutions. These solutions involve installing various sensors on the wheelchair to collect posture or center-of-gravity pressure data. When a center-of-gravity shift exceeds a safety threshold, braking or posture adjustment actions are triggered to actively prevent tipping. However, this active anti-tipping solution lacks effective synchronization and fusion processing of the collected multi-source data. It cannot dynamically adjust based on the validity of data in different scenarios, resulting in fused feature data that fails to accurately reflect the actual stability of the wheelchair's center of gravity and makes it difficult to accurately predict center-of-gravity shift trends. Furthermore, the anti-tipping control actions are mostly single emergency braking, lacking hierarchical control logic and unable to execute targeted anti-tipping actions based on the degree and trend of center-of-gravity shift.

[0031] To address the aforementioned issues, this embodiment provides an intelligent wheelchair anti-tipping control method and an intelligent wheelchair. By collecting multi-source sensor data related to the wheelchair's center of gravity stability, it achieves precise anti-tipping perception and executes targeted actions based on different center of gravity shifts, effectively reducing the risk of tipping and improving the safety and stability of wheelchair use. Example 1

[0032] like Figure 1As shown, this embodiment provides an intelligent wheelchair anti-tipping control method, which includes the following steps S100 to S400.

[0033] The S100 collects multi-source sensor data related to the stability of the wheelchair's center of gravity. The multi-source sensor data includes environmental data, wheelchair posture data, and user action data.

[0034] The S200 performs spatiotemporal synchronization and feature extraction on multi-source sensor data to obtain multiple single-mode feature data.

[0035] The S300 employs a cross-modal attention dynamic weight fusion method to fuse and decode multiple single-modal feature data to obtain the anti-tipping perception results for the wheelchair.

[0036] The S400 executes multi-level anti-tipping actions based on anti-tipping sensing results.

[0037] In step S100, environmental data reflects the external operating conditions of the wheelchair, such as road surface type (asphalt, gravel, slippery), slope, adhesion coefficient, bumpiness, and obstacles ahead. Environmental data can be acquired using sensors such as LiDAR, millimeter-wave radar, visual cameras, and road surface IMUs (Inertial Measurement Units). Wheelchair posture data reflects the current position and motion state of the wheelchair, such as seat height, rotation angle, chassis three-dimensional tilt angle, and drive wheel torque. Wheelchair posture data can be acquired using stroke sensors, angle sensors, chassis IMUs, and motor torque sensors. User motion data reflects the user's real-time center of gravity distribution and limb movement trends, such as the user's three-dimensional center of gravity coordinates, hip pressure distribution, limb force direction and intensity, and forward sway characteristics. This data is acquired using arrayed fiber optic pressure sensors, tension or resistance sensors, and limb posture IMUs.

[0038] Additionally, a pre-stored personalized tilting model is loaded synchronously to obtain the user's basic parameters for the center of gravity safety boundary, which will then be used in subsequent calculations.

[0039] The method for constructing the personalized tilting model for users is as follows: Upon first use, a basic model is generated based on the user's height, weight, hemiplegic side, limb muscle strength level, and center of gravity distribution in a standard sitting posture; combined with seat height, road conditions, etc., a unique three-dimensional center of gravity safety boundary is dynamically generated; among them, the safety threshold for the affected side of hemiplegic users is automatically narrowed by 15%-30%, the threshold for overweight users is narrowed by 10%-20%, and the threshold in all directions is narrowed by 5% for every 5cm increase in seat height; based on usage data, the basic model is gradually iterated to adapt to the user's limb rehabilitation progress and changes in usage habits.

[0040] like Figure 2As shown, in step S200, spatiotemporal synchronization and feature extraction are performed on the multi-source sensor data to obtain multiple single-mode feature data, including the following steps S201 to S203.

[0041] S201, timestamps the collected multi-source sensor data and map all the data to the predetermined coordinate system of the wheelchair chassis.

[0042] Different sensors have different sampling frequencies, trigger times, and coordinate system origins. Microsecond-level timestamp alignment eliminates time asynchrony; a pre-calibrated extrinsic parameter matrix transforms all data into a Cartesian coordinate system (X forward, Y right, Z upward) with the wheelchair chassis geometric center as the origin, ensuring spatial consistency—a prerequisite for multimodal fusion. During timestamp alignment, interpolation, such as linear or spline interpolation, aligns low-frequency sensor data to a common time reference with the highest sampling rate, ensuring that data from the same moment reflects the same physical scene. During spatial coordinate mapping, a pre-calibrated extrinsic parameter matrix from each sensor's coordinate system to the predetermined coordinate system of the wheelchair chassis is obtained. All sensor measurements are then mapped to this unified coordinate system through extrinsic parameter transformation, ensuring that the same physical point described by different sensors, such as the user's center of gravity in space, has a consistent spatial coordinate expression.

