Non-motor vehicle riding safety protection system based on multi-sensor fusion

Abstract

This invention discloses a non-motorized vehicle riding safety protection system based on multi-sensor fusion, belonging to the field of non-motorized vehicle safety technology. The system consists of sequentially connected modules for sensors, data acquisition, preprocessing, and feature extraction. Sensors collect riding data from multiple dimensions, and features are extracted after acquisition and preprocessing. By matching behavioral feature databases and mining temporal correlations of dangerous features to identify the status, a safety judgment module identifies anomalies early, adaptively adjusts thresholds, and finally, an early warning execution module outputs tiered warnings or interventions. This invention uses multi-sensor fusion to collect non-motorized vehicle riding data, and through preprocessing, sample matching, and temporal correlation identification, accurately determines the riding scenario and status. An adaptive threshold model is configured to achieve scenario-based adaptation, and tiered warnings combined with intervention operations respond promptly to dangers. Failure protection is set to avoid secondary risks. The behavioral feature database can be updated online, providing intelligent protection for non-motorized vehicle riding throughout the entire process.

Description

Technical Field

[0001] This invention relates to the field of non-motorized vehicle safety technology, specifically to a non-motorized vehicle riding safety protection system based on multi-sensor fusion. Background Technology

[0002] Non-motorized vehicles have become an important mode of transportation for short-distance travel, and their riding safety has received widespread attention. Fatigue riding, distracted driving, and improper operation are the core causes of accidents. Early identification and timely warning of these dangerous behaviors have become a key research focus in the field of non-motorized vehicle riding safety protection. The application of multi-sensor fusion technology in the field of condition monitoring is becoming increasingly mature, providing technical support for the accurate detection of riding behavior and conditions. At the same time, the development of artificial intelligence and time-series data analysis technology makes it possible to mine the correlation between behavioral characteristics and achieve intelligent judgment of dangerous conditions. Under this technological development trend, developing a safety protection system adapted to non-motorized vehicle riding scenarios has become an important direction for solving riding safety problems.

[0003] Traditional safety measures for non-motorized vehicle riding are mostly passive, relying solely on helmets and protective gear to reduce the severity of injuries after an accident. They lack the ability to actively identify and warn of dangerous behaviors during riding. Many existing active protection technologies use single sensors to collect data, detecting only a single riding state and failing to achieve multi-dimensional behavior and state monitoring. Furthermore, they do not consider the temporal correlations between different features, making them prone to biased identification of dangerous states due to incomplete data. Additionally, the judgment thresholds are often fixed and cannot be dynamically adjusted according to actual scenarios such as riding speed and road conditions, resulting in insufficient rationality and accuracy of the identification results. Moreover, the adaptability and optimizability of the protection systems are poor, making it difficult to continuously improve protective capabilities based on different riders' habits and changing scenarios. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a non-motorized vehicle riding safety protection system based on multi-sensor fusion. This system integrates multiple types of sensors to collect multi-dimensional data on rider's control behavior and state. After preprocessing and feature extraction, combined with an online-updable behavioral feature library, a temporal correlation recognition module mines the temporal correlations between dangerous features to accurately determine the riding scenario and state. The system is equipped with an adaptive threshold dynamic adjustment model to achieve scenario-based accurate determination of dangerous states such as fatigued riding and distracted driving. A tiered early warning mechanism outputs multiple forms of warning signals, and when necessary, executes safe and controllable riding intervention operations. Simultaneously, it sets failure protection and prohibition of intervention scenarios to avoid secondary risks. The overall system constructs a complete protection system from data collection and state recognition to early warning intervention, significantly improving the intelligence and effectiveness of non-motorized vehicle riding safety protection.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a non-motorized vehicle riding safety protection system based on multi-sensor fusion, comprising a sensor module, a data acquisition module, a data preprocessing module, and a feature extraction module connected in sequence, wherein the feature extraction module is connected to a behavior feature library module and a temporal correlation recognition module, the temporal correlation recognition module is connected to a safety determination module, and the safety determination module is connected to a warning execution module; The sensor module consists of a specially arranged grip force sensor, finger capacitive sensor, head posture sensor, and riding trajectory sensor, used to collect multi-dimensional data on the rider's control behavior and riding status. The data acquisition module is configured with different acquisition frequencies according to the sensor detection characteristics, and a data verification mechanism is set up to ensure the validity of data transmission. The acquired raw data is then transmitted to the data preprocessing module. The data preprocessing module performs noise reduction, normalization, and time alignment on the raw data to obtain standardized data. Time alignment is used to unify the time axis of the data from the four sources of sensors. The feature extraction module extracts grip force features, finger capacitance features, head posture features, and cycling trajectory features from standardized data; The behavior feature library module stores control behavior feature samples under different cycling scenarios, and establishes a rider control behavior feature library that supports online updates. The temporal correlation recognition module completes accurate scene determination through sample library matching, historical same scene baseline and fluctuation calculation, and DR deviation rate verification. At the same time, it explores the temporal correlation between three sets of dangerous features: grip force decay and abnormal head deflection, single hand detachment and trajectory drift, and abnormal head deflection and trajectory drift. It matches the features extracted in real time with samples in the behavioral feature library to identify the rider's current riding status. The safety determination module uses the temporal correlation recognition results and preset thresholds to make early determinations of cycling fatigue, distracted driving and illegal operation. It also configures an adaptive threshold dynamic adjustment model to achieve scenario-based adaptation of the determination threshold. The early warning execution module outputs corresponding graded early warning signals based on the safety assessment results, and performs cycling intervention operations when necessary.

[0006] Furthermore, the grip force sensor is a piezoresistive force sensor, symmetrically installed on the left and right grip areas of the electric bicycle handlebars, with at least two grip force sensors installed in each grip area; the finger capacitive sensor is a capacitive touch sensor, integrated on the surface of the grip area, and corresponding one-to-one with the grip force sensor; the head posture sensor is an inertial measurement unit integrating a gyroscope, accelerometer, and magnetometer, worn on the forehead of the rider's helmet, outputting three-dimensional posture data of head pitch angle, roll angle, and yaw angle; the riding trajectory sensor adopts a GPS and inertial measurement unit fusion architecture, installed at the front of the electric bicycle, the GPS unit collects riding position, speed, and heading data, and the inertial measurement unit performs dead reckoning when the GPS signal is interrupted.

[0007] Furthermore, the data preprocessing module uses wavelet transform to denoise the original data and employs a normalization method to map the denoised data to a preset numerical range of [0, 1]. The normalization mathematical expression is:

[0008] In the formula, The original physical quantity of the sensor, This is the minimum value of the sample for this physical quantity. This represents the maximum value of the sample for this physical quantity. These are the normalized feature values; the data preprocessing module uses a 3-layer attention network to construct an association matrix, with an input dimension of M×T, where M is the total number of features from the four source sensors, T is 30 time steps, and the output dimension is an M×M feature association matrix. The network training objective is to minimize the temporal alignment error of the multi-source sensor data, and the loss function is the mean square error function, mathematically expressed as:

[0009] In the formula, This is the mean squared error loss value. The total number of samples, For the actual value of timing alignment, The predicted values ​​are time-aligned. The time alignment is achieved using a sliding window mechanism with a window length of 10 seconds. The spatial and temporal correlation of each sensor's data is captured through an association matrix, and the sampling times of the four source sensors are aligned to the same time axis.

