Collision prevention method for electric bicycle and electric bicycle

Abstract

The invention discloses an electric bicycle collision prevention method and an electric bicycle, and the method comprises the steps: obtaining an acceleration sequence and an angular velocity sequence of the electric bicycle, generating a motion feature sequence through frequency domain analysis, and calculating a time sequence drift prediction index; performing grading processing on multi-sensor sampling data according to the time sequence drift prediction index to generate a to-be-corrected data sequence; performing time window segmentation on the to-be-corrected data sequence, generating a time sequence correlation matrix based on local features, and calculating an optimal time sequence mapping relation; performing nonlinear interpolation and time sequence remapping according to the time sequence correction parameter to obtain a corrected sensor data sequence; and extracting target features based on the corrected data sequence, calculating a collision risk level and generating an early warning signal. According to the technical scheme, the problem of inconsistent time sequence of multi-sensor data of the electric bicycle can be solved, the influence of data drift on obstacle position prediction is reduced, and the reliability of an anti-collision system in a high-dynamic scene is guaranteed.

Description

Technical Field

[0001] The invention relates to the technical field of electric bicycle safety, and in particular to an electric bicycle collision prevention method and an electric bicycle. Background Art

[0002] Electric bicycles are a light vehicle powered by electricity and assisted by human pedaling. They are becoming more and more popular in urban travel due to their easy control, relatively low price, and moderate mileage. As the urban road environment becomes increasingly complex, electric bicycles frequently travel through streets with dense traffic and narrow space, facing high-risk situations such as frequent starts and stops, sharp turns, or competing with pedestrians. In order to protect the personal safety of riders and surrounding vehicles and pedestrians, some high-end electric bicycles have begun to be equipped with anti-collision functions, hoping to collect and analyze environmental information around the vehicle through sensors, and issue warnings in time or assist in avoidance when potential dangers occur.

[0003] At present, the anti-collision function for electric bicycles on the market mostly draws on the sensor fusion technology in the automotive field. The common solution is to use multiple sensors such as millimeter-wave radar and cameras to collaboratively build a map of the vehicle's surrounding environment, and predict possible obstacles based on data fusion. However, automotive solutions often rely on high-performance computing platforms and precise timing synchronization modules, which are difficult to apply to scenarios such as electric bicycles, which are smaller in size and more sensitive to cost and power consumption. Especially in high-dynamic conditions such as rapid start and stop and sharp turns of electric bicycles, inconsistent sampling frequencies of various sensors will cause data timing misalignment, which will lead to significant deviations in obstacle position prediction. For high-dynamic scenarios of electric bicycles with limited hardware resources, the existing technology lacks a lightweight timing correction mechanism, and it is difficult to effectively suppress the reduction in recognition and perception accuracy caused by data drift, which has become a core issue restricting the reliability and practicality of the anti-collision function of electric bicycles. Summary of the invention

[0004] The main purpose of the present invention is to solve the data drift problem caused by inconsistent timing of multi-sensor data in high-dynamic scenarios of electric bicycles.

[0005] A first aspect of the present invention provides an electric bicycle collision prevention method and an electric bicycle, wherein the electric bicycle collision prevention method and the electric bicycle include: Obtaining an acceleration sequence and an angular velocity sequence of the electric bicycle, generating a motion feature sequence through frequency domain analysis according to the acceleration sequence and the angular velocity sequence, and calculating a timing drift prediction index according to the motion feature sequence; Performing hierarchical processing on the multi-sensor sampling data according to the timing drift prediction index to generate a data sequence to be corrected, wherein the hierarchical processing includes taking the ratio of the difference of the sampling frequency to the preset frequency threshold as a priority parameter, and compensating the data according to the priority parameter; The data sequence to be corrected is divided into time windows, a time series correlation matrix is ​​generated based on local features in each time window, an optimal time series mapping relationship is calculated according to the time series correlation matrix, and a time series correction parameter is obtained; Performing nonlinear interpolation and time sequence remapping on the to-be-corrected data sequence according to the time sequence correction parameters to obtain a corrected sensor data sequence; Target features are extracted based on the corrected sensor data sequence, a collision risk level is calculated according to the motion feature sequence and the target features, and a warning signal is generated.

[0006] Optionally, the obtaining of the acceleration sequence and the angular velocity sequence of the electric bicycle, generating a motion feature sequence through frequency domain analysis according to the acceleration sequence and the angular velocity sequence, and calculating the timing drift prediction index according to the motion feature sequence includes: Obtaining an acceleration sequence and an angular velocity sequence, and performing dual-frequency sampling on the acceleration sequence and the angular velocity sequence according to a preset sampling rule to obtain a high-frequency sequence segment and a low-frequency sequence segment; Performing Fourier transformation on the high-frequency sequence fragments and the low-frequency sequence fragments to obtain frequency domain feature data, and separating a high-frequency component representing a sharp turn feature and a low-frequency component representing a speed change from the frequency domain feature data; Calculate a sudden steering change parameter according to the high-frequency component, calculate an acceleration / deceleration parameter according to the low-frequency component, and combine the sudden steering change parameter and the acceleration / deceleration parameter to obtain a motion state feature; Performing time series correlation analysis on the motion state characteristics, calculating the rate of change of characteristic values ​​in adjacent time windows, and obtaining a state transition characteristic sequence; Identify the motion mode of the electric bicycle according to the state transition feature sequence, match the identified motion mode based on a preset motion mode feature library, and generate a motion feature sequence; Weight coefficients are set for the sudden turn parameters and acceleration / deceleration parameters in the motion feature sequence respectively, and weighted calculation is performed according to the weight coefficients and a preset drift risk threshold to obtain a time series drift prediction index.

