Intelligent helmet based on two-stage cooperative collision detection and control method

Abstract

The application belongs to the technical field of intelligent wearable devices, and particularly relates to an intelligent helmet based on two-stage cooperative collision detection and a control method, which comprises a data acquisition unit, a main control unit and a hierarchical early warning unit arranged on the helmet. The data acquisition unit is used for collecting real-time three-axis acceleration information of the head of a wearer, posture angle information of the helmet body relative to an initial reference angle, and real-time acquisition of position information and speed information of the current position of the helmet. The main control unit is configured to adopt the data collected by the data acquisition unit, perform two-stage cooperative collision detection, and output a hierarchical early warning signal according to the two-stage cooperative collision detection, wherein the two-stage cooperative collision detection comprises first-stage waveform pre-screening and second-stage multi-dimensional scoring decision. The hierarchical early warning unit is used for performing a three-stage differentiated processing scheme according to the hierarchical early warning signal output by the main control unit. Real collisions and road bumps can be effectively distinguished, and the false positive rate is significantly reduced.

Description

Technical Field

[0001] This application belongs to the field of smart wearable device technology, specifically relating to a smart helmet and control method based on two-level cooperative collision detection. Background Technology

[0002] With the increasing popularity of green travel, electric bicycles and bicycle riding safety have become a focus of social attention. As an important piece of equipment for riding safety, the market demand for smart helmets continues to expand.

[0003] The existing collision detection of smart helmets mostly adopts the method of comparing acceleration thresholds at a single moment, which has the following technical defects: (1) Periodic vibrations generated by bumpy roads such as speed bumps and gravel roads are prone to frequent false triggers; (2) There are significant differences in waveform morphology between real collisions and road bumps. Real collisions are manifested as a single huge pulse followed by rapid decay, while road bumps are manifested as a continuous periodic pulse group. However, the existing technology only focuses on the amplitude and ignores the key distinguishing information of waveform morphology; (3) Only outputs a binary result of safety / danger, which cannot support graded warning strategies, resulting in excessive interference to users when there are minor abnormalities and insufficient response when there are serious accidents.

[0004] Therefore, how to design a smart helmet that can effectively distinguish between real collisions and road bumps and achieve graded warnings is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] This application provides an intelligent helmet and control method based on two-level collaborative collision detection. By performing first-level waveform pre-screening and second-level multi-dimensional scoring decision-making, it effectively distinguishes between real collisions and road bumps, reduces the false alarm rate, and achieves a transformation from passive protection to active safety through three-level differentiated early warning processing, significantly improving emergency rescue efficiency.

[0006] The first aspect of this application provides a smart helmet based on two-level cooperative collision detection, comprising: The data acquisition unit and main control unit are mounted on the helmet. The data acquisition unit is used to collect in real time the three-axis acceleration information of the wearer's head, the attitude angle information of the helmet body relative to the initial reference angle, and to obtain the current position and velocity information of the helmet in real time.

[0007] The main control unit is configured to use the data collected by the data acquisition unit to perform two-level collaborative collision detection, which includes a first-level waveform pre-screening and a second-level multi-dimensional scoring decision, and output a graded early warning signal based on the two-level collaborative collision detection. The graded early warning unit is used to execute a three-level differentiated processing scheme based on the graded early warning signal output by the main control unit.

[0008] Furthermore, the first-level waveform pre-screening includes: Triaxial acceleration data is continuously collected at a preset sampling frequency. When the triaxial acceleration magnitude exceeds a preset threshold, acceleration waveform segments within a preset time window before and after the trigger moment are extracted. The acceleration waveform segment is preprocessed to extract time-domain features, including pulse width, rise time, and decay rate, to form a feature vector; The feature vector is matched with a pre-set accident waveform feature library and a turbulence waveform feature library using weighted Mahalanobis distance to output an accident type label; the event type label includes "collision candidate", "turbulence interception" and "uncertain".

[0009] Furthermore, the feature vector is matched with a pre-set accident waveform feature library and a turbulence waveform feature library using weighted Mahalanobis distance to output an accident type label, including: The main control unit has two pre-set waveform feature template libraries, including an accident waveform feature library and a turbulence waveform feature library. Each waveform feature template library contains a standard feature vector and statistical standard deviation for the corresponding event type. Based on the standard feature vector and statistical standard deviation, the matching distance between the real-time feature vector and each waveform feature template library is calculated using weighted Mahalanobis distance; The first-level event type label is output based on the matching distance, including: when the accident matching distance is less than the bump matching distance and the accident matching distance is less than 1.5, it is determined as "collision candidate"; when the bump matching distance is less than the accident matching distance and the bump matching distance is less than 1.5, it is determined as "bump interception". If neither of the above two conditions is met, it is judged as "uncertain".

[0010] Furthermore, the second-level multidimensional scoring decision includes: When the event type label is "collision candidate" or "uncertain", calculate the impact intensity score, attitude anomaly score, post-impact stillness score and velocity change rate score. The weighting coefficients of each score are adjusted according to the event type label: the default weight is maintained when the label is "collision candidate"; the weighting coefficient of the impact intensity score is increased when the label is "uncertain". The four modulated dimensions of scores are weighted and fused, and after time-series verification, a comprehensive confidence score is obtained and a graded early warning signal is output.