[0043] S202, the spatially aligned data is filtered, denoised and normalized to obtain environmental preprocessed data, wheelchair posture preprocessed data and user action preprocessed data.

[0044] Even after spatiotemporal synchronization, the multi-source data still contains noise and outliers, and the physical quantities have different dimensions. Therefore, the aligned data undergoes filtering, denoising, and normalization. First, filtering and denoising are performed. Based on the noise characteristics of different sensors, extended Kalman filtering is used to remove Gaussian white noise and vibration interference from signals such as IMU and motor torque. Median filtering or statistical outlier removal methods are used to remove non-Gaussian noise and outliers from sensors such as radar and ultrasound, thus obtaining smooth data that accurately reflects the physical changes related to the center of gravity. Subsequently, the modal data are normalized. Min-max normalization or Z-score standardization maps physical quantities with different dimensions to a unified numerical range, eliminating the adverse effects of dimensional differences on subsequent feature extraction and fusion.

[0045] S203 extracts features from the preprocessed environmental data, preprocessed wheelchair posture data, and preprocessed user motion data to obtain multiple single-mode feature data, including environmental feature data, wheelchair posture feature data, and user motion feature data.

[0046] For normalized environmental data, wheelchair posture data, and user motion data, lightweight neural networks were used to extract their respective single-modal spatiotemporal features. Environmental data was feature-encoded using a combination of a deep separable convolutional network and a unidirectional long short-term memory network to capture the dynamic trends of environmental parameters such as road surface type, slope, and adhesion coefficient. Wheelchair posture and user motion data were processed using a combination of convolutional networks and a bidirectional long short-term memory network. The bidirectional long short-term memory network can simultaneously utilize past and future temporal information to effectively capture the inertial effects of seat lifting and rotation, as well as the pre-exertion sway features of the user's movements. The neural networks of all three modalities output feature vectors of the same dimension, achieving both data dimensionality reduction and abstraction while preserving the spatiotemporal patterns strongly correlated with center of gravity stability for each modality. This lays the feature foundation for subsequent cross-modal attention dynamic weight fusion. The entire feature extraction process employs a lightweight design, with single-frame inference time controlled to the millisecond level, meeting the real-time requirements of the wheelchair embedded platform.

[0047] It should be noted that the aforementioned neural network needs to be trained offline beforehand. A labeled dataset containing typical scenarios (slopes, bumpy roads, seat height adjustment, user exertion during exercise, etc.) should be constructed. The future center of gravity shift trajectory or tipping risk level should be used as the supervision signal. The model should be trained using mean squared error or cross-entropy loss functions, and then deployed to the wheelchair's embedded main control chip through quantization and pruning techniques. Existing technologies can be used for offline training of the neural network; specific training methods will not be elaborated upon here, but they should at least include the following: The dataset was constructed by collecting more than 1,000 hours of real-world usage data, covering various typical scenarios such as flat ground, slopes, slippery roads, seat height adjustment, and user fitness, with at least 20 scenarios. The center of gravity shift trajectory and tipping risk level were labeled for each frame of data. The loss function is designed using a multi-task joint loss function, with the total loss being 0.6 × centroid trajectory prediction MSE loss + 0.3 × dumping risk classification cross-entropy loss + 0.1 × weight regularization loss. The model deployment employs INT8 quantization and structured pruning techniques to compress the model parameters to below 5MB, with a single-frame inference time of ≤8ms, adapting to the computing power requirements of the STM32H743 main control chip.

[0048] Timestamp alignment enables time synchronization of data from different types of sensors, avoiding delays in center of gravity status judgment caused by differences in data acquisition time. Mapping all data to the predetermined coordinate system of the wheelchair chassis solves the problem of inconsistent spatial references between different sensor data, ensuring spatial correlation and comparability of the data, and laying the foundation for subsequent feature extraction and fusion. Filtering and denoising effectively removes invalid data generated by environmental interference and equipment errors during sensor acquisition, reducing the impact of noise on feature extraction and improving data purity. Normalization eliminates dimensional differences between different types of data, preventing a single type of data from dominating the feature fusion result due to excessively large numerical ranges. Feature extraction is performed on the three types of preprocessed data separately, accurately capturing the core features related to center of gravity stability in each type of data, providing high-quality single-mode feature support for subsequent cross-modal fusion, and improving the accuracy of anti-tipping sensing results.