[0007] Optionally, the identifying the motion mode of the electric bicycle according to the state transition feature sequence, matching the identified motion mode based on a preset motion mode feature library, and generating a motion feature sequence includes: Dividing the state transition feature sequence into time windows, calculating a fluctuation parameter according to fluctuations of feature values ​​within the time window, and dividing the state transition feature sequence into a plurality of feature segments according to the fluctuation parameter; The characteristic segments are hierarchically clustered according to their amplitudes to obtain hierarchical characteristic groups, and characteristic templates are generated according to the temporal distribution rules of the hierarchical characteristic groups; Performing similarity matching between the feature template and a reference feature template in the preset motion pattern feature library to obtain a pattern matching score; Classifying the sports scenes according to the pattern matching scores, and dividing the sports scenes into straight-line riding scenes, turning riding scenes, accelerating riding scenes, and decelerating riding scenes; Calculating characteristic statistical parameters for the state transition characteristic sequences in different riding scenarios respectively, and generating a scene characteristic vector according to the time series variation trend of the characteristic statistical parameters; The scene feature vectors are spliced ​​according to a time sequence relationship to generate a motion feature sequence.

[0008] Optionally, the multi-sensor sampling data is hierarchically processed according to the timing drift prediction index to generate a data sequence to be corrected, wherein the hierarchical processing includes taking the ratio of the difference between the sampling frequencies and a preset frequency threshold as a priority parameter, and compensating the data according to the priority parameter, including: Obtaining a sampling time series of multi-sensor sampling data, performing correlation analysis on the sampling time series and the time series drift prediction index, and dividing the data into sharp turn scenario data, fast start-stop scenario data, and smooth riding scenario data according to the correlation degree; Calculating sampling frequencies for the sharp turn scenario data, the fast start-stop scenario data, and the steady riding scenario data respectively to obtain a scenario-specific sampling frequency matrix; Calculate the sampling frequency difference according to the sub-scenario sampling frequency matrix, perform a ratio operation on the sampling frequency difference and the preset frequency threshold to obtain a basic priority parameter; The basic priority parameter is weighted and adjusted according to the timing drift prediction index, a weight coefficient K1 is assigned to the sharp turn scene data, a weight coefficient K2 is assigned to the fast start and stop scene data, and a weight coefficient K3 is assigned to the smooth riding scene data, to obtain a priority parameter; The multi-sensor sampling data is compensated in sections according to the priority parameter, wherein compensation is performed every N data points in the sharp turn scene data, compensation is performed every 2N data points in the fast start-stop scene data, and compensation is performed every 4N data points in the smooth riding scene data, where N is a basic compensation interval parameter, to obtain compensated data; The compensated data are reorganized in time sequence to generate a data sequence to be corrected.

[0009] Optionally, the basic priority parameter is weightedly adjusted according to the timing drift prediction index, a weight coefficient K1 is assigned to the sharp turn scene data, a weight coefficient K2 is assigned to the fast start and stop scene data, and a weight coefficient K3 is assigned to the smooth riding scene data, to obtain the priority parameter, including: The steering process is segmented according to the change rate of the time series drift prediction index, and the correlation coefficient between the mean angular velocity of each steering segment and the drift prediction index is calculated to obtain the steering segment drift risk index; According to the cross-correlation value of the time series drift prediction index and the acceleration sequence, the fluctuation characteristics of the drift prediction index during acceleration and deceleration are calculated to obtain the start-stop drift risk index; Based on the turning section drift risk index, nonlinear mapping is performed on the basic priority parameter to generate the weighting coefficient K1; Based on the start-stop segment drift risk index, segmentally map the basic priority parameter to generate the weighting coefficient K2; Performing linear mapping on the basic priority parameter according to the steady-state value of the timing drift prediction index to generate the weighting coefficient K3; The weighting coefficient K1, the weighting coefficient K2 and the weighting coefficient K3 are weighted and superimposed on the data of the corresponding scene to obtain the priority parameter.

[0010] Optionally, dividing the data sequence to be corrected into time windows, generating a time series correlation matrix based on local features in each time window, calculating an optimal time series mapping relationship according to the time series correlation matrix, and obtaining a time series correction parameter includes: Calculating the steering angular acceleration and the speed change rate for the data sequence to be corrected, identifying the sharp turning point and the rapid start-stop point according to the steering angular acceleration and the speed change rate, and dividing the data sequence to be corrected into a turning section, an acceleration section, a deceleration section and a constant speed section; According to the characteristics of various motion segments, differentiated window expansion strategies are adopted for different segments. The window of the turning segment is expanded with a step length of T, the windows of the acceleration segment and the deceleration segment are expanded with a step length of 2T, and the window of the uniform speed segment is expanded with a step length of 4T, where T is the basic expansion step length parameter, to generate a multi-scale window sequence; Performing wavelet transform on the data in the multi-scale window sequence, extracting energy distribution features of different frequency bands, calculating the temporal consistency scores between windows according to the energy distribution features, and generating a segmented feature vector; The segmented feature vectors are combined to construct a block time series correlation matrix, and different weight coefficients are set for the block time series correlation matrix according to the motion characteristics of the electric bicycle to obtain a weighted correlation matrix; Calculating the timing mapping cost function according to the weighted correlation matrix, solving the optimal timing mapping function with steering continuity and speed smoothness as constraints; The data of each motion segment is resampled and calculated according to the optimal timing mapping function, and the timing correction parameters are generated in combination with the feature weights of the corresponding motion segments.