[0011] Furthermore, the impact strength score, attitude anomaly score, post-impact stillness score, and velocity change rate score are calculated, including: calculating the impact strength score. : , in, For acceleration magnitude, It is the acceleration due to gravity; Posture abnormality score : , where θ is the current tilt angle measured by the angle sensor; Post-impact static rating Based on the variance of acceleration modulus within 2 seconds after the impact. ,when hour =1, when hour =0, intermediate linear interpolation; Rate of change of velocity rating : ,in The average deceleration within 2 seconds before and after the impact.

[0012] Furthermore, the tiered early warning signal includes: When the overall confidence level C ≥ 0.8 and the event type label is "collision candidate", a level 1 emergency warning signal is output. The secondary warning signal is output when the event type label is "collision candidate" and 0.6≤C<0.8 or the event type label is "uncertain" and C≥0.7; The no-warning signal is output when the event type label is "collision candidate" and C < 0.6, or the event type label is "uncertain" and C < 0.7, or the event type label is "bump interception".

[0013] Furthermore, the main control unit calculates the comprehensive confidence score at intervals. Each calculation is based on the latest sensor data at the calculation time. The comprehensive confidence score is collected continuously for n scoring periods, and the time-weighted confidence score is calculated using an exponentially weighted moving average. Based on the first-level event type label and time-weighted confidence, the main control unit outputs a graded early warning signal.

[0014] A second aspect of this application provides a control method for a smart helmet based on two-level cooperative collision detection, including: It can collect the wearer's three-axis acceleration information, the attitude angle information of the helmet body relative to the initial reference angle, and the current position and velocity information of the smart helmet in real time. The collected data is used to perform a two-level collaborative collision detection, which includes a first-level waveform pre-screening and a second-level multi-dimensional scoring decision. Based on the results of the two-level collaborative collision detection, a graded early warning signal is output, and a three-level differentiated processing scheme is executed according to the graded early warning signal.

[0015] Compared with the prior art, the advantages of this application are as follows: By performing time-domain feature analysis and pattern matching on acceleration waveforms through first-level waveform pre-screening, the system can effectively distinguish between real collisions (single pulse fast decay) and road bumps (periodic pulse groups). The first level can intercept most bump events, significantly reducing the false alarm rate.

[0016] The second-level multidimensional scoring decision integrates scores from four dimensions: impact intensity, attitude anomaly, post-impact staticity, and velocity change rate. A time-series verification mechanism is also introduced to improve the accuracy and reliability of collision judgment.

[0017] By employing a parameter modulation mechanism between the two levels (dynamically adjusting the weight coefficients and trigger thresholds of the second level based on the event type labels of the first level), adaptive decision-making is achieved, balancing high recall and low false positive rate.

[0018] By implementing a three-tiered differentiated processing scheme, a complete early warning chain is achieved from emergency assistance to continued monitoring, avoiding excessive interference to users in the event of minor anomalies, while ensuring timely response in the event of serious incidents.

[0019] In the event of a severe collision, an emergency assistance process is automatically triggered, and location and speed information are uploaded via the communication module to improve the timeliness of rescue. Attached Figure Description

[0020] The accompanying drawings are used to provide a further understanding of the technical solutions of this application and constitute a part of the specification. They are used together with the embodiments of this application to explain the technical solutions of this application and do not constitute a limitation on the technical solutions of this application.

[0021] Figure 1 This is a block diagram of the data acquisition unit, main control unit, and graded early warning unit installed on the helmet according to an embodiment of this application; Figure 2 This is a flowchart illustrating the operation of a helmet provided in an embodiment of this application; Figure 3 This is a flowchart of the first-level waveform pre-screening process provided in the embodiments of this application; Figure 4 This is a flowchart of the second-level multidimensional scoring decision provided in the embodiments of this application; Figure 5 This is a flowchart of a control method for a smart helmet based on two-level cooperative collision detection, provided in an embodiment of this application. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0023] See Figure 1 As shown in the figure, an embodiment of this application provides a smart helmet based on two-level cooperative collision detection, including a data acquisition unit, a main control unit, and a hierarchical early warning unit disposed on the helmet, wherein, The data acquisition unit is used to collect in real time the three-axis acceleration information of the wearer's head, the attitude angle information of the helmet body relative to the initial reference angle, and to obtain the current position and velocity information of the helmet in real time.

[0024] The main control unit is configured to use the data collected by the data acquisition unit to perform two-level collaborative collision detection, which includes a first-level waveform pre-screening and a second-level multi-dimensional scoring decision, and output a graded early warning signal based on the two-level collaborative collision detection. The graded early warning unit is used to execute a three-level differentiated processing scheme based on the graded early warning signal output by the main control unit.