[0049] like Figure 3 As shown, in step S300, a cross-modal attention dynamic weight fusion method is used to perform feature fusion and decoding on multiple single-modal feature data to obtain the anti-tipping perception result of the wheelchair, including the following steps S301 to S303.

[0050] S301, based on a cross-modal attention fusion network, learns the correlation between multiple single-modal feature data through multiple attention heads, and determines the fusion weight of each single-modal feature data.

[0051] Specifically, the three single-modal feature vectors output in step 200—environmental feature data, wheelchair posture feature data, and user action feature data—are stacked to form a feature matrix, which is then used as the input to the cross-modal attention fusion network. The cross-modal attention fusion network employs a multi-head self-attention architecture with six parallel attention heads, each independently calculating the attention weights between modalities. In the specific calculation, the feature matrix is ​​first linearly transformed to obtain the query matrix, key matrix, and value matrix. Then, the attention score is calculated, and the attention score is multiplied by the value matrix to obtain the output of that head. The outputs of all heads are concatenated along the feature dimensions and then mapped back to the original feature dimensions through a linear transformation layer.

[0052] The six attention heads are divided into two groups: three specialized attention heads and three global attention heads. The specialized attention heads, through the design of different attention masks or the use of different query-key projection matrices, force each head to focus on learning the correlation between specific modal pairs. Specifically, the first specialized head focuses on the correlation between user action features and wheelchair posture features, used to capture the impact of user's active force on seat stability; the second specialized head focuses on the correlation between environmental features and wheelchair posture features, used to assess the effects of external conditions such as road slope, adhesion coefficient, and bumpiness on wheelchair posture, thereby indirectly reflecting the risk of center of gravity drift; the third specialized head focuses on the correlation between user action features and environmental features, used to determine the risk differences of the same user action under different road conditions, such as the different effects of exercising on the center of gravity on flat ground versus on a ramp. The global attention heads are used to complete information using other modal features when a certain modal feature data is missing or its confidence level is below a threshold, and automatically reduce the fusion weight of the missing or low-confidence modal feature, solving the problem of distorted fusion results and tipping control failure when a certain modal data is missing or invalid.

[0053] The outputs of the six attention heads are concatenated and fed into a small weight generation network consisting of two fully connected layers. Finally, a softmax function outputs the fusion weights for three modalities: environmental feature weights, wheelchair posture weights, and user action feature weights. Each weight ranges from 0 to 1, and the sum of the three weights is 1. These weights are dynamically determined by the current input features. When the user's exertion during exercise is obvious, the user action feature weights automatically increase; when the wheelchair is traveling on a slippery, steep slope, the environmental feature weights increase; and when the seat is raised or lowered, the wheelchair posture weights increase. Furthermore, the confidence level or data validity indicators of each sensor are monitored in real time. When a modal feature is missing or its confidence level falls below a preset threshold (e.g., below 0.5) due to sensor failure, communication interruption, or environmental interference (such as a blocked camera or an offline fiber optic pressure sensor), the global attention head automatically completes the feature using the feature information from the other two modalities. Through attention weight redistribution, the missing modal information is implicitly recovered from the healthy modality. Simultaneously, the weight generation network reduces the fusion weight of the missing modality to near 0, ensuring it has almost no impact on subsequent fusion. This mechanism ensures that even in the extreme case of one or two sensor failures, the system can still output reasonable anti-tipping sensing results, significantly improving overall robustness.

[0054] S302, based on the fusion weight, performs weighted fusion of multiple single-mode feature data to generate a fused feature vector.

[0055] The environmental feature data, wheelchair posture data, and user action feature data are weighted and fused based on dynamically generated environmental feature weights, wheelchair posture weights, and user action feature weights. First, each feature vector is multiplied by its corresponding weight. Then, the three product vectors are summed element-wise to obtain the fused feature vector. To ensure numerical stability, each single-modal feature vector is normalized before weighted summation. The fused feature vector has the same dimension as the individual modal feature vectors, which condenses the key information from the environment, wheelchair posture, and user actions while maintaining a compact representation, facilitating efficient processing by the subsequent decoder.

[0056] S303, decode the fused feature vector to obtain the anti-tipping perception result; the anti-tipping perception result includes the current center of gravity state, the future center of gravity offset trend and the dynamic safety boundary.