[0011] Optionally, performing nonlinear interpolation and time sequence remapping on the to-be-corrected data sequence according to the time sequence correction parameter to obtain a corrected sensor data sequence includes: Classifying the data sequence to be corrected according to the motion characteristics of the electric bicycle, dividing the data into steering process data, acceleration and deceleration process data and uniform speed process data, and grouping the timing correction parameters accordingly to obtain the classified correction parameters; The time axis of the steering process data is non-uniformly divided, and the division density is proportional to the steering angle change rate; the time axis of the acceleration and deceleration process data is variable-density divided, and the division density is proportional to the acceleration amplitude, to generate a differentiated time grid; Generate interpolation control points according to the differentiated time grid and the classification correction parameters, and set different interpolation constraints for different types of data, wherein the control point interval of the steering process data is t1, the control point interval of the acceleration and deceleration process data is t2, and the control point interval of the uniform speed process data is t3; A piecewise nonlinear interpolation function is constructed based on the interpolation control points, cubic spline interpolation is used for the steering process data, quadratic spline interpolation is used for the acceleration and deceleration process data, and linear interpolation is used for the uniform speed process data to obtain a multi-mode interpolation function; Calculating a time series mapping matrix according to the multi-mode interpolation function, substituting the classification correction parameters into the time series mapping matrix, and generating a correction remapping function; The correction remapping function is used to resample and timestamp the data sequence to be corrected to obtain the corrected sensor data sequence.

[0012] Optionally, the time axis of the steering process data is non-uniformly divided, the division density is proportional to the steering angle change rate, the time axis of the acceleration and deceleration process data is variable-density divided, the division density is proportional to the acceleration amplitude, and the differentiated time grid is generated, including: Performing motion scene analysis on the steering process data and the acceleration / deceleration process data, dividing the steering process data into a large-angle turning segment, a small-angle turning segment, and a transition turning segment, and dividing the acceleration / deceleration process data into an emergency braking segment, a slow deceleration segment, and a uniform acceleration segment, to obtain a scene segment sequence; Calculating the steering angle change rate and the acceleration amplitude according to the scene segment sequence, assigning a weight coefficient W1 to the steering angle change rate of the large-angle turning segment, assigning a weight coefficient W2 to the steering angle change rate of the small-angle turning segment, and assigning a weight coefficient W3 to the steering angle change rate of the transition turning segment, to obtain a weighted steering feature; A weight coefficient V1 is assigned to the acceleration amplitude of the emergency braking section, a weight coefficient V2 is assigned to the acceleration amplitude of the slow deceleration section, and a weight coefficient V3 is assigned to the acceleration amplitude of the uniform acceleration section, to obtain a weighted acceleration feature; Calculating the division step length of the steering process according to the weighted steering feature, wherein the step length is inversely proportional to the weighted steering feature, and generating a steering process time grid; Calculating the division step length of the acceleration and deceleration process according to the weighted acceleration feature, wherein the step length is inversely proportional to the weighted acceleration feature, and generating a time grid of the acceleration and deceleration process; According to the motion continuity constraint of the electric bicycle, the steering process time grid and the acceleration / deceleration process time grid are boundary optimized and merged to generate a differentiated time grid.

[0013] Optionally, the extracting target features based on the corrected sensor data sequence, calculating the collision risk level according to the motion feature sequence and the target features, and generating a warning signal includes: Performing target detection on the corrected sensor data sequence, dividing the detection scene into a turning scene, a fast start-stop scene, and a straight-ahead scene according to the motion feature sequence, extracting position features and motion features of the targets in different scenes to obtain scenario-based target features; Classifying the targets according to the scenario-based target features, focusing on lateral approaching targets in turning scenarios, focusing on longitudinal moving targets in fast start-stop scenarios, and focusing on omnidirectional moving targets in straight-ahead scenarios, and generating a target threat index; The target threat index and the motion feature sequence are coupled and calculated, the steering angular velocity is used as a weighting factor in a turning scenario, the acceleration is used as a weighting factor in a fast start-stop scenario, and the speed is used as a weighting factor in a straight-ahead scenario, to obtain a scenario weight coefficient; Dividing the target into an emergency avoidance zone, a steering avoidance zone, a deceleration avoidance zone and a safety zone according to the scenario weight coefficient, assigning collision risk levels to different zones, and generating a zone risk level sequence; The partition risk level sequence is matched with the maneuverability parameter of the electric bicycle, and an acoustic and light warning with an interval time of P is issued for the emergency avoidance zone, an acoustic and light warning with an interval time of 2P is issued for the turning avoidance zone, and an acoustic and light warning with an interval time of 4P is issued for the deceleration avoidance zone, where P is a basic warning cycle parameter, to obtain a partition warning signal; The partition warning signals are fused to generate a warning signal.

[0014] A second aspect of the present invention provides an electric bicycle, comprising: A data acquisition module, used to obtain an acceleration sequence and an angular velocity sequence of the electric bicycle, generate a motion feature sequence through frequency domain analysis according to the acceleration sequence and the angular velocity sequence, and calculate a timing drift prediction index according to the motion feature sequence; A hierarchical processing module, used for hierarchically processing the multi-sensor sampling data according to the timing drift prediction index to generate a data sequence to be corrected, wherein the hierarchical processing includes taking the ratio of the difference of the sampling frequency to the preset frequency threshold as a priority parameter, and compensating the data according to the priority parameter; A data segmentation module is used to segment the data sequence to be corrected into time windows, generate a time series correlation matrix based on local features in each time window, calculate the optimal time series mapping relationship according to the time series correlation matrix, and obtain a time series correction parameter; A data correction module, used for performing nonlinear interpolation and time sequence remapping on the to-be-corrected data sequence according to the time sequence correction parameters to obtain a corrected sensor data sequence; The risk warning module is used to extract target features based on the corrected sensor data sequence, calculate the collision risk level according to the motion feature sequence and the target features, and generate a warning signal.