[0025] The data acquisition unit includes an attitude perception module and a GPS module, electrically connected to the main control unit. The attitude perception module preferably uses an MPU6050, including an accelerometer and an angle sensor. The accelerometer is used to collect real-time triaxial acceleration information of the wearer's head; in this embodiment, a sampling frequency of 200Hz is used. The angle sensor is used to detect the helmet's attitude angle information relative to an initial reference angle, which in this embodiment is defined as the helmet's posture when the wearer is riding normally upright. The GPS module is preferably an ATGM336H module, used to acquire the current position and speed information of the smart helmet in real time; the sampling frequency is preferably 1Hz. The main control unit is configured to perform two-level cooperative collision detection. The main control unit integrates a microprocessor, memory, and non-volatile memory with a firmware program. In this embodiment, the main control unit preferably uses an ARM Cortex-M series microcontroller to balance computing performance and power consumption. The graded warning unit includes a voice broadcast module, a confirmation button module, and a communication module. The preferred voice broadcast module is the SYN6288 module, which is used to play voice prompts and alarm information. The confirmation button module is used for users to manually confirm and cancel after triggering an alarm. The communication module uses the Air724UG module, which is used to establish a data connection with a remote host computer.

[0026] See Figure 2As shown, after power-on, the main control unit initializes the parameters of each module and starts the data acquisition unit for real-time data acquisition. The accelerometer continuously outputs triaxial acceleration data at a frequency of 200Hz, the angle sensor continuously outputs attitude angle data, and the GPS module outputs velocity data at a frequency of 1Hz. The above data is transmitted to the main control unit via I²C bus or serial port. The main control unit preprocesses the acquired raw data, including: performing median filtering on the acceleration data to eliminate instantaneous noise, performing Kalman filtering on the angle data to obtain stable attitude angle estimates, and removing abnormal jumps in the GPS data. After completing the preprocessing, the main control unit executes a two-level cooperative collision detection: when the acceleration magnitude exceeds a preset threshold, the two-level cooperative collision detection process is triggered, as detailed below.

[0027] Figure 2 Combination Figure 4 As shown, the first-level waveform pre-screening includes: Triaxial acceleration data is continuously collected at a preset sampling frequency. When the triaxial acceleration magnitude exceeds a preset threshold, acceleration waveform segments within a preset time window before and after the trigger moment are extracted. The acceleration waveform segment is preprocessed to extract time-domain features, including pulse width, rise time, and decay rate, to form a feature vector; The feature vector is matched with a pre-set accident waveform feature library and a turbulence waveform feature library using weighted Mahalanobis distance to output an accident type label; the event type label includes "collision candidate", "turbulence interception" and "uncertain".

[0028] Specifically, the main control unit maintains a circular buffer in memory, caching the triaxial acceleration data of the most recent 1000 sampling points in a cyclic overwrite manner, with a corresponding set duration; in one example, 5 seconds is selected. The main control unit calculates the acceleration magnitude of each sampling point in real time. . For acceleration magnitude, , , These represent the acceleration components measured by the accelerometer along the X, Y, and Z axes, respectively. In one example, when the acceleration magnitude at a certain sampling point When the first threshold (preferably 2g in this embodiment, where g represents gravitational acceleration) is reached or exceeded, the sampling point is marked as a trigger point and enters the first-level waveform pre-screening stage. The main control unit then extracts a sequence of continuous sampling points from the circular buffer, centered on the trigger point, within a time window that extends backward by 100ms and forward by 300ms. Since the sampling interval is 5ms, this sequence contains a total of 80 sampling points, constituting a discrete signal segment for subsequent analysis.

[0029] The extracted waveform segments are preprocessed as follows: First, the mean of the waveform is subtracted to eliminate the influence of the gravitational component; then, a 4th-order Butterworth low-pass filter is used for filtering with a cutoff frequency of 30Hz to remove high-frequency noise and retain the collision frequency component; finally, the amplitude of the filtered waveform is normalized and scaled to the [0,1] interval to obtain a normalized waveform sequence. , n=0,1,...,79.

[0030] The following time-domain features are extracted from the normalized waveform sequence: Pulse width The shorter the pulse width, the more concentrated the impact, indicating the time interval between the rise from the 10% peak value to the peak value and then the drop to the 90% peak value. Ascent Time The time required for the peak value to rise from 10% to 90% is the indicator of the impact. The shorter the rise time, the more severe the impact. Attenuation rate ,in The average amplitude within a 100ms window after the peak point. The closer the decay rate is to 1, the faster the energy dissipates after the impact, which is consistent with the characteristics of a real collision. peak The maximum value of the waveform amplitude represents the instantaneous impact intensity of the collision. Oscillation count The number of times the waveform crosses zero after the peak point. In a real collision, there are usually only 1-2 zero crossings, while road bumps show multiple periodic oscillations. Energy Center of Gravity This reflects the concentrated position of waveform energy on the time axis. The more violent the collision, the more concentrated the energy is near the moment of impact, and the smaller the energy center of gravity. The above six features are combined into a feature vector. .

[0031] The main control unit has two pre-built waveform feature template libraries: an accident waveform feature library and a bump waveform feature library. Each waveform feature template library contains a standard feature vector and statistical standard deviation for the corresponding event type. In this embodiment, the above-mentioned waveform feature template libraries are pre-calibrated based on a large amount of real vehicle collision test data and measured data under different road conditions.