[0010] Furthermore, the grip force characteristics include grip force amplitude and grip force decay rate; the finger contact characteristics include finger contact state and contact area; the head posture characteristics include head deflection angle and deflection frequency; and the cycling trajectory characteristics include trajectory offset and offset rate. The initial benchmark for the head reference posture and normal cycling trajectory is determined based on the feature data collected during the 5-minute initial normal cycling phase of the cyclist. Subsequently, a sliding window mechanism with a window length of 10 seconds is used to dynamically iterate and update the benchmark value in combination with sample weights. The weight of recent samples is set to 0.8, and the weight of long-term samples is set to 0.2. The cycling fatigue level determination supports dual determination criteria of raw physical quantities and normalized feature values. The two criteria are converted bidirectionally through the normalization method in the data preprocessing stage, and the determination threshold of fatigue level is adjusted synchronously according to the conversion relationship.

[0011] Furthermore, the process of establishing a rider control behavior feature database by the behavior feature database module includes sample collection, sample preprocessing, feature database establishment, and feature matching, and the feature database update trigger threshold is when the number of newly added valid samples reaches 1000 sets. Sample collection: Recruit cyclists of different ages and cycling proficiency levels, and collect data from four sources of sensors in four scenarios: normal cycling, fatigued cycling, distracted driving, and illegal operation, in indoor simulated scenarios and outdoor real cycling scenarios. Sample preprocessing: The collected sample data is denoised, normalized, and time-aligned. After feature extraction, the sample is labeled. Feature library establishment: Train the labeled feature samples to build and store the manipulation behavior feature model. Incremental network is introduced to realize online update of feature library. Incremental learning is triggered by a preset threshold of 1000 new effective samples. During update, the backbone model is frozen and only the incremental network is trained and its output is fused with the backbone model. Feature matching: Configure an environmental noise detection unit and a dynamic weight allocation model. The environmental noise detection unit quantifies the cycling environment noise into three levels: low, medium, and high. The dynamic weight allocation model adaptively updates the matching weights of each sensor feature based on the noise quantization results. At the low noise level, the matching weights of each sensor feature are equal. At the medium and high noise levels, the feature matching weights of the head posture sensor and cycling trajectory sensor are increased to 0.35, while the feature matching weights of the grip force sensor and finger capacitance sensor are reduced to 0.15.

[0012] Furthermore, the temporal association recognition module is based on an LSTM-based 3-layer temporal association model. The input dimension is N×T, where N is the total dimension of the four core feature classes and T is the 30 time steps. The output dimension is a D-dimensional feature association vector, where D is the 64-dimensional hidden layer mapping dimension. The model training objective is to maximize the feature association recognition accuracy, and the loss function is the cross-entropy loss function, mathematically expressed as:

[0013] In the formula, This represents the cross-entropy loss value. The total number of samples, The true label value of the sample.

[0014] This represents the predicted probability value of the sample. The temporal correlation recognition module sets up dedicated correlation mining branches for three groups of dangerous features: grip force attenuation and abnormal head deflection, single hand disengagement and trajectory drift, and abnormal head deflection and trajectory drift. Each branch is a combination structure of a single fully connected layer and an attention mechanism. The attention mechanism assigns dynamic weights of 0-1 to features at different time points, strengthening the weights of feature nodes with correlation indicative during the occurrence of dangerous behavior and weakening unrelated noisy feature nodes. The correlation strength of the three groups of dangerous features is calculated using the Pearson correlation coefficient, with a correlation strength threshold δ set to 0.6. When the Pearson correlation coefficient ≥ 0.6, a significant temporal correlation is determined between the features. The mathematical expression is:

[0015] In the formula, The Pearson correlation coefficient is used. The first set of features Each sample value The sample mean of the first set of features. The second set of features Each sample value The sample mean of the second set of features. The number of feature samples.

[0016] Furthermore, the step of the temporal correlation recognition module in matching the real-time extracted features with samples in the behavior feature library is as follows: the real-time extracted features from the four sources of sensors are serialized and normalized to construct a real-time feature vector A of a unified dimension; and the sample feature vectors of the four types of cycling scenarios stored in the behavior feature library are retrieved. The cosine similarity algorithm is used to calculate the similarity value between the real-time feature vector and the feature vector of each sample one by one. The similarity threshold T is set to 0.75. When the similarity value of a sample is ≥0.75, it is determined that the sample has initially matched the real-time feature. The mathematical expression of cosine similarity is:

[0017] In the formula, Similarity is the cosine similarity value. This is the fusion feature vector extracted in real time from four sensor sources. For vectors The 3D eigenvalues For the first in the behavioral feature database Sample feature vectors for cycling scenarios For vectors The i-th eigenvalue, The total dimension of the feature vectors; Preset similarity threshold When the similarity value of the corresponding samples is greater than or equal to At that time, it is determined that the sample has successfully matched the real-time features.

[0018] Furthermore, the temporal correlation recognition module retrieves feature data from historical similar scenes in the feature library, calculates the baseline values ​​and standard deviations of the fluctuations of each dimension of features under historical similar scenes using a sliding window mechanism, calculates the overall deviation rate between the initially successfully matched samples and the features of historical similar scenes based on the deviation rate formula, and compares the overall deviation rate with a preset preference feature threshold. contrast, Set to 0.2, if Then the matching result is confirmed to be valid. Then discard the sample and re-screen until the current cycling scenario type is determined; Baseline value calculation formula:

[0019] In the formula, For the first time in the same historical scene Baseline value of dimensional feature, This represents the number of valid feature samples from the same historical scenarios for this cyclist. This is the first time in history that this cyclist has been in the same situation. The first sample 3D eigenvalues; Formula for calculating the standard deviation of volatility:

[0020] In the formula, This is the first time in history that this cyclist has been in the same situation. Standard deviation of dimensional features; Deviation rate calculation formula:

[0021] In the formula, The overall deviation rate. The total dimension of the feature vectors. For the first successful match In the nth sample 3D eigenvalues It is a local minimum.

[0022] Furthermore, the determination logic of the security determination module is as follows: Fatigue riding judgment: When a decrease in grip strength is detected and lasts for a preset time of 3 seconds, accompanied by abnormal head deflection of ≥15° and lasting for a preset time of 2 seconds, and the correlation strength between the two reaches a preset threshold of 0.6, fatigue riding is judged, and the severity of fatigue riding is distinguished according to the degree of characteristic changes. Distracted driving determination: When an abnormal head deflection of ≥15° is detected and lasts for a preset time of 2 seconds, accompanied by a trajectory drift of ≥0.5m, and the correlation strength between the two reaches a preset threshold of 0.6, it is determined to be distracted driving. If a finger is detected to be off the handlebar at the same time, it is determined to be serious distracted driving. Judgment of illegal operation: When it is detected that one hand is off the handlebars for a preset time of 1 second, or both hands are off the handlebars for a preset time of 0.5 seconds, regardless of whether it is accompanied by track drift, it is judged as illegal operation; when it is detected that track drift is ≥0.5m and lasts for a preset time of 2 seconds, and no abnormal head turning or grip loss is detected, it is judged as illegal lane change or unstable riding. The adaptive threshold dynamic adjustment model takes the cyclist's historical cycling data, current cycling speed, and real-time road conditions as input features, and outputs a scenario-adaptive judgment threshold through feature mapping relationship. When the cycling speed is ≥20km / h, the trajectory offset threshold is adjusted to 0.3m.