[0057] The generated fused feature vector is input into the decoder network, which employs a multi-task learning architecture consisting of a shared fully connected layer and three independent task-branch fully connected layers. The first task branch outputs the current center of gravity state, including the three-dimensional coordinates (X, Y, Z) of the user's center of gravity in the wheelchair chassis coordinate system, the real-time pitch and roll angles of the chassis, and the offset vector of the center of gravity relative to the stable center point. The second task branch outputs the future center of gravity offset trend, i.e., predicts the trajectory of the center of gravity within a preset time window. Specifically, this is expressed as a sequence of three-dimensional coordinates or tilt angles of the center of gravity at fixed time intervals, such as every 50 milliseconds. The duration of the prediction window is dynamically adjusted according to the current speed of the wheelchair; the higher the speed, the longer the window, in order to obtain more sufficient response time in advance. The third task branch outputs dynamic safety boundaries, including center-of-gravity safety thresholds and center-of-gravity safety thresholds. The center-of-gravity safety threshold is the safe center-of-gravity offset or safe tilt angle value in all directions under the current scene, ensuring no tipping. Examples include forward tilt safety thresholds, backward tilt safety thresholds, left tilt safety thresholds, and right tilt safety thresholds. The center-of-gravity safety threshold is the maximum allowable center-of-gravity offset or maximum tilt angle value in all directions under the current scene, such as forward tilt safety thresholds, backward tilt safety thresholds, left tilt safety thresholds, and right tilt safety thresholds. These thresholds integrate user-personalized model parameters and real-time environmental information, automatically narrowing or widening as the scene changes. The three branches of the decoder compute in parallel, ultimately packaging the output results into structured data, i.e., the anti-tipping perception results.

[0058] By using multiple attention heads in a cross-modal attention fusion network, the intrinsic relationships between various single-modal feature data can be accurately learned, enabling dynamic adaptive adjustment of fusion weights. This solves the problem in existing technologies where fixed weights cannot adapt to changes in data validity across different scenarios. For example, when driving on a slope, the weights of environmental data and wheelchair posture data are automatically increased, while the weights of user action data are appropriately reduced, ensuring that the fusion result matches the actual scenario requirements. Based on dynamic weights, weighted fusion generates a fusion feature vector that integrates the advantages of various single-modal features, accurately reflecting the wheelchair's center of gravity stability. The anti-tipping perception result obtained after decoding not only includes the current center of gravity state but also predicts future center of gravity shift trends and dynamic safety boundaries, providing lead time for subsequent multi-level anti-tipping actions. The output of dynamic safety boundaries can dynamically adjust the safety threshold according to the real-time scenario, adapting to safety requirements under different driving speeds and user actions, thus improving the flexibility and adaptability of anti-tipping control.

[0059] It should be noted that the prediction time window for the future center of gravity shift trend is adjusted according to the wheelchair's speed, with a prediction time window range of 0.3s-1.2s. This prediction time window adaptively adjusts based on the wheelchair's speed, using a longer window in high-speed or high-risk scenarios to allow for more time for compensation and response; and a shorter window in low-speed or low-risk scenarios to reduce computational consumption and unnecessary intervention. This dynamic adjustment mechanism balances safety and system efficiency, avoiding the problems of over-protection or insufficient response that can occur with a fixed window.

[0060] It should be noted that the intelligent wheelchair in this embodiment integrates functions such as multi-degree-of-freedom seat movement, wheel drive, and audible and visual alarms. The multi-degree-of-freedom seat movement includes actions such as reclining or tilting forward, rotating, raising and lowering, and locking. The wheel drive includes rotation and braking, and differential torque adjustment of the left and right drive wheels. Based on the anti-tipping perception results, multi-level anti-tipping actions are executed, including the following three levels.

[0061] The first level involves performing pre-shift compensation actions based on the predicted future center of gravity shift trend before the actual shift occurs.

[0062] Based on the future trend of center of gravity shift, the system proactively compensates for this shift before the center of gravity begins to move. Specific compensation actions include slightly tilting the seat backward or forward, adjusting the differential torque of the left and right drive wheels to generate a counter-torque, or performing both actions simultaneously. The magnitude and direction of the compensation are calculated using inverse kinematics based on the predicted trajectory of the future center of gravity shift. The goal is to ensure that the change in center of gravity caused by the compensation action precisely offsets the expected shift, thus preventing the center of gravity from exceeding the safety boundary in the first place. When the wheelchair travels at a high speed, the road surface adhesion coefficient is low, or the seat height is significantly increased, the system automatically extends the prediction time window to allow for a more sufficient compensation response time; under low-speed or static conditions, the window is shortened to reduce unnecessary intervention.