[0032] Specifically, the accident waveform feature library stores the standard values ​​of six time-domain features corresponding to collision accident events, as well as the statistical standard deviations of each of these six features. The bump waveform feature library stores the standard values ​​of six time-domain features corresponding to road bump events, as well as the statistical standard deviations of each of these six features. These six time-domain features are pulse width, rise time, decay rate, peak value, number of oscillations, and energy centroid.

[0033] Each element in the standard feature vector represents the statistical mean of the corresponding feature across a large number of samples, reflecting the typical shape of the waveform characteristics of this type of event. The statistical standard deviation reflects the degree of dispersion of each feature under this type of event and is used for normalization processing in subsequent weighted Mahalanobis distance calculation.

[0034] In this embodiment, the two waveform feature template libraries mentioned above are not arbitrarily set, but are obtained through pre-calibration based on a large amount of real vehicle collision test data and measured data under different road conditions. The specific calibration process is as follows.

[0035] Standard test dummies wearing smart helmets were used to conduct multi-condition crash tests in a compliant crash laboratory. The test conditions covered a variety of typical crash scenarios, including but not limited to frontal collisions, side collisions, rear collisions, and rollover falls. Acceleration waveforms were recorded at a sampling frequency of 200 Hz for each experiment.

[0036] Real-world road tests were conducted under various road conditions, covering a wide range of surface types encountered in daily riding, including urban roads, unpaved roads, mountain roads, and highways. Urban road conditions specifically included typical bump sources such as speed bumps, manhole covers, and road joints; unpaved roads included gravel roads, dirt roads, and washboard roads; mountain roads included continuous curves and steep, undulating sections; and highways included road expansion joints and bridge approach bounces. At least 200 valid bump waveform samples were collected for each road condition.

[0037] For each set of waveform samples collected, following the aforementioned preprocessing procedure and time-domain feature extraction method, the specific values ​​of the six time-domain features are calculated to form a set of sample feature vectors. For the collision sample set and the turbulence sample set, the arithmetic mean of each feature is calculated to obtain the standard value of each feature under this event type; at the same time, the standard deviation of each feature is calculated to obtain the statistical dispersion of each feature under this event type.

[0038] The specific calculation method is as follows: For a sample set under a certain event type, the sum of the values ​​of each feature in all samples is divided by the total number of samples to obtain the standard value of the feature; the sum of the squares of the differences between the feature value and the standard value of each sample is divided by the total number of samples minus one, and the square root is taken to obtain the standard deviation of the feature.

[0039] After calibration, the calculated standard feature vector and statistical standard deviation are stored as constant data in the main control unit's internal non-volatile memory, such as flash memory. During the first-level waveform pre-screening process, the main control unit directly reads these preset parameters to calculate the weighted Mahalanobis distance without online training or dynamic adjustment, thus ensuring the real-time performance and computational efficiency of collision detection.

[0040] The matching distance between the real-time feature vector and each template library is calculated using weighted Mahalanobis distance and expressed as follows: ， ， in, For the i-th component of the real-time feature vector, These are the standard values ​​of the i-th feature in the accident template library and the bump template library, respectively. and The corresponding standard deviation is... For accident matching distance, This represents the turbulence matching distance. The weight coefficients of each feature. Based on its importance in distinguishing between accidents and bumps, the preferred settings in this embodiment are: pulse width 0.25, rise time 0.20, decay rate 0.20, peak value 0.15, number of oscillations 0.15, and energy center of gravity 0.05.

[0041] Output the first-level event type label based on the matching distance: like and If the value is less than 1.5, output "collision candidate", indicating that the current waveform shape is close to the actual collision characteristics; like and If the value is less than 1.5, the output will be "Bump Interception", indicating that the current waveform matches the characteristics of road bumps. Otherwise, the output will be "uncertain", indicating that the waveform characteristics are unclear.

[0042] When a "bump interception" message is output, the system directly intercepts the event without further processing and returns to normal monitoring status. This mechanism effectively filters out periodic vibrations from bumpy road surfaces such as speed bumps and gravel roads, significantly reducing the false alarm rate.

[0043] See Figure 4 As shown, the second-level multidimensional scoring decision includes: When the event type label is "collision candidate" or "uncertain", calculate the impact intensity score, attitude anomaly score, post-impact stillness score and velocity change rate score. The weighting coefficients of each score are adjusted according to the event type label: the default weight is maintained when the label is "collision candidate"; the weighting coefficient of the impact intensity score is increased when the label is "uncertain". The four modulated dimensions of scores are weighted and fused, and after time-series verification, a comprehensive confidence score is obtained and a graded early warning signal is output.

[0044] In the second-level multidimensional scoring decision, the system first dynamically modulates the weight coefficients of each dimension's score based on the event type labels output from the first-level waveform pre-screening. Specifically: When the event type label of the first-level output is "collision candidate," it indicates that the temporal characteristics of the current waveform have matched the accident waveform template library well. At this time, the four-dimensional scores use the default weight configuration. The default weight configuration reflects the basic importance of each dimension in real collision judgment. Among them, the impact intensity score is the core indicator for judging the severity of the impact event, so it is given the highest weight; the posture abnormality score reflects whether the wearer has lost balance and is an important piece of evidence of common post-collision states, so it is given the second highest weight; the post-impact stillness score is used to determine whether the wearer has stopped moving after the impact, helping to confirm the abnormal state; and the velocity change rate score provides supplementary information from the perspective of kinematic changes.