[0023] Furthermore, the warning signal output by the warning execution module is divided into three forms: visual warning, auditory warning, and tactile warning. Each warning form can be output individually or in combination. The warning device is set at the handlebars and the front of the electric bicycle. For minor danger, only tactile warning is output; for moderate danger, tactile and auditory warnings are output; and for severe danger, tactile, auditory, and visual warnings are output. In the event that the rider is in a severely dangerous situation and does not respond to the warning signal within 3 seconds, the electric power steering control signal and / or micro electromagnetic braking control signal connected to the system are triggered. The electric power steering angle is ≤5° and the maximum is no more than 10°. The micro electromagnetic braking deceleration is ≤5km / h and the maximum is no more than 10km / h per cycle, so as to assist in the control of the electric bicycle's direction of travel and / or speed. Set intervention execution boundaries and limit parameters, establish failure protection mechanisms for sensors, communication, and actuators, and immediately stop intervention and output fault warning when any module fails; define prohibited intervention scenarios such as riding speed ≥25km / h and rider's active strong control, and achieve secondary risk avoidance by real-time detection of rider's control behavior; When the system detects that the rider has resumed normal control for 2 seconds, all riding intervention operations will immediately cease, and the normal control of the electric bicycle will be restored.

[0024] Beneficial effects Compared with existing technologies, this non-motorized vehicle riding safety protection system based on multi-sensor fusion has the following beneficial effects: I. This invention utilizes a multi-sensor fusion architecture to achieve multi-dimensional data collection of non-motorized vehicle riding data. Combined with differentiated collection frequencies and data verification mechanisms, it ensures the effectiveness of data transmission. A professional preprocessing process aligns the temporal sequence of multi-source data, enabling more comprehensive and accurate acquisition of rider control behavior and status data. Based on sample matching and temporal correlation recognition models using a behavioral feature library, it mines the temporal correlations between multiple sets of dangerous features, achieving accurate determination of riding scenarios. Simultaneously, it combines historical baseline analysis of the same scenario to verify the matching results, making the identification of riding states more closely aligned with actual riding scenarios. This significantly improves the accuracy and relevance of identifying dangerous states such as fatigued riding and distracted driving, forming a closed-loop precision processing system from data collection to state recognition.

[0025] Second, this invention achieves scenario-based adaptation of judgment thresholds by configuring an adaptive threshold dynamic adjustment model. This allows the judgment criteria for dangerous cycling conditions to flexibly change according to actual cycling situations, improving the rationality of the judgment results. Based on a tiered early warning mechanism, it outputs multiple forms of early warning signals, combined with cycling intervention operations to achieve timely responses to dangerous situations. Simultaneously, it sets up a failure protection mechanism and prohibited intervention scenarios, defining intervention execution boundaries and limiting parameters, ensuring the effectiveness of intervention while avoiding secondary risks caused by excessive intervention. Furthermore, the behavioral feature database supports online updates, continuously optimizing the accuracy of sample matching, allowing the system's protective capabilities to continuously improve with use, providing a comprehensive, intelligent, and highly safe protection solution for non-motorized vehicle cycling.

[0026] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0028] Figure 1 This is a structural block diagram of a non-motorized vehicle riding safety protection system based on multi-sensor fusion. Detailed Implementation

[0029] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0030] Example 1: The non-motorized vehicle riding safety protection system based on multi-sensor fusion of the present invention consists of a sensor module, a data acquisition module, a data preprocessing module, a feature extraction module, a behavior feature library module, a temporal correlation recognition module, a safety judgment module, and a warning execution module, as follows: Figure 1 As shown, the sensor module and the data acquisition module are connected by signals. The data acquisition module, the data preprocessing module, and the feature extraction module are connected by signals in sequence. The feature extraction module is also connected by signals to the behavior feature library module and the temporal correlation recognition module. The temporal correlation recognition module is connected by signals to the safety judgment module. The safety judgment module is connected by signals to the early warning execution module. Each module performs its own function and works together to achieve full-process safety monitoring and protection of non-motorized vehicle riding, and completes accurate identification of riding status, early judgment of dangerous behavior, and graded early warning and intervention operations.

[0031] Sensor module: The sensor module is a combination of specially arranged grip force sensors, finger capacitive sensors, head posture sensors, and riding trajectory sensors. Through the coordinated work of multiple types of sensors, it achieves multi-dimensional and comprehensive data collection of the rider's control behavior, body posture, and vehicle riding status, providing raw data support for subsequent data processing and status recognition. The specific settings and working methods of each sensor are as follows: Handle grip force sensor: A piezoresistive force sensor is selected and fixed to the left and right grip areas of the electric bicycle handlebars in a symmetrical installation manner. At least two of these sensors are installed in each grip area. The sensors are evenly distributed at the main contact points of the palm in the grip area, which can detect the rider's grip force on the handlebars in real time, convert the mechanical detection signal into an electrical signal output, and accurately capture the real-time changes in grip force.

[0032] Finger capacitive sensor: A capacitive touch sensor is used, directly integrated into the surface of the handlebar grip area. It is configured in a one-to-one correspondence with the handlebar grip force sensor, and the sensor is flush with the handlebar grip surface, not interfering with the rider's normal grip operation. This sensor determines the contact or detachment state of the finger from the handlebar by detecting the change in capacitance when the finger contacts the sensor. Simultaneously, based on the amplitude range of the capacitance change and a preset calibration curve, it calculates the actual contact area between the finger and the handlebar, and outputs the corresponding capacitance change electrical signal.

[0033] Head posture sensor: Employing an inertial measurement unit integrating a gyroscope, accelerometer, and magnetometer, this waterproof and shockproof device is worn on the forehead of the cyclist's helmet. The sensor is fixedly connected to the helmet and kept horizontal, allowing it to move synchronously with the cyclist's head. Through the coordinated detection of the gyroscope, accelerometer, and magnetometer, it outputs real-time posture data in three dimensions: pitch angle, roll angle, and yaw angle, accurately reflecting the head's yaw angle, direction, and frequency of movement.

[0034] Cycling trajectory sensor: Utilizing a hardware architecture integrating GPS and an inertial measurement unit (IMU), the sensor is mounted in a fixed position on the front of the electric bicycle, ensuring its horizontal and stable operation. The GPS unit uses satellite positioning to collect real-time data on the rider's location, speed, and heading during the ride. The IMU serves as a supplementary data source; when GPS signals are interrupted due to building obstruction, tunnel travel, electromagnetic interference, or other factors, it immediately activates a dead reckoning algorithm. Combining this with data such as the vehicle's angular velocity and acceleration, it calculates the rider's real-time position and motion state, ensuring the continuity and integrity of the cycling trajectory data.

[0035] Data acquisition module: The data acquisition module has the functions of multi-channel signal reception, frequency adaptation and data transmission. It adapts to the output signal type of each sensor in the sensor module, establishes a stable signal connection with each sensor through wired communication, receives the raw electrical signals and data information output by each sensor in real time, and performs preliminary format conversion on the received data, converting different formats of analog signals and digital signals into a standardized digital signal format for easy processing by subsequent modules.