[0015] In scenarios such as electric bicycles that are sensitive to cost and power consumption, this solution uses acceleration sequences and angular velocity sequences to perform frequency domain analysis, generate motion sequences that reflect the characteristics of rapid starting, stopping and sharp turns of vehicles, and calculate a prediction index that can predict the degree of data misalignment. Through this prediction index, the multi-sensor sampling data is graded and processed, and the data source can be compensated, interpolated and time-remapped according to the differences in different sampling frequencies and potential drift risks. Since the speed, acceleration and steering angular velocity of electric bicycles may change dramatically under high dynamic conditions, the graded compensation process introduced in this solution can capture these changes in time, align data at different frequencies to a more consistent time scale, and fundamentally reduce the large-scale timing misalignment of multiple sensors under high-speed motion conditions. In particular, after the graded processing, the data sequence to be corrected will be divided into time windows, and a correlation matrix will be constructed based on local features, and then the optimal timing mapping relationship will be solved, so that the final corrected sensor data can maintain high accuracy and real-time performance in sudden change scenarios.

[0016] Throughout the entire process, in order to further adapt to the limited computing resources of electric bicycles, the solution does not use an overly large algorithm structure or rely on high-precision clock synchronization hardware, but simplifies the capture process of high-dynamic behaviors through multi-band analysis of acceleration and angular velocity. In other words, after obtaining the motion sequence, the solution will compensate and interpolate the data of different sensors in a targeted manner according to the corresponding prediction index, and timely correct the timing misalignment caused by the sampling frequency difference. The nonlinear interpolation and time remapping principles used here can flexibly correct local moments such as steering, acceleration and deceleration with higher accuracy while ensuring that the amount of calculation is controllable, so that the perception and prediction of the obstacle position of the electric bicycle during rapid start and stop or sharp turn will not be distorted due to timing drift. In this way, the solution effectively alleviates the error accumulation caused by data drift in multi-sensor fusion, and provides a practical technical path for the stable collision prevention of electric bicycles in complex urban traffic environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying creative work.

[0018] Figure 1 A schematic diagram of an embodiment of a method for preventing collision of an electric bicycle in an embodiment of the present invention; Figure 2 FIG. 1 is a schematic diagram of an electric bicycle according to an embodiment of the present invention.

[0019] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0020] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0021] It should be noted that if the embodiments of the present invention involve directional indications (such as up, down, left, right, front, back...), the directional indications are only used to explain the relative position relationship, movement status, etc. between the components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication will also change accordingly.

[0022] In addition, the descriptions of "first", "second", etc. in the present invention are only used for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In addition, "and / or" in the full text includes three solutions. Taking A and / or B as an example, it includes technical solution A, technical solution B, and technical solution that satisfies both A and B. In addition, the technical solutions between the various embodiments can be combined with each other, which must be based on the ability of ordinary technicians in the field to implement. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0023] An embodiment of the present application provides an electric bicycle collision prevention method and an electric bicycle. Figure 1 A method for preventing collision of an electric bicycle and a flow chart of an electric bicycle provided in an embodiment of the present application. In this embodiment, the method includes: See also Figure 1 , obtaining an acceleration sequence and an angular velocity sequence of the electric bicycle, generating a motion feature sequence through frequency domain analysis according to the acceleration sequence and the angular velocity sequence, and calculating a timing drift prediction index according to the motion feature sequence; In one embodiment of the present invention, the obtaining of the acceleration sequence and the angular velocity sequence of the electric bicycle, generating a motion feature sequence through frequency domain analysis according to the acceleration sequence and the angular velocity sequence, and calculating the timing drift prediction index according to the motion feature sequence includes: Obtaining an acceleration sequence and an angular velocity sequence, and performing dual-frequency sampling on the acceleration sequence and the angular velocity sequence according to a preset sampling rule to obtain a high-frequency sequence segment and a low-frequency sequence segment; Performing Fourier transformation on the high-frequency sequence fragments and the low-frequency sequence fragments to obtain frequency domain feature data, and separating a high-frequency component representing a sharp turn feature and a low-frequency component representing a speed change from the frequency domain feature data; Calculate a sudden steering change parameter according to the high-frequency component, calculate an acceleration / deceleration parameter according to the low-frequency component, and combine the sudden steering change parameter and the acceleration / deceleration parameter to obtain a motion state feature; Performing time series correlation analysis on the motion state characteristics, calculating the rate of change of characteristic values ​​in adjacent time windows, and obtaining a state transition characteristic sequence; Identify the motion mode of the electric bicycle according to the state transition feature sequence, match the identified motion mode based on a preset motion mode feature library, and generate a motion feature sequence; Weight coefficients are set for the sudden turn parameters and acceleration / deceleration parameters in the motion feature sequence respectively, and weighted calculation is performed according to the weight coefficients and a preset drift risk threshold to obtain a time series drift prediction index.

[0024] Specifically, the acquisition of acceleration sequence and angular velocity sequence can be achieved by installing devices with acceleration and angular velocity measurement functions on the vehicle body. These devices will produce raw data in real time during the vehicle's driving process, and these data will be split according to the established dual-frequency sampling rules. The dual-frequency sampling rules include two different sampling frequencies: the higher sampling frequency is used to capture the impact or steering information with significant amplitude changes in a short time, and the lower sampling frequency is used to record the acceleration and deceleration trend of the vehicle over a longer time range. If the vehicle is driving slowly on a straight road, the angular velocity and acceleration fluctuations collected by the high-frequency sampling channel are relatively small, and the sequence data of the low-frequency sampling channel will not change significantly between adjacent sampling points; if the vehicle needs to turn quickly or dodge urgently at an intersection, the high-frequency sampling channel will record abnormally rising acceleration values ​​or angular velocity values ​​in a very short time interval, forming a high-frequency sequence segment, while the data recorded by the low-frequency sampling channel is more focused on the smooth evolution of speeds above several seconds, forming a low-frequency sequence segment.