[0045] When the event type label of the first-level output is "uncertain," it indicates that the time-domain characteristics of the current waveform are between accident waveforms and turbulence waveforms, making it difficult to make a clear judgment based solely on the waveform morphology. In this case, the system automatically increases the weighting coefficient of the impact intensity score while correspondingly decreasing the weighting coefficients of other dimensions. The technical consideration behind this is that when waveform characteristics are unclear, the absolute intensity of the acceleration modulus becomes the most reliable basis for judgment—if the impact intensity is large enough, even if the waveform morphology has some ambiguity, a higher-level warning should still be triggered; conversely, if the impact intensity is small, it tends to be judged as a state without warning.

[0046] After weighting, the system sums the scores of the four dimensions with their corresponding weight coefficients to obtain the comprehensive credibility score at the current time point. The specific calculation method is as follows: multiply the impact intensity score by the impact intensity weight, add the attitude anomaly score by the attitude anomaly weight, add the post-impact stillness score by the post-impact stillness weight, and add the velocity change rate score by the velocity change rate weight.

[0047] The overall credibility score ranges from 0 to 1. A higher score indicates that the detected event is more likely to be a real collision and requires an early warning response. A lower score indicates that the event is more likely to be a non-collision event or an event with insufficient impact intensity, requiring no early warning or only a low-level response.

[0048] To avoid misjudgments caused by single noise interference or transient anomalies, the system introduces a timing verification mechanism. The core idea of ​​this mechanism is that the acceleration waveform changes and attitude changes generated by real collision events often have a certain continuity, while accidental noise spikes or transient disturbances do not have this temporal continuity.

[0049] The specific implementation is as follows: The system continuously calculates the comprehensive confidence score at 5 time points, and the total length of the time window corresponding to these 5 time points is 1 second. Since the sampling frequency of the accelerometer is 200 Hz, the system calculates the comprehensive confidence score at regular time intervals, and the results of 5 consecutive collections constitute a time-series score sequence.

[0050] To strike a balance between smoothing noise and maintaining response sensitivity, the system employs an exponentially weighted moving average method to fuse the five scores. The basic principle of the exponentially weighted moving average is that scores closer to the current time are assigned higher weights, and scores farther from the current time are assigned lower weights. Specifically, in order of increasing time, the five scores are multiplied by 0.95 to the power of 4, 0.95 to the power of 3, 0.95 squared, 0.95 to the power of 1, and 0.95 to the power of 0, respectively, to obtain their respective weighting coefficients. The weighted scores are then summed and divided by the sum of all weighting coefficients to obtain the final time-weighted confidence score.

[0051] The system outputs tiered early warning signals based on a combination of event type labels and time-weighted confidence levels, according to preset rules. The specific judgment logic is as follows: When the event type label is "collision candidate" and the time-weighted confidence level is greater than or equal to 0.8, the system determines that a real collision has occurred and immediately outputs a Level 1 emergency warning signal.

[0052] When the event type label is "collision candidate" and the time-weighted confidence level is greater than or equal to 0.6 but less than 0.8, the system determines that there is a high suspicion of collision, but the confidence level has not yet reached the emergency threshold. At this time, a level 2 warning signal is output. At the same time, the system will activate the time window extension mechanism for time-series verification, extending the time window for subsequent time-series verification from 1 second to 3 seconds to obtain more time-series data for continuous observation.

[0053] When the event type label is "uncertain" and the time-weighted confidence level is greater than or equal to 0.7, the system determines that although the waveform shape is not typical, the confidence level is high after multi-dimensional score fusion, and there is still a possibility of collision. At this time, a level 2 warning signal is output, and the time window extension mechanism is also activated.

[0054] When the event type label is "bump interception", the system determines that the current event is a road bump rather than a collision accident and directly outputs no warning signal.

[0055] When the event type label is "collision candidate" but the time-weighted confidence level is less than 0.6, or when the event type label is "uncertain" but the time-weighted confidence level is less than 0.7, the system determines that the current impact intensity is insufficient or the characteristics deviate from the typical collision mode, and outputs no warning signal.

[0056] In one example, when the first-level output is "collision candidate" or "uncertain," the master control unit initiates the second-level analysis to calculate scores across four dimensions: (1) Impact strength score This non-linear design makes the scoring more reasonable in differentiating between impacts of varying intensities. ; (2) Posture abnormality score : , where θ is the current tilt angle measured by the angle sensor; (3) Post-impact staticity score Based on the variance of acceleration modulus within 2 seconds after the impact. ,when hour =1, when hour =0, intermediate linear interpolation. The smaller the variance, the closer the helmet is to a stationary state, the greater the likelihood of the rider losing mobility, and the higher the score; (4) Rate of change of velocity score : ,in The average deceleration over 2 seconds before and after the impact, expressed in m / s². This score reflects the overall severity of the collision and complements the impact intensity score.