[0036] This module configures the acquisition frequency differently based on the detection characteristics of each sensor. The grip force sensor and finger capacitive sensor are configured with a high acquisition frequency to meet the detection requirements of rapid changes in grip force and finger contact state. The head posture sensor and cycling trajectory sensor are configured with a low acquisition frequency to match the rate of change in head posture and cycling trajectory, ensuring detection accuracy while reducing data redundancy. The data acquisition module incorporates a data verification mechanism, employing data frame verification and retransmission mechanisms to perform real-time verification of data during transmission, promptly identifying and correcting data transmission errors to ensure the effectiveness and accuracy of data transmission. After format conversion and data verification, the data acquisition module transmits the acquired raw data to the data preprocessing module in real-time without loss.

[0037] Data preprocessing module: The data preprocessing module receives the raw data transmitted from the data acquisition module and performs three processes in sequence: denoising, normalization, and time alignment. This eliminates environmental interference, dimensional differences, and time axis deviations in the raw data, resulting in standardized multi-source sensor data. This provides a high-quality, highly consistent data source for the subsequent feature extraction module. The specific implementation methods for each processing step are as follows: Denoising Processing: Wavelet transform algorithm is used to denoise the raw data from each sensor. Based on the different noise characteristics of the grip force sensor, finger capacitance sensor, head posture sensor, and riding trajectory sensor, appropriate wavelet basis functions and wavelet decomposition levels are selected to decompose the raw data into noise signal components and effective signal components. Threshold quantization is used to suppress and eliminate noise signal components, retaining effective signal components that can truly reflect the riding state. Wavelet reconstruction is then performed on the processed signals to obtain clean data after removing environmental vibration, electromagnetic interference, and signal jitter.

[0038] Normalization processing: The minimum-maximum normalization method is adopted to map the effective data of each sensor after noise reduction to the preset numerical range of [0, 1]. By unifying the numerical range, the dimensional and order-of-magnitude differences between different sensor data are eliminated, making different types of data such as grip force, capacitance change, head posture angle, and trajectory offset comparable and fusionable.

[0039] Temporal alignment processing: A three-layer attention network is used to construct an association matrix for multi-source sensor data. The input dimension of this attention network is M×T, where M is the total number of features of the four sensors and T is 30 time steps. The network output dimension is an M×M feature association matrix. The network training objective is to minimize the temporal alignment error of the multi-source sensor data, and the mean squared error function is selected as the loss function. This association matrix accurately captures the spatial and temporal correlations between the data of each sensor. At the same time, a sliding window mechanism is used to achieve temporal alignment. The window length is set to 10s. Using the time axis of the high-frequency acquisition sensor as the reference, the data of the low-frequency acquisition sensor is interpolated and matched with the time axis. Finally, the sampling time of the four-source sensor data is uniformly aligned to the same time axis, resulting in standardized data with temporal synchronization.

[0040] Feature extraction module: The feature extraction module takes the standardized data output by the data preprocessing module as input and uses a feature extraction algorithm to extract feature parameters that reflect the riding state and have obvious recognizability and distinguishability from the standardized data of the four types of sensors. Finally, four core features are obtained: grip force features, finger capacitance features, head posture features, and riding trajectory features. The specific extraction content and methods of each feature are as follows: Grip force characteristics: Two feature parameters are extracted: grip force amplitude and grip force decay rate. The grip force amplitude is the real-time grip force value after normalization, which directly reflects the rider's grip strength on the handlebars. The grip force decay rate is the change in grip force amplitude per unit time, which is obtained by calculating the ratio of the difference in grip force amplitude between two adjacent time points to the time interval, reflecting the trend and speed of grip force change.

[0041] Finger capacitance characteristics: Two feature parameters are extracted: finger contact state and contact area. The finger contact state is determined as a binary state of contact or detachment based on the output signal of the capacitance sensor, which directly reflects whether the finger is in contact with the handlebar. The contact area is calculated by converting the capacitance change with a preset capacitance-contact area calibration relationship, which reflects the degree of contact between the finger and the handlebar.

[0042] Head posture features: Two feature parameters are extracted: head deflection angle and deflection frequency. The head deflection angle is the difference between the real-time collected head posture angle and the reference head posture angle during normal riding, reflecting the degree to which the head deviates from the normal riding posture. The deflection frequency is the number of times the head deflection angle exceeds the normal reference range per unit time, reflecting the frequency of abnormal head deflection.

[0043] Cycling trajectory features: Two feature parameters are extracted: trajectory offset and offset rate. The trajectory offset is the straight-line distance between the real-time cycling position and the normal cycling trajectory, which directly reflects the degree of deviation of the vehicle's driving trajectory. The offset rate is the change in trajectory offset per unit time, which is obtained by the ratio of the difference in trajectory offset between adjacent time nodes to the time interval, reflecting the speed and trend of trajectory drift.

[0044] Behavioral Feature Library Module: The behavior feature library module is used to build, store, and update the rider's control behavior feature library, providing basic samples and data support for sample matching and scene determination in the time-series correlation recognition module. Its implementation process includes four steps: sample collection, sample preprocessing, feature library establishment, and feature matching. The specific implementation methods for each step are as follows: Sample Collection: Cyclists of different ages and skill levels were recruited as experimental subjects, covering youth, middle-aged, and elderly individuals, as well as beginners, average cyclists, and experienced cyclists at various skill levels. Sample collection was conducted in both indoor simulated cycling scenarios and outdoor real-world cycling scenarios. The indoor simulated scenarios included different road conditions and environments such as flat roads, slopes, turns, congestion, rain, and nighttime. The outdoor real-world scenarios covered different road types, including urban non-motorized vehicle lanes, rural roads, mixed traffic roads, and urban expressway non-motorized vehicle lanes. Raw data from four sensor sources were collected from the subjects under four cycling scenarios: normal riding, fatigued riding, distracted riding, and illegal operation, ensuring the diversity, representativeness, and comprehensiveness of the collected samples.

[0045] Sample preprocessing: The collected sample data undergoes the same denoising, normalization, and time-series alignment processes as the real-time cycling data to eliminate interference factors and format differences in the sample data. Then, the four core features of the sample data are extracted through the feature extraction module mentioned above. Professional traffic engineering and sensor technicians manually annotate the extracted feature samples according to the collection scenario. The annotation information includes key information such as cycling scenario type, feature value range, and dangerous behavior level, forming standardized annotated samples.

[0046] Feature Library Establishment: The labeled feature samples are used to train a model, constructing a cyclist control behavior feature model. This model and the labeled samples are then stored together to form a cyclist control behavior feature library. An incremental network is introduced into the feature library to enable online updates. The feature library update trigger threshold is set at 1000 new valid samples. When this threshold is reached, the feature library update process is automatically triggered. During the update, the parameters of the backbone model are frozen, and only the incremental network is trained. After training, the output of the incremental network is fused with the backbone model to achieve dynamic iterative updates of the feature library, ensuring the timeliness and adaptability of the feature library samples.

[0047] Feature Matching: The feature library module is configured with an environmental noise detection unit and a dynamic weight allocation model to achieve adaptive adjustment of feature matching. The environmental noise detection unit detects the noise level of the cycling environment in real time and quantifies it into three levels: low, medium, and high. The dynamic weight allocation model adaptively updates the matching weights of each sensor feature based on the noise quantification results. At low noise levels, all sensor features are assigned equal matching weights to ensure comprehensive feature matching. At medium and high noise levels, the matching weights for the head posture sensor and cycling trajectory sensor, which have strong anti-interference capabilities, are increased to 0.35, while the matching weights for the grip force sensor and finger capacitive sensor, which are susceptible to environmental noise interference, are decreased to 0.15, improving the accuracy and anti-interference capability of feature matching.