[0063] The second level is to adjust the wheelchair posture to pull the center of gravity back to the safe range when the center of gravity shifts beyond the center of gravity safety threshold but does not reach the center of gravity safety critical value, based on the real-time target stable center of gravity. If the current compensation control fails to completely prevent center of gravity shift, or if the center of gravity shift exceeds the preset safety threshold but has not yet reached the preset safety critical value, this level of control is activated. Using the optimal stable center of gravity calculated in real-time as the closed-loop control target, an adaptive proportional-integral-derivative (PID) control algorithm is employed to continuously adjust the wheelchair's seat posture and drive wheel status, dynamically pulling the real-time center of gravity back to the safe range. Specific actions include controlling the seat to rotate in the opposite direction, adjusting the seat height, and generating a corrective torque through differential drive. The PID parameters are not fixed values ​​but are tuned in real-time based on the current anti-tipping sensing results: when the center of gravity shift speed is high or the road surface adhesion coefficient is low, the proportional gain increases to improve response speed; when the steady-state error is small, the integral gain decreases to prevent overshoot. This process continues until the center of gravity shift is less than 80% of the safety threshold. Throughout the correction process, the system simultaneously monitors the control effect; if the center of gravity shift further deteriorates and reaches the critical value, the next level of control is immediately triggered.

[0064] The third level involves locking the wheelchair's degrees of freedom and / or applying emergency braking when the center of gravity shifts to the critical safety value.

[0065] When the center of gravity shift reaches a dynamically generated preset safety threshold, or when the system detects a sudden road risk (such as a sudden pothole or slippery ice surface) or sensor malfunction (such as loss of critical sensor data), this level of control is immediately triggered. First, the seat's lifting and rotational degrees of freedom are locked via an electromagnetic locking mechanism to prevent further deterioration of the center of gravity position due to continued seat raising or rotation. Second, maximum braking torque is applied to the left and right drive wheels to achieve emergency braking. Finally, if the wheelchair is equipped with retractable anti-rollover outriggers, they automatically extend to increase support width. These actions are completed within 10 milliseconds of detecting the trigger condition and are unaffected by any parameter adjustments in the algorithm's closed loop. Once triggered, this level will not automatically deactivate unless manually reset by the user or the system detects that the center of gravity has fully stabilized.

[0066] The pre-emptive compensation mechanism proactively intervenes before the actual shift in the center of gravity occurs, suppressing sudden changes in the center of gravity at its source. This departs from the passive approach of existing technologies that only remedy the situation after the tilt exceeds a threshold. Adjusting the wheelchair's posture performs a closed-loop correction when the center of gravity exceeds the threshold but has not reached its limit, pulling the center of gravity back to a safe range and achieving stable dynamic balance. Locking and emergency braking provide ultimate safety assurance when critical values ​​are reached. These three levels of actions are progressive and coordinated, avoiding frequent triggering of limit protection that could disrupt normal use while ensuring user safety in extreme situations.

[0067] In addition, based on the anti-tipping perception results, multi-level anti-tipping actions are executed, including triggering an audible and visual alarm when the center of gravity shifts beyond a preset safety threshold. Specifically, a buzzer on the wheelchair emits a tiered audible alarm, such as intermittent short beeps for slight shifts and continuous long beeps for severe shifts, while simultaneously illuminating LED warning lights on the armrests or dashboard, for example, changing color from yellow to red. Triggering the audible and visual alarms can promptly remind users of the risks and encourage them to proactively adjust their actions, such as stopping exertion or changing their posture, while also alerting those around them, further enhancing human-machine collaborative safety. This is a low-cost, high-efficiency auxiliary protection measure.

[0068] It should be noted that during the execution of multi-level anti-tipping actions, execution process data can be collected and used as optimization signals to inversely optimize the feature fusion and decoding process of multiple single-mode feature data. This involves inversely optimizing the parameters of the cross-modal attention fusion network model in step S300, such as fusion weights and dynamic safety boundaries, forming a bidirectional closed-loop adaptive optimization mechanism between perception and control. The execution process data includes center of gravity prediction error, control steady-state error, false / missed warning events, actual risk events, and user center of gravity change patterns. Specific optimization includes the following three levels: The first level is real-time micro-closed-loop feedback. In each control cycle (e.g., 10ms), the error between the actual center of gravity compensation effect and the predicted center of gravity trajectory is calculated. Based on the center of gravity prediction error and the control steady-state error, the weight allocation of the cross-modal attention head is fine-tuned in real time. For example, if the prediction error is too large for several consecutive cycles (the predicted value is greater than the actual value), the attention weight of the user's action modality is appropriately reduced, because the intensity of the user's force may have been overestimated.

[0069] The second level is short-term adaptive closed-loop feedback. During a single use, it accumulates statistics on false alarms (warnings issued even though no actual risk occurred) and missed alarms (risks occurred but no warning was issued). The warning trigger threshold is adjusted: if false alarms are frequent, the threshold is appropriately increased; if missed alarms are frequent, the threshold is decreased. Road surface recognition sensitivity is optimized; for example, if the system is too sensitive on specific road surfaces (such as slippery surfaces), the risk assessment parameters for that scenario are adjusted.