[0057] Modulate the weighting coefficients based on the event type labels output from the first level: If the label is "collision candidate", maintain the default weight. =0.45, =0.30, =0.15, =0.10; If the label is "uncertain," the weight of the impact intensity score will be increased. =0.55, other values ​​adjusted accordingly. =0.30, =0.10, =0.05; When the first stage cannot be clearly determined from the waveform (output "uncertain"), the system relies more on the most direct indicator of impact intensity, while reducing its reliance on auxiliary indicators to avoid introducing more uncertainty.

[0058] The four modulated dimensions are then weighted and fused to obtain the overall credibility score: , To filter out transient fluctuations and single-point noise, the main control unit introduces a timing verification mechanism. The main control unit calculates the overall reliability score every 200ms. Each calculation is based on the latest sensor data at that moment. A comprehensive score is collected for five consecutive scoring periods, and an exponentially weighted moving average is used to calculate the time-weighted confidence level. , Based on first-level event type labels and time-series weighted confidence. The main control unit outputs tiered warning signals: when the time-weighted confidence level C ≥ 0.8 and the event type label is "collision candidate", a level 1 emergency warning signal is output; when (event type label is "collision candidate" and 0.6 ≤ C < 0.8) or (event type label is "uncertain" and C ≥ 0.7), a level 2 warning signal is output; when (event type label is "collision candidate" and C < 0.6) or (event type label is "uncertain" and C < 0.7) or (event type label is "bump interception"), no warning signal is output.

[0059] The graded early warning unit executes a three-level differentiated processing scheme based on the received early warning signal.

[0060] Specifically, upon receiving a Level 1 emergency warning signal, the main control unit immediately initiates a 10-second timer confirmation process. The voice broadcast module repeatedly plays the message, "Severe collision detected. Please press the confirmation button within 10 seconds to cancel the alarm," repeating every 2 seconds. If no confirmation signal is detected within 10 seconds, the main control unit determines that the user may have lost mobility and automatically triggers the emergency assistance process. Otherwise, the main control unit terminates the alarm process, and the system returns to normal detection status.

[0061] Upon receiving a Level 2 warning signal, the voice broadcast module reminds the user, "An abnormal state has been detected; please be aware of safety." Simultaneously, the main control unit activates a time-window extension mechanism for timing verification, extending the time window from 1 second to 3 seconds. Within this 3-second window, the main control unit continuously calculates the time-weighted confidence score C. If the time-weighted confidence score C continuously decreases and falls below the Level 2 warning trigger threshold, indicating that the abnormal state is easing, the main control unit stops the voice broadcast and returns to normal detection status. If the time-weighted confidence score C continuously increases and reaches the Level 1 emergency trigger condition, indicating that the situation is deteriorating, the main control unit upgrades the warning level to Level 1 emergency and executes the complete handling plan corresponding to Level 1 emergency.

[0062] When no warning signal is received, the system will not trigger any voice prompts, will not upload any data, and will immediately return to normal detection status.

[0063] Specifically, the emergency assistance process includes the GPS module acquiring the current geographical location information, the main control unit parsing the received NMEA format data to extract the current longitude and latitude coordinates, the main control unit uploading the speed information and GPS positioning information to the host computer monitoring system via the 4G communication module based on the TCP / IP protocol, and the host computer receiving the information parsing, displaying and storing the data to facilitate a rapid rescue response by relevant personnel.

[0064] Through the above implementation methods, this application employs a two-level collaborative collision detection mechanism. The first level utilizes waveform morphology analysis to quickly filter non-collision events such as road bumps, while the second level uses multi-dimensional feature fusion scoring for accurate evaluation. The two levels coordinate parameters through event type labels to form a collaborative decision-making process. In complex road conditions, the system can effectively distinguish between real collisions and road bumps, significantly reducing the false alarm rate. Simultaneously, through a tiered early warning mechanism, it achieves a shift from passive protection to proactive safety, significantly improving emergency rescue efficiency.

[0065] In another embodiment, see Figure 5 As shown, this application provides a control method for a smart helmet based on two-level cooperative collision detection, including: S101 collects real-time three-axis acceleration information of the wearer's head, attitude angle information of the helmet body relative to an initial reference angle, and real-time position and velocity information of the smart helmet. Specifically, it collects acceleration and attitude angle information through an attitude sensing module installed inside the helmet. The attitude sensing module includes an accelerometer and an angle sensor. The accelerometer continuously collects the acceleration components of the wearer's head in three orthogonal directions at a sampling frequency of 200 Hz, namely, acceleration in the X-axis direction, acceleration in the Y-axis direction, and acceleration in the Z-axis direction. The angle sensor is used to detect the attitude angle information of the helmet body relative to a preset initial reference angle, which includes one or more of pitch angle, roll angle, and yaw angle, to determine whether the wearer's head is in an abnormal tilt state.

[0066] Meanwhile, the system uses a GPS module mounted on the helmet to obtain real-time information about the helmet's location and speed. This information is used to calculate the rate of change of speed and to provide location data to the host computer when an emergency call is triggered.

[0067] S102, two-level collaborative collision detection is performed using the collected data. The two-level collaborative collision detection includes a first-level waveform pre-screening and a second-level multi-dimensional scoring decision. The first-level waveform pre-screening is used for fast pattern matching based on the time-domain characteristics of the acceleration waveform, and the second-level multi-dimensional scoring decision is used for multi-dimensional comprehensive evaluation based on the first-level output.