[0048] Temporal correlation identification module: The temporal correlation recognition module receives real-time features from the feature extraction module and sample data from the behavioral feature library module. Through sample library matching, calculation of historical baselines and fluctuations for the same scenario, and DR deviation rate verification, it accurately determines the cycling scenario. Simultaneously, it mines the temporal correlation relationships among three sets of hazard features, ultimately accurately identifying the cyclist's current cycling state. The specific implementation method consists of three parts: temporal correlation mining, sample library matching, and DR deviation rate verification. Temporal correlation mining: A three-layer temporal association model is built based on LSTM. The model input dimension is N×T, where N is the total dimension of the four core feature classes and T is the time step size of 30. The model output is a 64-dimensional feature association vector, with 64 dimensions representing the hidden layer mapping dimension. The model training objective is to maximize the feature association recognition accuracy. The cross-entropy loss function is used, and its mathematical expression is:

[0049] In the formula, This represents the cross-entropy loss value. The total number of samples, The true label value of the sample.

[0050] This represents the predicted probability value for the sample.

[0051] For three groups of hazardous features—grip weakening and abnormal head deflection, single-hand disengagement and trajectory drift, and abnormal head deflection and trajectory drift—dedicated association mining branches were established for each group. Each branch is a combination structure of a single fully connected layer and an attention mechanism. The attention mechanism assigns dynamic weights (0-1) to features at different time points, increasing the weight of feature nodes with indicative association during the occurrence of hazardous behaviors to strengthen their role, and decreasing the weight of unrelated noisy feature nodes to weaken their interference. The association strength of the three groups of hazardous features was calculated using the Pearson correlation coefficient, with the mathematical expression as follows:

[0052] In the formula, The Pearson correlation coefficient is used. The first set of features Each sample value The sample mean of the first set of features. The second set of features Each sample value The sample mean of the second set of features. Given the number of feature samples, the preset association strength threshold δ is 0.6. When the calculated Pearson correlation coefficient is ≥0.6, it is determined that there is a significant temporal association between the features in this group.

[0053] Sample library matching: First, the features extracted in real time from the four sensor sources by the feature extraction module are serialized and normalized to eliminate dimensional differences and sequence disorder, constructing a real-time feature vector A with uniform dimensions. Then, sample feature vectors for four riding scenarios—normal riding, fatigued riding, distracted riding, and improper operation—are retrieved from the behavior feature library. The cosine similarity algorithm is used to calculate the real-time feature vector A and the feature vectors of each sample one by one. The similarity value, the mathematical expression for cosine similarity, is:

[0054] In the formula, Similarity is the cosine similarity value. This is the fusion feature vector extracted in real time from four sensor sources. For vectors The 3D eigenvalues For the first in the behavioral feature database Sample feature vectors for cycling scenarios For vectors The i-th eigenvalue, The total dimension of the feature vector is denoted as T. The preset similarity threshold T is 0.75. When the similarity value of a sample is ≥0.75, the sample is considered to have initially matched the real-time feature.

[0055] DR deviation rate verification: To adapt to the personalized control habits of different cyclists and improve the accuracy of scene determination, the DR deviation rate of initially matched samples is reviewed. First, feature data of the cyclist's historical data in the same scene are retrieved from the behavioral feature database. Then, the baseline values ​​and standard deviations of the fluctuations of each dimension of features in the historical same scene are calculated using a sliding window mechanism. Baseline value calculation formula:

[0056] In the formula, For the first time in the same historical scene Baseline value of dimensional feature, This represents the number of valid feature samples from the same historical scenarios for this cyclist. This is the first time in history that this cyclist has been in the same situation. The first sample 3D eigenvalues; Formula for calculating the standard deviation of volatility:

[0057] In the formula, This is the first time in history that this cyclist has been in the same situation. Standard deviation of dimensional features; Deviation rate calculation formula:

[0058] In the formula, The overall deviation rate. The total dimension of the feature vectors. For the first successful match In the nth sample 3D eigenvalues It is a local minimum.

[0059] Preset preference feature threshold If the threshold is 0.2, the calculated overall deviation rate DR is compared with this threshold. If DR≤0.2, it means that the deviation between the initial matched sample and the rider's historical operating habits is within an acceptable range, and the matching result is confirmed to be valid. If DR>0.2, the sample is discarded, and the sample with the second highest similarity is selected from the feature library again. The above sample matching and DR deviation rate verification process is repeated until the current riding scenario type is determined.

[0060] Security determination module: The safety assessment module, based on the cycling scene recognition results and feature temporal correlation relationships output by the temporal correlation recognition module, combined with preset assessment thresholds and logic, achieves early and accurate assessment of cycling fatigue, distracted driving, and illegal operation. Simultaneously, it classifies the danger level according to the degree of feature change and configures an adaptive threshold dynamic adjustment model to achieve scenario-based adaptation of the assessment thresholds, improving the adaptability and accuracy of the assessment results. The specific assessment logic and threshold adjustment method are as follows: Specific judgment logic: Fatigue cycling is determined as follows: when a decrease in grip strength is detected and this decrease lasts for a preset time of 3 seconds, accompanied by an abnormal head deflection of ≥15° for a preset time of 2 seconds, and the temporal correlation strength between the two reaches a preset threshold of 0.6, fatigue cycling is identified. The severity of fatigue cycling is differentiated based on the rate of grip strength decrease and the magnitude of the abnormal head deflection angle; the faster the decrease rate and the larger the deflection angle, the higher the severity of fatigue cycling.

[0061] Distracted driving determination: When an abnormal head deflection of ≥15° is detected for a preset time of 2 seconds, accompanied by a trajectory drift of ≥0.5m, and the temporal correlation strength between the two reaches a preset threshold of 0.6, it is determined to be distracted driving; if, under the above conditions, a finger detachment from the handlebars is detected at the same time, it is directly determined to be severe distracted driving.

[0062] Judgment of illegal control: When it is detected that one hand is off the handlebars for a preset time of 1 second, or both hands are off the handlebars for a preset time of 0.5 seconds, regardless of whether it is accompanied by track drift, it is directly judged as illegal control; when it is detected that track drift is ≥0.5m and lasts for a preset time of 2 seconds, and no abnormal head tilt or grip loss is detected, it is judged as illegal lane change or unstable riding, which falls under the category of illegal control.

[0063] Adaptive threshold dynamic adjustment: The adaptive threshold dynamic adjustment model takes the cyclist's historical riding data, current riding speed, and real-time road conditions as input features. Through feature mapping relationships, it outputs scenario-adapted judgment thresholds to achieve dynamic adjustment of the judgment thresholds. The core adjustment rules are as follows: based on the cyclist's historical riding data, it adapts the feature benchmark threshold to match the cyclist's personalized driving habits, reducing misjudgments caused by differences in personal habits; based on the complexity of real-time road conditions, it adjusts the judgment sensitivity of trajectory drift and head turn, increasing the judgment sensitivity for complex road conditions and appropriately relaxing it for simple road conditions; when the riding speed is ≥20km / h, the trajectory deviation threshold is adjusted from 0.5m to 0.3m to improve the sensitivity of danger judgment under high-speed riding conditions and promptly detect dangerous behaviors in high-speed riding.