[0025] After obtaining the high-frequency sequence fragment and the low-frequency sequence fragment, the two parts of the time domain signal need to be input into the Fourier transform processing flow respectively. The Fourier transform can be realized with the help of the fast Fourier algorithm, mapping the discrete time domain sampling points to the frequency domain to obtain the amplitude distribution of each frequency component. At this time, the high-amplitude peak area in the high-frequency sequence fragment will be concentrated in a relatively high frequency range, and can better reflect the characteristics of sharp turns or short-term violent shaking; the low-frequency sequence fragment shows the overall trend of acceleration and deceleration in a relatively low frequency range. In order to further refine the available information, it is necessary to define a certain frequency bandwidth in the frequency domain feature data to separate the high-frequency component that can represent the sudden change in steering and the low-frequency component that can represent the speed change. If the vehicle continues to change direction quickly for a period of time, the energy peak in the high-frequency component will appear as multiple concentrated and prominent peaks; if the vehicle is in the process of acceleration or deceleration, the overall amplitude of the low-frequency component will show a clear upward or downward trajectory over time.

[0026] Subsequently, the sudden turn parameter needs to be calculated from the high-frequency component. The sudden turn parameter can be quantified by the number of peaks in the high-frequency component, the peak amplitude, and the distribution density of the peaks on the time axis. For example, in a crowded urban road section, if the vehicle has several consecutive sharp turns in a short period of time, the peaks of the high-frequency component will be relatively dense and have large amplitudes. The characteristic values ​​of these peaks can be accumulated or statistically averaged to obtain the sudden turn parameter reflecting the intensity of the turn. In parallel, the amplitude tracking or slope analysis of the low-frequency component is performed to calculate the acceleration and deceleration parameters. The acceleration and deceleration parameters can be determined by comparing the amplitude change rate of consecutive low-frequency data points. If the amplitude rises at a large rate, it means that the vehicle is accelerating rapidly. If the amplitude continues to decrease, it reflects that the vehicle is in a clear deceleration stage. After combining the sudden turn parameter with the acceleration and deceleration parameter, the motion state feature can be obtained. At this time, the motion state feature in each time segment will identify the change in steering and the change in speed. For example, if the sudden turn parameter and the acceleration and deceleration parameter are both higher than a specific threshold within a certain period of time, it means that the vehicle is in a compound action of sharp turn and rapid speed change.

[0027] In order to grasp the specific trajectory of the evolution of these motion state characteristics over time, it is necessary to introduce time series correlation analysis. Time series correlation analysis can segment the motion state characteristics according to windows of fixed length or adaptive length, such as segmenting every tens of milliseconds or longer time periods, and then calculate the rate of change of the characteristic values ​​in adjacent windows. The rate of change is the value obtained by normalizing or comparing the difference between the steering mutation parameter and the acceleration and deceleration parameter between windows. If the steering mutation parameter shows a rapid increase trend and the acceleration and deceleration parameter is also increasing in several consecutive windows, it can be inferred that the vehicle is in a stage where the speed and steering are rising synchronously; if the steering mutation parameter gradually falls back and the acceleration and deceleration parameter stabilizes at a low level, it means that the vehicle has returned to a relatively stable working condition. The sequence formed by these change rates is called the state transition feature sequence, which can map the moment and intensity of the vehicle's riding mode switching to a certain extent.

[0028] Next, the state transition feature sequence needs to be compared with the pre-established motion mode feature library, which contains several common riding scenarios, such as large-angle turns, emergency avoidance, rapid start and stop, and smooth cruising. Each scenario corresponds to the range and jump rules of the steering mutation parameters and acceleration and deceleration parameters on different time scales. By matching the numerical distribution of the state transition feature sequence with these rules one by one, the vehicle's most suitable motion mode can be identified and the motion feature sequence can be output. For example, if a certain feature value sequence shows that the steering mutation parameters and acceleration and deceleration parameters fluctuate greatly in a short period of time, which coincides with the emergency avoidance mode curve in the feature library, it will be determined that the vehicle is in a scenario of avoiding risks. In this way, the motion feature sequence can accurately identify the current dynamic state of the vehicle.

[0029] Finally, it is necessary to set weight coefficients for the steering mutation parameter and the acceleration / deceleration parameter respectively, and perform weighted calculations in combination with the pre-defined drift risk threshold to obtain the time series drift prediction index. If the vehicle makes multiple sharp turns in a very short period of time at a densely trafficked intersection, the importance of the steering mutation parameter should be relatively amplified. If the vehicle also has rapid acceleration or emergency stop at this time, the acceleration / deceleration parameter will also have a higher weight coefficient. The two together will generate a time series drift prediction index with a higher value. The higher the index, the more likely the vehicle's current motion characteristics are to cause the risk of misalignment in the time alignment of different sensor outputs. Once the timing is inconsistent, the obstacle position prediction will have significant deviations, affecting the accuracy of subsequent collision avoidance judgments. Through this indexation method, the system can quantitatively describe the data drift risks in a high dynamic environment, and then give priority to these high index periods in the subsequent data correction and fusion links to ensure that the multi-sensor fusion process will not cause large errors in obstacle perception due to frequency differences.

[0030] In an embodiment of the present invention, the method for identifying the motion mode of an electric bicycle according to the state transition feature sequence and matching the identified motion mode based on a preset motion mode feature library to generate a motion feature sequence includes: Dividing the state transition feature sequence into time windows, calculating a fluctuation degree parameter according to the fluctuation situation of the feature values within the time windows, and dividing the state transition feature sequence into multiple feature segments according to the fluctuation degree parameter; Performing hierarchical clustering on the feature segments according to the amplitude size to obtain a hierarchical feature group, and generating a feature template according to the timing distribution law of the hierarchical feature group; Performing similarity matching between the feature template and a reference feature template in the preset motion mode feature library to obtain a pattern matching score; Classifying the motion scenarios according to the pattern matching score, and classifying the motion scenarios into a straight-line riding scenario, a turning riding scenario, an accelerating riding scenario, and a decelerating riding scenario; Calculating feature statistical parameters for the state transition feature sequence in different riding scenarios respectively, and generating a scenario feature vector according to the timing change trend of the feature statistical parameters; Stitching the scenario feature vectors according to the timing relationship to generate a motion feature sequence.