[0070] The third level involves a long-term incremental learning loop. After a user has used the wheelchair multiple times, for example, 10 times, the offline model is updated when the wheelchair is idle. This includes long-term recording of risk events (such as actual tipping warnings and locking events), patterns of user center of gravity changes (such as improved center of gravity control in hemiplegic patients after rehabilitation), and user action feature statistics. A low-rank adaptation algorithm is used to fine-tune the parameters of the cross-modal attention fusion network model and other relevant model parameters, such as the user-personalized tipping model, to adapt to user habits.

[0071] In summary, by collecting multi-source data directly related to wheelchair center of gravity stability—environment, wheelchair posture, and user actions—comprehensive coverage of factors influencing wheelchair center of gravity is achieved, avoiding the bias in center of gravity judgment caused by incomplete collection of single-type data. Spatiotemporal synchronization and feature extraction solve the problems of asynchronous time and inconsistent spatial coordinate systems among multi-source data, ensuring the accuracy of single-modal feature data. The adoption of a cross-modal attention-based dynamic weight fusion method overcomes the limitations of existing fixed-weight fusion, dynamically adjusting weights based on the effectiveness of each modality's data in different scenarios. This makes the fused anti-tipping perception results more closely match the actual center of gravity state of the wheelchair, accurately capturing the trend of center of gravity shift. Executing multi-level anti-tipping actions based on the perception results replaces the existing single emergency braking control method. Targeted actions can be performed according to different center of gravity shift situations, effectively reducing the risk of tipping and improving the safety and stability of wheelchair use. Specifically, the accuracy of center of gravity trajectory prediction is improved by more than 60%, and the prediction error is reduced from more than 20% in the existing technology to less than 8%; the anti-tipping response speed is improved by more than 60%, and center of gravity compensation can be achieved in advance of 0.3-1.2s; the system robustness is greatly improved, and the system can operate stably for ≥30s when a single sensor fails; the false alarm rate is reduced by 75%, from more than 20% in the existing technology to less than 5%. Example 2

[0072] like Figure 4 As shown in the figure, this embodiment discloses an intelligent wheelchair, the specific structure of which includes: a wheelchair body, and a data acquisition module 100, a processing module 200, a storage module 300, and a control execution module 400 integrated on the wheelchair body. It also includes a power supply module 500 and a communication module 600. All modules work together to ensure the safety and convenience of wheelchair operation. The intelligent wheelchair integrates multi-degree-of-freedom seat movement, automatic driving, user exercise, and other auxiliary functions; the specific structural design of these functions can utilize existing mature technologies in the field, and will not be elaborated upon in this embodiment.

[0073] The data acquisition module 100 is used to collect multi-source sensor data related to the stability of the wheelchair's center of gravity. The multi-source sensor data includes environmental data, wheelchair posture data, and user action data.

[0074] The data acquisition module 100 consists of multiple types of sensors and data acquisition interfaces, specifically configured as follows: Environmental data acquisition unit 110, including LiDAR, millimeter-wave radar, a vision camera, and a road surface IMU (inertial measurement unit), used to collect data on external operating conditions such as road surface type, slope, adhesion coefficient, bumpiness, and obstacles ahead; Wheelchair posture data acquisition unit 120, including a stroke sensor, angle sensor, chassis IMU, and motor torque sensor, used to collect data on wheelchair posture and motion status such as seat height adjustment, rotation angle, chassis three-dimensional tilt angle, and drive wheel torque; User motion data acquisition unit 130, including an array of fiber optic pressure sensors, tension / resistance sensors, and a limb posture IMU, used to collect user-related data such as three-dimensional center of gravity coordinates, hip pressure distribution, limb force direction and intensity, and pre-motion characteristics. All sensors communicate with the processing module via high-speed data interfaces to ensure real-time data acquisition. The sampling frequency is dynamically adjusted according to the sensor type, with the core sensor sampling frequency not less than 100Hz to meet microsecond-level timestamp alignment requirements.