[0064] Early warning execution module: The early warning execution module outputs corresponding graded early warning signals based on the judgment results and danger levels of the safety assessment module. When the cyclist is in a severely dangerous state and fails to respond in a timely manner, necessary cycling intervention operations are performed. At the same time, a failure protection mechanism and a scenario prohibiting intervention are set up to achieve dual protection of early warning and intervention, and avoid the occurrence of secondary risks. The specific implementation method includes three parts: graded early warning, cycling intervention, and safety protection. Tiered early warning: Warning signals are divided into three forms: visual warning, auditory warning, and tactile warning. Each warning form can be output individually or in combination. The warning devices are respectively installed at the handlebars and the front of the electric bicycle, and are output in a graded manner according to the level of danger. Mild hazard: Only provides tactile warning, using a vibration device at the handlebar grip to emit a slight, intermittent vibration to alert the rider without causing strong disturbance; Moderate danger: Outputs tactile and auditory warnings. The vibration device at the handlebar grip vibrates continuously, and the buzzer at the front of the bike emits intermittent auditory alarms, providing a double reminder to the rider to correct their riding posture. Severe danger: Outputs tactile, auditory, and visual warnings. The vibration device at the handlebars vibrates at full load, the buzzer continuously emits a high-frequency auditory alarm, and the warning light on the front of the vehicle flashes at a high frequency, achieving a triple-linked warning to alert the rider to the greatest extent possible.

[0065] Cycling intervention: In cases where a cyclist is in a severely dangerous situation and fails to respond to the warning signal within 3 seconds, the warning execution module triggers the electric power steering control signal and / or micro electromagnetic brake control signal connected to the system to perform appropriate cycling intervention. The intervention should be mild and moderate to avoid secondary accidents caused by excessive intervention. Electric power steering intervention: The electric power steering angle is ≤5° and the maximum is no more than 10°. By slightly adjusting the handlebar direction, the vehicle's driving trajectory is guided back to the normal range. Miniature electromagnetic braking intervention: Miniature electromagnetic braking reduces speed by ≤5km / h per cycle and not exceeding 10km / h per maximum. By gradually reducing the vehicle's speed, it reduces the risk of accidents caused by dangerous behavior.

[0066] When it is detected that the rider has returned to normal control and this state lasts for 2 seconds, all riding intervention operations will be stopped immediately, the normal control of the electric bicycle will be restored, and the control of the vehicle will be completely returned to the rider.

[0067] Safety precautions: The warning execution module sets strict intervention boundaries and limit parameters, strictly limiting the angle of electric power steering and the deceleration of electromagnetic braking to avoid over-range intervention. Simultaneously, a failure protection mechanism is established for sensors, communication, and actuators, monitoring the working status of each module in real time. If any module fails or malfunctions, all riding intervention operations are immediately stopped, and a fault warning signal is output to remind the rider to check the vehicle and system.

[0068] In addition, clearly defined scenarios where intervention is prohibited are established. When the system detects a cycling speed of ≥25km / h or strong active control by the cyclist, it will not perform any cycling intervention operations. By detecting the cyclist's control strength and actions in real time, it can achieve secondary risk avoidance and protect the cyclist's active control and cycling safety.

[0069] This invention relates to a non-motorized vehicle riding safety protection system based on multi-sensor fusion. It utilizes a variety of sensors arranged in a specific manner to collect riding data from multiple dimensions. Combined with differentiated data acquisition, multi-step preprocessing, and precise feature extraction techniques, the system ensures the validity and usability of the data. Furthermore, it uses temporal correlation recognition to accurately determine riding scenarios and hazardous characteristics. Combined with adaptive threshold adjustment safety judgment logic, it enables early identification of dangerous riding behaviors such as fatigue, distraction, and violations. Finally, through a dual protection system of tiered early warning and appropriate intervention, it promptly reminds riders to correct dangerous behaviors and, when necessary, reduces accident risks through minor interventions. This effectively improves the safety of non-motorized vehicle riding and reduces the incidence of non-motorized vehicle traffic accidents caused by fatigue riding, distracted driving, and improper operation.

[0070] Example 2: Detection, early warning, and intervention for fatigue cycling during daily urban road cycling.

[0071] This embodiment is applied to ordinary asphalt roads in the city. The rider is an experienced adult cyclist, driving an electric bicycle at a speed of 15 km / h straight in the non-motorized vehicle lane. This is a low-noise riding environment during the day, and all modules of the system are in normal working condition. The specific implementation process is as follows: The sensor module collects data as follows: Piezoresistive grip force sensors (two on each side) are symmetrically installed on the left and right grip areas of the handlebars to collect the grip force data of the rider's hands; capacitive finger sensors, which correspond one-to-one with the grip force sensors, collect the finger contact state and contact area data; a head posture sensor (integrating a gyroscope, accelerometer, and magnetometer) worn on the forehead of the helmet outputs three-dimensional posture data of head pitch angle, roll angle, and yaw angle; and a riding trajectory sensor with a GPS + inertial measurement unit fusion architecture installed on the front of the bike collects real-time data of riding position, speed, and heading via GPS. The four-source sensors complete the synchronous collection of multi-dimensional rider control behavior and riding status data.

[0072] Data acquisition module processing and transmission: The data acquisition module configures the acquisition frequency according to the different detection characteristics of each sensor, performs a data verification mechanism on the acquired raw data, and after confirming that the data transmission is without loss or error, transmits all raw data to the data preprocessing module.

[0073] The data preprocessing module performs standardization: First, it uses wavelet transform to denoise the original data. Then, it normalizes the denoised data to a preset numerical range of [0, 1]. The normalization mathematical expression is as follows:

[0074] In the formula, The original physical quantity of the sensor, This is the minimum value of the sample for this physical quantity. This represents the maximum value of the sample for this physical quantity. The normalized feature values ​​are then used to construct the correlation matrix using a 3-layer attention network. The input dimension is the total number of features from the four sensors × 30 time steps, and the output is an M×M feature correlation matrix. The temporal alignment error is minimized using the mean squared error function as the loss function. The mathematical expression of the mean squared error function is as follows:

[0075] In the formula, This is the mean squared error loss value. The total number of samples, For the actual value of timing alignment, The predicted values ​​are time-aligned, and time alignment is achieved through a 10-second sliding window mechanism to capture the spatial and temporal correlation of data from each sensor. The sampling times of the four sensors are aligned to the same time axis, and finally standardized data is obtained.

[0076] The feature extraction module extracts features: It accurately extracts four core features from the standardized data: grip force features (grip force amplitude, grip force decay rate), finger contact features (finger contact state, contact area), head posture features (head deflection angle, deflection frequency), and cycling trajectory features (trajectory offset, offset rate).

[0077] The behavior feature library module completes feature matching: The behavior feature library module retrieves stored control behavior feature samples from different cycling scenarios. Since the current environment is a low-noise level environment, the dynamic weight allocation model assigns equal matching weights to the features of each sensor, and performs preliminary matching between the features extracted in real time and the feature library samples.