[0031] Specifically, when dividing the state transition feature sequence into time windows, it is necessary to first set a rolling or segmented time interval on the entire feature sequence and read the continuous change of the feature values within this time interval. If the vehicle experiences frequent turning or acceleration / deceleration operations during a certain period, the values of the steering mutation parameter and the acceleration / deceleration parameter often show large fluctuations, and the state transition feature sequence will also synchronously show obvious peak-valley alternations. In order to quantify this peak-valley alternation, it is necessary to extract the fluctuation range from the set of feature values within each time window and regard it as the basis for the fluctuation degree parameter. The calculation method of the fluctuation degree parameter can be completed by analyzing the range, variance, or standard deviation of the feature values within the time window. If the distribution amplitude of the feature values in some windows is high, the value of the fluctuation degree parameter for this window will also increase. By continuously monitoring the fluctuation degree parameters of these windows, the overall fluctuation situation of the state transition feature sequence on the time axis can be grasped. Subsequently, according to the fluctuation degree parameter corresponding to each window, the state transition feature sequence is divided into multiple feature segments. A high fluctuation degree parameter in some windows means that the vehicle has sharp turns, rapid acceleration / deceleration, or other intense actions during this stage, while windows with a low fluctuation degree parameter indicate that the vehicle may be in a stage of steady riding or small adjustments. After segmentation, the feature segments can more precisely present the dynamic behavior characteristics of the vehicle at different time periods.

[0032] After the feature segments are divided, they need to be hierarchically clustered according to the amplitude. In this process, the overall amplitude level of each segment is first counted, for example, by calculating the average or peak value of the steering mutation parameter and the acceleration and deceleration parameter in the segment, and these amplitude sets are regarded as the initial clustering data. Next, these amplitude data are hierarchically processed using clustering algorithms such as hierarchical clustering or K-means to obtain several hierarchical feature groups. The generation of hierarchical feature groups helps to distinguish between segments that show a sharp change trend in motion and segments with relatively medium or gentle amplitudes, and can also classify segments with the smallest fluctuations into a separate category. If a vehicle is detected to make repeated large-angle turns in a short period of time at an urban intersection, the fluctuation value of the steering mutation parameter will reach a high amplitude several times in a row, and the corresponding feature segments will be clustered into a high amplitude group; if there is only slight acceleration or deceleration in the subsequent section of the journey, those segments are more likely to be classified into medium amplitude or low amplitude groups. According to the temporal distribution law of the hierarchical feature group, a feature template can be further generated. This step mainly extracts the distribution form of the start time, end time and amplitude for each hierarchical feature group, and integrates this information into a template that can represent the typical temporal characteristics of the hierarchical feature group. The feature template can retain the key peak and valley positions and amplitude change laws on the time axis. For example, the template of the high amplitude group may show more obvious peak structures, while the template of the low amplitude group is relatively stable within the same time span.

[0033] After obtaining the feature template, it is necessary to perform similarity matching with the benchmark feature template in the preset motion mode feature library to obtain the pattern matching score. The motion mode feature library usually contains benchmark templates under various typical riding conditions, such as feature curves for straight riding, feature curves for turning riding, feature curves for accelerating riding, and feature curves for decelerating riding. Similarity matching can be completed based on methods such as dynamic time warping algorithm or correlation coefficient calculation. By comparing the shape and amplitude trend of the feature template and the benchmark feature template on the time axis, the similarity can be quantitatively evaluated. If a high-amplitude group template has multiple sharp turning mutation parameter peaks in a short period of time, and these peaks are highly consistent with the benchmark template for turning riding, the matching algorithm will assign a higher pattern matching score. On the contrary, if a low-amplitude group template is significantly different from the benchmark feature template, the matching score will be low.

[0034] When classifying motion scenarios based on the pattern matching score, it is necessary to combine the time period in which the feature group obtained from the previous hierarchical clustering is located and the specific value of the matching score, and classify the motion scenarios of the vehicle in each time interval into a straight-line riding scenario, a turning riding scenario, an accelerating riding scenario, or a decelerating riding scenario. If the steering mutation parameter in the state transition feature sequence fluctuates significantly during this time period and has the highest similarity with the turning riding benchmark template, then this time period is determined to be a turning riding scenario; if the acceleration / deceleration parameter has the highest matching score with the accelerating riding benchmark template, it is judged as an accelerating riding scenario; if both values are relatively smooth, it is closer to a straight-line riding scenario or a decelerating riding scenario. In this way, the entire state transition feature sequence is marked with the corresponding scenario categories in different time periods.

[0035] After obtaining these preliminary motion scenario determinations, it is necessary to calculate the feature statistical parameters for the state transition feature sequences under different riding scenarios respectively. The feature statistical parameters can include the number of peaks of the steering mutation parameter, the average change rate of the acceleration / deceleration parameter, the maximum fluctuation range, or the stability index, etc. in this scenario. If a turning riding scenario contains three obvious turning peaks and two moderate speed drops, the number of peaks of the steering mutation parameter will be higher than that of the straight-line riding scenario, while the average change rate of the acceleration / deceleration parameter may show a relatively stable trend. By tracking the temporal variation trend of these parameters, a numerical distribution reflecting the feature evolution within the scenario can be obtained, and then a scenario feature vector is generated. The scenario feature vector often consists of several normalized parameters. For example, the turning scenario vector will include the number of peaks of the steering mutation parameter, the average fluctuation value, and the duration, etc.