[0068] The specific process of the first-level waveform pre-screening is as follows: The main control unit maintains a circular buffer in memory to cache the most recent 1000 acceleration sampling points, corresponding to historical data for 5 seconds. The main control unit calculates the acceleration magnitude value of each sampling point in real time, which is the square root of the sum of the squares of the acceleration components in the three orthogonal directions. When the acceleration magnitude value reaches or exceeds twice the gravitational acceleration, the system enters the first-level waveform pre-screening process.

[0069] After triggering the screening, the system extracts acceleration waveform segments from the circular buffer, 100 milliseconds before and 300 milliseconds after the trigger moment, totaling 80 sampling points. These waveform segments are preprocessed: first, the arithmetic mean of the waveform is subtracted to eliminate the influence of the gravitational component; then, a fourth-order Butterworth low-pass filter is used, with a cutoff frequency set to 30 Hz, to filter out high-frequency noise while retaining the main frequency components of the collision; finally, the amplitude of the filtered waveform is normalized, linearly scaling the amplitude to the range of 0 to 1.

[0070] For the normalized waveform sequence, the system extracts six time-domain features: pulse width, rise time, decay rate, peak value, number of oscillations, and energy centroid. These six features are combined into a feature vector. Subsequently, the system uses weighted Mahalanobis distance to match the real-time feature vector with two internally pre-set waveform feature template libraries: an accident waveform feature library and a turbulence waveform feature library. Each template library contains a standard feature vector and statistical standard deviation for the corresponding event type. Based on the comparison results of the matching distance, the first level outputs an event type label of "collision candidate," "turbulence interception," or "uncertain." The second-level multidimensional scoring decision is triggered only when the event type label is "collision candidate" or "uncertain." At this level, the system calculates scores in four dimensions: impact intensity score, attitude anomaly score, post-impact stillness score, and velocity change rate score. Among them, the impact intensity score is calculated in a piecewise linear manner based on the numerical range of the acceleration modulus; the posture abnormality score is calculated using an exponential decay model based on the current tilt angle; the post-impact stillness score is calculated based on the variance of the acceleration modulus within 2 seconds after the impact, with a smaller variance indicating that the wearer is more likely to be still; and the velocity change rate score is calculated based on the average deceleration within 2 seconds before and after the impact.

[0071] Based on the event type labels output from the first level, the system modulates the weight coefficients of each dimension's score: the default weight configuration is used when the label is "collision candidate"; the weight coefficient of the impact intensity score is increased when the label is "uncertain". Then, the weighted scores of the four dimensions are weighted and fused to obtain the comprehensive credibility score at the current time point.

[0072] To avoid misjudgments caused by single-point noise, the system introduces a time-series verification mechanism: a comprehensive confidence score is calculated for five consecutive time points, corresponding to a one-second time window, and a time-series weighted confidence score is calculated using an exponentially weighted moving average method. Under specific conditions, the system will activate a time window extension mechanism, extending the verification window from one second to three seconds. Finally, based on the combination of event type label and time-series weighted confidence score, the system outputs a tiered early warning signal.

[0073] S103, output a graded early warning signal based on the result of the two-level cooperative collision detection, and execute a three-level differentiated processing scheme based on the graded early warning signal.

[0074] When a Level 1 emergency warning signal is output, the system determines that a real collision has occurred and immediately starts a 10-second timer. The system then prompts the user via a voice broadcast module to confirm the incident manually by pressing a confirmation button. If no confirmation signal is detected within the 10-second timer, the system automatically triggers an emergency assistance process. This includes sending a distress signal to the outside world via the voice broadcast module, obtaining the current location information via the GPS module, and transmitting the speed and location information to the host computer via the mobile communication network, enabling automatic alarm, data display, and storage. If a confirmation signal is detected within the 10-second timer, the system determines that the wearer is conscious and does not require external assistance, and returns to normal monitoring status.

[0075] When a Level 2 warning signal is output, the system determines that an abnormal state has been detected, but the severity has not yet reached the emergency level. The system then uses a voice broadcast module to remind the user, "An abnormal state has been detected; please be careful." Simultaneously, the main control unit activates a time-window extension mechanism for timing verification, extending the time window from one second to three seconds to continue monitoring the subsequent confidence level changes, so that the warning level can be upgraded promptly if the situation worsens.

[0076] When no warning signal is output, the system determines that the current event is a road bump, a minor impact, or noise interference. Without any response, the system returns to the normal detection state and continues to monitor subsequent data.

[0077] The intelligent helmet control method provided in this application significantly improves the accuracy of collision recognition while ensuring real-time detection through a two-level collaborative architecture of first-level waveform pre-screening and second-level multi-dimensional scoring decision-making. The first level employs lightweight temporal feature extraction and template matching, which can quickly filter out most road bump events. The second level introduces a multi-dimensional scoring fusion and temporal verification mechanism to perform refined evaluation of suspicious events. The two levels communicate and modulate weights through event type labels, forming a collaborative and complementary judgment logic. A three-level differentiated processing scheme executes different response strategies based on risk levels, ensuring timely alarm and assistance in case of severe collisions while avoiding frequent false alarms that disrupt user experience during daily use.