[0078] The temporal correlation recognition module completes the scene and feature association determination: It serializes and normalizes the features extracted from the four sensor sources in real time, constructing a unified-dimensional real-time feature vector A. It then retrieves sample feature vectors from the behavior feature library for four types of cycling scenarios and calculates the similarity value using a cosine similarity algorithm. The mathematical expression for cosine similarity is:

[0079] In the formula, Similarity is the cosine similarity value. This is the fusion feature vector extracted in real time from four sensor sources. For vectors The 3D eigenvalues For the first in the behavioral feature database Sample feature vectors for cycling scenarios For vectors The i-th eigenvalue, The total number of dimensions of the feature vector is given. The similarity value between the real-time features and the fatigue cycling sample features reaches 0.82, which is higher than the similarity threshold of 0.75, thus completing the initial matching. Subsequently, historical feature data of the cyclist traveling straight on urban roads in the same scene are retrieved from the feature library. The baseline value and standard deviation of the fluctuation of each dimension feature are calculated through a sliding window mechanism. Then, the overall deviation rate between the initial matching sample and the historical features of the same scene is calculated through the deviation rate formula. The obtained deviation rate is 0.15, which is lower than the preset threshold of 0.2, confirming that the matching result is valid and determining that the current cycling scene is a fatigue cycling related scene. Simultaneously, this module, based on a 3-layer LSTM temporal correlation model, mines the temporal correlation between grip strength decay and abnormal head deflection through a dedicated correlation mining branch. It assigns dynamic weights (0-1) to features at different time points using an attention mechanism, strengthening the weights of indicative feature nodes. The Pearson correlation coefficient shows a correlation strength of 0.72, higher than the correlation strength threshold of 0.6, indicating a significant temporal correlation. The mathematical expression for the Pearson correlation coefficient is:

[0080] In the formula, The Pearson correlation coefficient is used. The first set of features Each sample value The sample mean of the first set of features. The second set of features Each sample value The sample mean of the second set of features. The number of feature samples.

[0081] The safety assessment module completes the fatigue riding assessment: Based on the temporal correlation recognition results and combined with the preset threshold, the safety assessment module detects that the rider's grip strength has significantly decreased and this state lasts for 3 seconds, accompanied by an abnormal head tilt angle of 18° lasting for 2 seconds. The correlation strength between the two reaches the preset threshold of 0.6, which is consistent with the fatigue riding assessment logic. Therefore, the rider is determined to be in a state of moderate fatigue riding. Since the current riding speed is 15km / h, the adaptive threshold dynamic adjustment model does not adjust the assessment threshold.

[0082] The warning execution module outputs tiered warnings and monitors responses: Based on the safety assessment result of moderate fatigue, the warning execution module outputs a combined tactile and auditory warning signal. Vibration is generated through the tactile device at the handlebar grip, and intermittent audible alerts are emitted from the auditory device at the front of the bicycle. The system monitors the rider's response to the warning signal in real time. If the rider perceives the warning within 2 seconds and adjusts their riding posture, with hand grip strength returning to normal and head posture returning to stability, the system detects this normal operating state and, after 2 seconds, immediately stops all warning signals and restores normal operating control of the electric bicycle. This completes the fatigue riding detection and warning process. In summary, this embodiment fully implements the entire operational logic of a non-motorized vehicle riding safety protection system based on multi-sensor fusion in a low-noise urban road environment where non-motorized vehicles travel straight. Riding data is collected from four sensors from multiple dimensions. After standardized processing including data acquisition and preprocessing, core features are extracted. Equal-weight feature matching is performed using a behavioral feature library. Then, a temporal correlation recognition module accurately determines the scene and mines temporal correlations with fatigue-related hazard features. Finally, a safety judgment module identifies moderate fatigue riding based on preset thresholds. After the warning execution module outputs a combined warning, the system terminates the warning and returns control of the vehicle because the rider responds promptly and resumes normal operation. This verifies the effectiveness and timeliness of the system's fatigue detection and warning in typical riding scenarios.

[0083] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A non-motorized vehicle riding safety protection system based on multi-sensor fusion, characterized in that, It includes a sensor module, a data acquisition module, a data preprocessing module, and a feature extraction module connected in sequence. The feature extraction module is connected to a behavior feature library module and a temporal correlation recognition module, respectively. The temporal correlation recognition module is connected to a security determination module, and the security determination module is connected to an early warning execution module. The sensor module consists of a specially arranged grip force sensor, finger capacitive sensor, head posture sensor, and riding trajectory sensor, used to collect multi-dimensional data on the rider's control behavior and riding status. The data acquisition module is configured with different acquisition frequencies according to the sensor detection characteristics, and a data verification mechanism is set up to ensure the validity of data transmission. The acquired raw data is then transmitted to the data preprocessing module. The data preprocessing module performs noise reduction, normalization, and time alignment on the raw data to obtain standardized data. Time alignment is used to unify the time axis of the data from the four sources of sensors. The feature extraction module extracts grip force features, finger capacitance features, head posture features, and cycling trajectory features from standardized data; The behavior feature library module stores control behavior feature samples under different cycling scenarios, and establishes a rider control behavior feature library that supports online updates. The temporal correlation recognition module completes accurate scene determination through sample library matching, historical same scene baseline and fluctuation calculation, and DR deviation rate verification. At the same time, it explores the temporal correlation between three sets of dangerous features: grip force decay and abnormal head deflection, single hand detachment and trajectory drift, and abnormal head deflection and trajectory drift. It matches the features extracted in real time with samples in the behavioral feature library to identify the rider's current riding status. The safety determination module uses the temporal correlation recognition results and preset thresholds to make early determinations of cycling fatigue, distracted driving and illegal operation. It also configures an adaptive threshold dynamic adjustment model to achieve scenario-based adaptation of the determination threshold. The early warning execution module outputs corresponding graded early warning signals based on the safety assessment results, and performs cycling intervention operations when necessary.

2. The non-motorized vehicle riding safety protection system based on multi-sensor fusion according to claim 1, characterized in that, The grip force sensors are piezoresistive force sensors, symmetrically installed on the left and right grip areas of the electric bicycle handlebars, with at least two sensors installed in each grip area. The finger capacitive sensors are capacitive touch sensors integrated into the surface of the grip area and are configured one-to-one with the grip force sensors. The head posture sensor is an inertial measurement unit integrating a gyroscope, accelerometer, and magnetometer, worn on the forehead of the rider's helmet, outputting three-dimensional posture data of head pitch, roll, and yaw angles. The riding trajectory sensor adopts a GPS and inertial measurement unit fusion architecture, installed at the front of the electric bicycle. The GPS unit collects riding position, speed, and heading data, while the inertial measurement unit performs dead reckoning when the GPS signal is interrupted.

3. The non-motorized vehicle riding safety protection system based on multi-sensor fusion according to claim 1, characterized in that, The data preprocessing module uses wavelet transform to denoise the original data and normalizes the data to a preset value range of [0, 1]. The data preprocessing module uses a 3-layer attention network to construct an association matrix with an input dimension of M×T, where M is the total number of features of the four source sensors and T is 30 time steps. The output dimension is an M×M feature association matrix. The network training objective is to minimize the temporal alignment error of the multi-source sensor data, and the loss function is the mean square error function. The temporal alignment is achieved using a sliding window mechanism with a window length of 10s. The spatial and temporal correlation of each sensor data is captured through the association matrix, and the sampling times of the four source sensors are uniformly aligned to the same time axis.

4. The non-motorized vehicle riding safety protection system based on multi-sensor fusion according to claim 1, characterized in that, The grip force characteristics include grip force amplitude and grip force decay rate; the finger contact characteristics include finger contact state and contact area; the head posture characteristics include head deflection angle and deflection frequency; and the cycling trajectory characteristics include trajectory offset and offset rate.