[0075] The processing module 200 is electrically connected to the storage module 300, the data acquisition module 100, and the control execution module 400. The storage module 300 uses a non-volatile memory chip to store computer programs, control instructions, and various threshold parameters that can run on the processing module 200. The processing module 200 uses an embedded main control chip, which has the characteristics of fast operation speed and low power consumption. When executing the computer program or instructions, it specifically implements the following steps: performing spatiotemporal synchronization processing on the multi-source sensor data transmitted by the data acquisition module 100 to eliminate the data delay difference between different sensors, and then extracting the core features of various types of data through feature extraction algorithms to obtain multiple single-mode feature data; using a cross-modal attention dynamic weight fusion method, performing feature fusion and decoding on the multiple single-mode feature data, dynamically allocating weights according to the degree of influence of different data on the center of gravity stability, and accurately obtaining the anti-tipping perception result of the wheelchair; based on the anti-tipping perception result, sending corresponding control instructions to the control execution module 400 to control it to execute corresponding multi-level anti-tipping actions, thereby achieving precise anti-tipping control.

[0076] The control execution module 400 is electrically connected to the processing module 200. Its core function is to execute multi-level anti-tipping actions based on the control commands sent by the processing module 200, and at the same time realize the warning reminder of abnormal status. Specifically, it includes an actuator drive unit 410 and an audible and visual alarm unit 420. The two units work together to realize three-level anti-tipping control.

[0077] The actuator drive unit 410 includes a seat height adjustment mechanism, a seat rotation adjustment mechanism, a drive wheel differential mechanism, an electromagnetic locking mechanism, and retractable anti-tipping outriggers. All mechanisms utilize existing mature technologies. This unit executes anti-tipping actions in stages according to instructions from the processing module 200: The first stage is a pre-compensation action, including slightly tilting the seat backward or forward, adjusting the differential torque of the left and right drive wheels to generate a counter-torque, or simultaneously performing both actions. The second stage is a posture adjustment action, including controlling the seat to rotate in the opposite direction, adjusting the seat height, and generating a corrective torque through differential drive. The third stage is an emergency protection action, including locking the wheelchair's degrees of freedom and / or emergency braking. The audible and visual alarm unit 420 includes a buzzer and LED warning lights. When the processing module 200 detects that the wheelchair's center of gravity shift exceeds a preset safety threshold, this unit simultaneously activates an alarm: the buzzer emits a continuous or intermittent warning sound, and the LED warning light flashes. The frequency of the warning sound and the flashing frequency of the LED light are adjusted according to the severity of the center of gravity shift to promptly alert the user and surrounding personnel to safety. It should be noted that the specific structure and driving method of the actuator drive unit 410 and the audible and visual alarm unit 420 can be achieved using existing technologies in the field, and this embodiment does not make any improvements in this regard.

[0078] The power module 500 provides a stable and continuous power supply to the data acquisition module 100, processing module 200, storage module 300, control execution module 400, and communication module 500. It also integrates a battery life monitoring function, which can monitor the remaining battery power in real time and transmit the power information to the processing module 200 so that users can understand the battery status in a timely manner. The communication module 600 supports common wireless communication methods such as Bluetooth and WiFi, which can realize wireless connection with mobile terminals (mobile phones, tablets, etc.). Users can view the real-time operating status of the wheelchair, center of gravity offset data, and other information through the mobile terminal, and can also remotely adjust the anti-tipping threshold parameters. At the same time, when the wheelchair malfunctions or the center of gravity offset exceeds the safety threshold, the communication module 600 can push fault alarm information and safety warning information to the caregiver's mobile terminal, so that the caregiver can promptly detect and deal with related problems, improving the safety and convenience of wheelchair use.

[0079] Although exemplary embodiments have been described herein with reference to the accompanying drawings, it should be understood that the above exemplary embodiments are merely illustrative and are not intended to limit the scope of this application. Various changes and modifications can be made therein by those skilled in the art without departing from the scope and spirit of this application. All such changes and modifications are intended to be included within the scope of this application as claimed in the appended claims.

[0080] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0081] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another device, or some features may be ignored or not executed.

[0082] The various component embodiments of this application can be implemented in hardware, or as software modules running on one or more processors, or a combination thereof. Those skilled in the art will understand that microprocessors or digital signal processors (DSPs) can be used in practice to implement some or all of the functions of some modules according to the embodiments of this application. This application can also be implemented as an apparatus program (e.g., a computer program and computer program product) for performing part or all of the methods described herein. Such an implementation of this application can be stored on a computer-readable medium, or can be in the form of one or more signals. Such signals can be downloaded from an Internet website, provided on a carrier signal, or provided in any other form.

[0083] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0084] Although the description of this application has been made in conjunction with the specific embodiments described above, it will be apparent to those skilled in the art that many substitutions, modifications, and variations can be made based on the foregoing. Therefore, all such substitutions, modifications, and variations are included within the spirit and scope of the appended claims.