[0036] By splicing these scenario feature vectors in chronological order, the scenario information scattered in different time segments can be merged into a complete and continuous motion feature sequence. If a long riding process includes multiple turns and several segments of acceleration, then the spliced scenario feature vectors will present an overall manifold with multi-scenario orderly connection such as turning - accelerating - straight - decelerating, which can more intuitively reflect the motion state switching of the vehicle during the entire journey. For example, if the vehicle only starts slowly at the beginning, then makes one or two large-angle turns at the street corner in the urban area, then rides at a constant speed on a relatively spacious section and decelerates urgently at the traffic light intersection, and finally moves forward smoothly again, then these links will be reflected in turn in the splicing result of the scenario feature vectors, and each link carries the corresponding feature statistical parameters and fluctuation record. In this way, the motion information of the vehicle in various scenarios is unified and dynamically integrated into the same sequence, providing a more targeted decision-making basis for subsequent analysis or anti-collision logic.

[0037] Please continue to refer to Figure 1, perform hierarchical processing on the multi-sensor sampling data according to the timing drift prediction index to generate a data sequence to be corrected, where the hierarchical processing includes using the ratio of the difference in sampling frequencies to a preset frequency threshold as a priority parameter and compensating the data according to the priority parameter; In an embodiment of the present invention, the performing hierarchical processing on the multi-sensor sampling data according to the timing drift prediction index to generate a data sequence to be corrected, where the hierarchical processing includes using the ratio of the difference in sampling frequencies to a preset frequency threshold as a priority parameter and compensating the data according to the priority parameter, includes: Obtain the sampling time sequence of the multi-sensor sampling data, perform correlation analysis on the sampling time sequence and the timing drift prediction index, and classify the data into sharp turn scenario data, rapid start-stop scenario data, and steady riding scenario data according to the degree of correlation; Calculate the sampling frequencies for the sharp turn scenario data, the rapid start-stop scenario data, and the steady riding scenario data respectively to obtain a sub-scenario sampling frequency matrix; Calculate the sampling frequency difference according to the sub-scenario sampling frequency matrix, perform a ratio operation on the sampling frequency difference and the preset frequency threshold to obtain a basic priority parameter; Perform weighted adjustment on the basic priority parameter according to the timing drift prediction index, assign a weighting coefficient K1 to the sharp turn scenario data, assign a weighting coefficient K2 to the rapid start-stop scenario data, and assign a weighting coefficient K3 to the steady riding scenario data to obtain a priority parameter; Perform segmented compensation on the multi-sensor sampling data according to the priority parameter, where compensation is performed every N data points in the sharp turn scenario data, every 2N data points in the rapid start-stop scenario data, and every 4N data points in the steady riding scenario data, N is a basic compensation interval parameter, to obtain compensated data; Reorganize the compensated data in chronological order to generate a data sequence to be corrected.

[0038] Specifically, the sampling time series of multi-sensor sampling data includes data points generated by each sensor at different times and their timestamps. To relate this data to the timing drift risk of the vehicle under different operating conditions, it is necessary to first perform a correlation analysis on the sampling time series and the timing drift prediction index. During the correlation analysis, a certain step size or window can be selected on the time axis, and the change of the timing drift prediction index within this window is compared with the distribution of acceleration or angular velocity in the sampling time series, and the correlation coefficient or other coupling degree indicators are used to measure the corresponding relationship between the two. If the timing drift prediction index remains in a high numerical range within a certain window, and there are significant instantaneous steering fluctuations or signs of rapid acceleration and deceleration in the sampling time series, it is determined that the data set corresponding to this window is more likely to exhibit high dynamic misalignment phenomena, and this set can be marked as sharp turn scenario data or rapid start / stop scenario data; if the value of the timing drift prediction index is stable within a window of the same size and the change of the sampling time series is small, this part is classified as stable riding scenario data. In this way, the three types of segmented scenario data can more accurately reflect the multi-sensor information of the vehicle during high-dynamic and low-dynamic periods.

[0039] After obtaining the sharp turn scenario data, rapid start / stop scenario data, and stable riding scenario data, it is necessary to further calculate the sampling frequency of each of these three types of data and organize it into a sub-scenario sampling frequency matrix. The process of calculating the sampling frequency can first count the number of data points contained in each scenario segment within a unit time, and then perform an average or median estimation on it to obtain the average sampling frequency of the sharp turn scenario data, rapid start / stop scenario data, and stable riding scenario data. If an electric bicycle frequently turns left and right on a narrow road, the sharp turn scenario data will concentrate on large-angle turn records, and the data points generated by the sensors are relatively closely spaced from each other, resulting in a relatively higher sampling frequency; if a section of the journey is mainly stable driving, the sensor output of the stable riding scenario data fluctuates less, and the sampling frequency will also be closer to the set reference value.

[0040] Next, it is necessary to calculate the sampling frequency difference based on the sub-scenario sampling frequency matrix and perform a ratio operation on this difference and a preset frequency threshold to obtain a basic priority parameter. The sampling frequency difference can be regarded as the difference between the actual sampling frequency and the ideal synchronous frequency. If the sampling frequency in a certain scenario far exceeds or is far lower than this ideal value, it indicates that the risk of timing inconsistency between sensors in this scenario is greater. The preset frequency threshold can be set as a scalar for evaluating the tolerable degree of sampling frequency dispersion. When the sampling frequency difference of a certain type of scenario significantly exceeds this threshold, the result of the corresponding ratio operation will show a value higher than 1 or other obvious over-standard levels, indicating that there is a high risk of multi-sensor misalignment in this scenario. After obtaining the basic priority parameter in this way, the correction requirements for the three types of scenarios can be sorted based on this.

[0041] On this basis, it is also necessary to perform weighted adjustment on the basic priority parameters in combination with the time series drift prediction index to obtain the priority parameters. The time series drift prediction index is usually a comprehensive value obtained by considering factors such as the sharp turn amplitude and the severity of acceleration and deceleration during the previous analysis of vehicle motion characteristics. If the index shows that the vehicle frequently exhibits turning peaks or rapid acceleration and deceleration within a certain period of time, the importance of sharp turn scenario data and rapid start-stop scenario data should be amplified. To achieve this amplification or reduction effect, different weighting coefficients K1, K2, and K3 can be set for different scenarios. When the coefficient K1 corresponding to the sharp turn scenario data is significantly greater than 1, it means that the proportion of data in this type of scenario is increased when calculating the priority parameters; when the coefficient K3 corresponding to the smooth riding scenario data is relatively small, it indicates that the urgency of such data for subsequent time series correction is lower. Through such a refined weighting process, the priority parameters can more accurately reflect the differences between high-dynamic scenarios and low-dynamic scenarios, enabling the system to allocate precious computing resources to high-risk periods.