5. The non-motorized vehicle riding safety protection system based on multi-sensor fusion according to claim 1, characterized in that, The process of establishing the cyclist control behavior feature database module includes sample collection, sample preprocessing, feature database establishment and feature matching, and the feature database update trigger threshold is when the number of newly added valid samples reaches 1000. Sample collection: Recruit cyclists of different ages and cycling proficiency levels, and collect data from four sources of sensors in four scenarios: normal cycling, fatigued cycling, distracted driving, and illegal operation, in indoor simulated scenarios and outdoor real cycling scenarios. Sample preprocessing: The collected sample data is denoised, normalized, and time-aligned. After feature extraction, the sample is labeled. Feature library establishment: Train the labeled feature samples to build and store the manipulation behavior feature model. Incremental network is introduced to realize online update of feature library. Incremental learning is triggered by a preset threshold of 1000 new effective samples. During update, the backbone model is frozen and only the incremental network is trained and its output is fused with the backbone model. Feature matching: Configure an environmental noise detection unit and a dynamic weight allocation model. The environmental noise detection unit quantifies the cycling environment noise into three levels: low, medium, and high. The dynamic weight allocation model adaptively updates the matching weights of each sensor feature based on the noise quantization results. At the low noise level, the matching weights of each sensor feature are equal. At the medium and high noise levels, the feature matching weights of the head posture sensor and cycling trajectory sensor are increased to 0.35, while the feature matching weights of the grip force sensor and finger capacitance sensor are reduced to 0.

15.

6. The non-motorized vehicle riding safety protection system based on multi-sensor fusion according to claim 1, characterized in that, The temporal association recognition module is based on an LSTM-based 3-layer temporal association model. The input dimension is N×T, where N is the total dimension of the four core features and T is 30 time steps. The output dimension is a D-dimensional feature association vector, where D is the 64-dimensional hidden layer mapping dimension. The model training objective is to maximize the feature association recognition accuracy, and the loss function is the cross-entropy loss function. The temporal correlation recognition module sets up dedicated correlation mining branches for three groups of dangerous features: grip force attenuation and abnormal head deflection, single hand disengagement and trajectory drift, and abnormal head deflection and trajectory drift. Each branch is a combination structure of a single fully connected layer and an attention mechanism. The attention mechanism assigns dynamic weights of 0-1 to features at different time points, strengthening the weights of feature nodes that have correlation indicativeness during the occurrence of dangerous behavior and weakening unrelated noisy feature nodes. The correlation strength of the three groups of dangerous features is calculated using the Pearson correlation coefficient, and the correlation strength threshold δ is set to 0.

6. When the Pearson correlation coefficient is ≥0.6, it is determined that there is a significant temporal correlation between the features.

7. The non-motorized vehicle riding safety protection system based on multi-sensor fusion according to claim 1, characterized in that, The temporal correlation recognition module matches the real-time extracted features with samples in the behavior feature library as follows: The real-time extracted features from the four sensor sources are serialized and normalized to construct a unified-dimensional real-time feature vector A; the sample feature vectors of the four cycling scenarios stored in the behavior feature library are then retrieved. The cosine similarity algorithm is used to calculate the similarity value between the real-time feature vector and the feature vector of each sample. The similarity threshold T is set to 0.

75. When the similarity value of the sample is ≥0.75, it is determined that the sample has initially matched the real-time feature.

8. The non-motorized vehicle riding safety protection system based on multi-sensor fusion according to claim 1, characterized in that, The temporal correlation recognition module retrieves feature data from historical similar scenarios in the feature library, calculates the baseline values ​​and standard deviations of the fluctuations of each dimension of features under historical similar scenarios using a sliding window mechanism, calculates the overall deviation rate between the initially successfully matched samples and the features of historical similar scenarios based on the deviation rate formula, and compares the overall deviation rate with a preset preference feature threshold. contrast, Set to 0.2, if Then the matching result is confirmed to be valid. Then discard the sample and re-screen until the current cycling scenario type is determined; Baseline value calculation formula: In the formula, For the first time in the same historical scene Baseline value of dimensional feature, This represents the number of valid feature samples from the same historical scenarios for this cyclist. This is the first time in history that this cyclist has been in the same situation. The first sample 3D eigenvalues; Formula for calculating the standard deviation of volatility: In the formula, This is the first time in history that this cyclist has been in the same situation. Standard deviation of dimensional features; Deviation rate calculation formula: In the formula, The overall deviation rate. The total dimension of the feature vectors. For the first successful match In the nth sample 3D eigenvalues It is a local minimum.

9. The non-motorized vehicle riding safety protection system based on multi-sensor fusion according to claim 1, characterized in that, The determination logic of the security determination module is as follows: Fatigue riding judgment: When a decrease in grip strength is detected and lasts for a preset time of 3 seconds, accompanied by abnormal head deflection of ≥15° and lasting for a preset time of 2 seconds, and the correlation strength between the two reaches a preset threshold of 0.6, fatigue riding is judged, and the severity of fatigue riding is distinguished according to the degree of characteristic changes. Distracted driving determination: When an abnormal head deflection of ≥15° is detected and lasts for a preset time of 2 seconds, accompanied by a trajectory drift of ≥0.5m, and the correlation strength between the two reaches a preset threshold of 0.6, it is determined to be distracted driving. If a finger is detected to be off the handlebar at the same time, it is determined to be serious distracted driving. Unauthorized handling judgment: When it is detected that one hand is off the handlebars for a preset time of 1 second, or both hands are off the handlebars for a preset time of 0.5 seconds, regardless of whether it is accompanied by trajectory drift, it is judged as unauthorized handling; when it is detected that the trajectory drift is ≥0.5m and lasts for a preset time of 2 seconds, and no abnormal head turning or grip weakening is detected, it is judged as unauthorized lane change or unstable riding; The adaptive threshold dynamic adjustment model uses the rider's historical riding data, current riding speed, and real-time road conditions as input features, and outputs a scenario-adapted judgment threshold through feature mapping relationship. When the riding speed is ≥20km / h, the trajectory deviation threshold is adjusted to 0.3m.

10. The non-motorized vehicle riding safety protection system based on multi-sensor fusion according to claim 1, characterized in that, The warning signals output by the warning execution module are divided into three forms: visual warning, auditory warning, and tactile warning. Each warning form can be output individually or in combination. The warning device is set at the handlebars and the front of the electric bicycle. For minor danger, only tactile warning is output; for moderate danger, tactile and auditory warnings are output; and for severe danger, tactile, auditory, and visual warnings are output. In the event that the rider is in a severely dangerous situation and does not respond to the warning signal within 3 seconds, the electric power steering control signal and / or micro electromagnetic braking control signal connected to the system are triggered. The electric power steering angle is ≤5° and the maximum is no more than 10°. The micro electromagnetic braking deceleration is ≤5km / h and the maximum is no more than 10km / h per cycle, so as to assist in the control of the electric bicycle's direction of travel and / or speed. Set intervention execution boundaries and limit parameters, establish failure protection mechanisms for sensors, communication, and actuators, and immediately stop intervention and output fault warning when any module fails; When the system detects that the rider has resumed normal control for 2 seconds, all riding intervention operations will immediately cease, and the normal control of the electric bicycle will be restored.

Images