Claims

1. A method for controlling the tipping of an intelligent wheelchair, characterized in that, The method includes: Collect multi-source sensor data related to wheelchair center of gravity stability, including environmental data, wheelchair posture data, and user action data; The multi-source sensor data is spatiotemporally synchronized and feature extracted to obtain multiple single-mode feature data. A cross-modal attention dynamic weight fusion method is used to fuse and decode the multiple single-modal feature data to obtain the anti-tipping perception results of the wheelchair. Based on the anti-tipping perception results, multi-level anti-tipping actions are executed.

2. The intelligent wheelchair anti-tipping control method according to claim 1, characterized in that, The process of spatiotemporal synchronization and feature extraction of the multi-source sensor data yields multiple single-mode feature data, including: The collected multi-source sensor data is timestamped and aligned, and all data is uniformly mapped to the predetermined coordinate system of the wheelchair chassis; The spatially aligned data is filtered, denoised, and normalized to obtain preprocessed environmental data, preprocessed wheelchair posture data, and preprocessed user action data. Feature extraction is performed on the environmental preprocessing data, wheelchair posture preprocessing data, and user action preprocessing data respectively to obtain the multiple single-mode feature data.

3. The intelligent wheelchair anti-tipping control method according to claim 2, characterized in that, The process involves extracting features from the environmental preprocessing data, wheelchair posture preprocessing data, and user motion preprocessing data to obtain the multiple single-mode feature data, including: The environmental preprocessing data, wheelchair posture preprocessing data, and user action preprocessing data are used to extract features from each other to obtain the corresponding single-mode feature data.

4. The intelligent wheelchair anti-tipping control method according to claim 1, characterized in that, The method employs cross-modal attention dynamic weight fusion to fuse and decode the multiple single-modal feature data, obtaining the wheelchair anti-tipping perception results, including: Based on a cross-modal attention fusion network, the association between the multiple single-modal feature data is learned through multiple attention heads, and the fusion weight of each single-modal feature data is determined. Based on the fusion weights, the multiple single-mode feature data are weighted and fused to generate a fused feature vector; The fused feature vector is decoded to obtain the anti-tipping sensing result; The anti-tipping sensing results include the current center of gravity status, future center of gravity shift trend, and dynamic safety boundary.

5. The intelligent wheelchair anti-tipping control method according to claim 4, characterized in that, The multiple attention heads include a set of specialized attention heads and a set of global attention heads; The specific attention head is used to learn the correlation between the multiple single-modality feature data; the global attention head is used to complete the information through the other modality features when a certain modality feature data is missing or the confidence level is lower than the threshold, and automatically reduce the fusion weight of the missing or low-confidence modality feature.

6. The intelligent wheelchair anti-tipping control method according to claim 4, characterized in that, Based on the anti-tipping sensing results, the execution of multi-level anti-tipping actions includes: Before the actual shift of the center of gravity occurs, a pre-shift compensation action is performed based on the predicted future shift trend of the center of gravity. When the center of gravity shifts beyond the center of gravity safety threshold but does not reach the center of gravity safety critical value, the wheelchair posture is adjusted to pull the center of gravity back to the safe range, based on the real-time target stable center of gravity. When the center of gravity shifts to the critical value for safety, the wheelchair degrees of freedom are locked and / or emergency braking is applied.

7. The intelligent wheelchair anti-tipping control method according to claim 6, characterized in that, The forward center of gravity compensation action includes at least one of pre-deploying the anti-rollover outriggers, slightly tilting the seat backward or forward, and adjusting the differential speed of the drive wheels; the adjustment of the wheelchair posture includes at least one of controlling the seat to rotate in the opposite direction and controlling the seat height adjustment.

8. The intelligent wheelchair anti-tipping control method according to claim 6, characterized in that, The step of executing multi-level anti-tipping actions based on the anti-tipping perception results also includes: triggering an audible and visual alarm when the center of gravity shifts beyond a preset safety threshold.

9. The intelligent wheelchair anti-tipping control method according to claim 6, characterized in that, The prediction time window for the future center of gravity shift trend is adjusted according to the wheelchair's driving speed, and the prediction time window ranges from 0.3s to 1.2s.

10. An intelligent wheelchair, characterized in that, include: The data acquisition module is used to collect multi-source sensor data related to the stability of the wheelchair's center of gravity. The multi-source sensor data includes environmental data, wheelchair posture data, and user action data. The processing module is used to perform spatiotemporal synchronization and feature extraction on the multi-source sensor data to obtain multiple single-mode feature data. A cross-modal attention dynamic weight fusion method is used to fuse and decode the multiple single-modal feature data to obtain the anti-tipping perception results of the wheelchair. The control execution module is used to perform multi-level anti-tipping actions.

Images