[0078] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A smart helmet based on two-level cooperative collision detection, characterized in that, include: The helmet includes a data acquisition unit, a main control unit, and a tiered early warning unit. The data acquisition unit is used to collect the wearer's three-axis acceleration information, the attitude angle information of the helmet body relative to the initial reference angle, and to obtain the helmet's current position and velocity information in real time. The main control unit is configured to use the data collected by the data acquisition unit to perform two-level collaborative collision detection, which includes a first-level waveform pre-screening and a second-level multi-dimensional scoring decision, and output a graded early warning signal based on the two-level collaborative collision detection. The graded early warning unit is used to execute a three-level differentiated processing scheme based on the graded early warning signal output by the main control unit.

2. The smart helmet based on two-level cooperative collision detection according to claim 1, characterized in that, The first-level waveform pre-screening includes: Triaxial acceleration data is continuously collected at a preset sampling frequency. When the triaxial acceleration magnitude exceeds a preset threshold, acceleration waveform segments within a preset time window before and after the trigger moment are extracted. The acceleration waveform segment is preprocessed to extract time-domain features, including pulse width, rise time, and decay rate, to form a feature vector; The feature vector is matched with a pre-set accident waveform feature library and a turbulence waveform feature library using weighted Mahalanobis distance to output an accident type label; the event type label includes "collision candidate", "turbulence interception" and "uncertain".

3. A smart helmet based on two-level cooperative collision detection according to claim 2, characterized in that, The feature vectors are matched with a pre-set accident waveform feature library and a turbulence waveform feature library using weighted Mahalanobis distance to output accident type labels, including: The main control unit has two pre-set waveform feature template libraries, including an accident waveform feature library and a turbulence waveform feature library. Each waveform feature template library contains a standard feature vector and statistical standard deviation for the corresponding event type. Based on the standard feature vector and statistical standard deviation, the matching distance between the real-time feature vector and each waveform feature template library is calculated using weighted Mahalanobis distance; The first-level event type label is output based on the matching distance, including: when the accident matching distance is less than the bump matching distance and the accident matching distance is less than 1.5, it is determined as "collision candidate"; when the bump matching distance is less than the accident matching distance and the bump matching distance is less than 1.5, it is determined as "bump interception". If neither of the above two conditions is met, it is judged as "uncertain".

4. A smart helmet based on two-level cooperative collision detection according to claim 2, characterized in that, The second-level multidimensional scoring decision includes: When the event type label is "collision candidate" or "uncertain", calculate the impact intensity score, attitude anomaly score, post-impact stillness score and velocity change rate score. The weighting coefficients of each score are adjusted according to the event type label: the default weight is maintained when the label is "collision candidate"; the weighting coefficient of the impact intensity score is increased when the label is "uncertain". The four modulated dimensions of scores are weighted and fused, and after time-series verification, a comprehensive confidence score is obtained and a graded early warning signal is output.

5. A smart helmet based on two-level cooperative collision detection according to claim 4, characterized in that, Calculation of impact strength score, attitude anomaly score, post-impact stillness score, and velocity change rate score includes: Calculation of impact strength score : ， in, For acceleration magnitude, It is the acceleration due to gravity; Posture abnormality score : , where θ is the current tilt angle measured by the angle sensor; Post-impact static rating Based on the variance of acceleration modulus within 2 seconds after the impact. ,when hour =1, when hour =0, intermediate linear interpolation; Rate of change of velocity rating : ,in The average deceleration within 2 seconds before and after the impact.

6. A smart helmet based on two-level cooperative collision detection according to claim 1, characterized in that, The tiered early warning signals include: When the time-weighted confidence level C ≥ 0.8 and the event type label is "collision candidate", a level 1 emergency warning signal is output. The secondary warning signal is output when the event type label is "collision candidate" and 0.6≤C<0.8 or the event type label is "uncertain" and C≥0.7; The no-warning signal is output when the event type label is "collision candidate" and C < 0.6, or the event type label is "uncertain" and C < 0.7, or the event type label is "bump interception".

7. A smart helmet based on two-level cooperative collision detection according to claim 5, characterized in that, The main control unit calculates the comprehensive confidence score at intervals. Each calculation is based on the latest sensor data at the calculation time. The comprehensive confidence score is collected continuously for n scoring periods, and the time-weighted confidence score is calculated using an exponentially weighted moving average. Based on the first-level event type label and time-weighted confidence, the main control unit outputs a graded early warning signal.

8. A control method for a smart helmet based on two-level cooperative collision detection, employing the smart helmet described in any one of claims 1-7, characterized in that, include: It can collect the wearer's three-axis acceleration information, the attitude angle information of the helmet body relative to the initial reference angle, and the current position and velocity information of the smart helmet in real time. The collected data is used to perform a two-level collaborative collision detection, which includes a first-level waveform pre-screening and a second-level multi-dimensional scoring decision. Based on the results of the two-level collaborative collision detection, a graded early warning signal is output, and a three-level differentiated processing scheme is executed according to the graded early warning signal.

Images