[0042] Once the priority parameters are determined, segmented compensation can be performed on the multi-sensor sampling data. Segmented compensation refers to setting specific compensation intervals in the data point sequence according to the scenario category and performing interpolation or alignment processing at the intervals. If the sharp turn scenario data has the highest priority, a compensation operation will be performed every N data points. For example, when the vehicle is making continuous sharp turns, the high-frequency sampled angular velocity data will be aligned to a unified time axis using interpolation methods to reduce misalignment; if the rapid start-stop scenario data has a slightly lower priority, compensation can be performed every 2N data points; if the smooth riding scenario data has the lowest priority, interpolation can be performed only every 4N data points. N is the basic compensation interval parameter, which can be jointly determined by system resources and real-time requirements. If a processor with limited performance is used on a certain electric bicycle, N can be appropriately increased, but due to the relatively large time series risk in the sharp turn scenario, the compensation frequency will still be denser. This measure can better correct the errors caused by the mismatch of sensor sampling frequencies during the critical periods of high-speed turning or frequent start-stop of the vehicle, while saving a large amount of computing overhead during smooth riding.

[0043] After completing the segmented compensation, it is necessary to merge the data corresponding to the sharp turn scenario, the rapid start and stop scenario, and the smooth riding scenario back into the original time series to form a relatively complete data sequence to be corrected. This step usually involves reordering the compensated and interpolated data of each scenario in timestamp order to ensure the coherence of the actual driving process of the vehicle on the time axis. If the first half of a riding journey contains multiple sharp turns and the second half only has one short acceleration and deceleration, then the scenario data in the first half will be compensated and interpolated at a high frequency, significantly reducing the timing error, while the second half will complete the basic alignment operation at a lower compensation frequency. The data sequence to be corrected spliced in this way can effectively eliminate the significant misalignment in high-dynamic scenarios and also minimize the occupation of system resources in low-dynamic scenarios, providing more accurate basic data support for subsequent fusion algorithms and collision warning logics. Through this complete process, the multi-sensor anti-collision system of an electric bicycle can obtain more reliable timing matching performance at a relatively small cost of hardware and computing power, thereby improving the ability to discriminate the position of obstacles and driving risks.

[0044] In one embodiment of the present invention, the weighted adjustment of the basic priority parameter according to the timing drift prediction index, assigning a weighting coefficient K1 to the sharp turn scenario data, a weighting coefficient K2 to the rapid start and stop scenario data, and a weighting coefficient K3 to the smooth riding scenario data to obtain the priority parameter includes: Segment the steering process according to the change rate of the timing drift prediction index, calculate the mean angular velocity of each steering segment and the correlation coefficient of the drift prediction index, and obtain the steering segment drift risk index; According to the cross-correlation value between the timing drift prediction index and the acceleration sequence, calculate the fluctuation characteristics of the drift prediction index during the acceleration and deceleration processes, and obtain the start-stop segment drift risk index; Based on the steering segment drift risk index, perform a non-linear mapping on the basic priority parameter to generate the weighting coefficient K1; Based on the start-stop segment drift risk index, perform a segmented mapping on the basic priority parameter to generate the weighting coefficient K2; According to the steady-state value of the timing drift prediction index, perform a linear mapping on the basic priority parameter to generate the weighting coefficient K3; Superimpose the weighting coefficients K1, the weighting coefficient K2, and the weighting coefficient K3 with the data of the corresponding scenarios to obtain the priority parameter.

[0045] Specifically, when segmenting the steering process according to the change rate of the time-series drift prediction index, it is possible to first identify the obvious steering periods in the vehicle angular velocity data, and set several sub-intervals within the steering periods according to the change rate of the time-series drift prediction index on the time axis. If the instantaneous value of the angular velocity fluctuates significantly multiple times within a certain period of time, and at the same time the value of the time-series drift prediction index also climbs rapidly with time, then this period is determined as a high-risk steering interval and distinguished from the adjacent relatively stable steering intervals. After segmenting these intervals, it is necessary to calculate the correlation coefficient between the mean angular velocity and the drift prediction index in each steering segment to measure the coupling degree between the angular velocity and the time-series drift prediction index. If the mean angular velocity is high during the sharp turn stage at a certain intersection and the correlation coefficient with the time-series drift prediction index is close to a strong positive correlation, it indicates that the potential impact of the time-series difference within this steering segment is relatively greater, and the drift risk index of the steering segment increases accordingly. The formed drift risk index of the steering segment can compare the dynamic characteristics of different turning intervals, so as to screen out the high-risk steering segments that need to be prioritized for attention and correction.

[0046] During the acceleration and deceleration phases, it is necessary to measure the drift risk of the vehicle during start-stop periods based on the cross-correlation value between the time-series drift prediction index and the acceleration sequence. The calculation of the cross-correlation value generally involves sliding alignment of the time-series drift prediction index and the acceleration data, and then statistical analysis of their correlation at different time offsets. If the acceleration sequence shows a jump within a certain period of time, and the time-series drift prediction index also shows a significant jump in the same period, then the cross-correlation value will increase significantly, reflecting that the vehicle has a higher requirement for multi-sensor synchronization during acceleration or deceleration. These moments are identified as high-risk start-stop segments. After calculating the fluctuation characteristics of the drift prediction index during acceleration and deceleration, the drift risk index of the start-stop segment can be obtained. If a vehicle brakes or accelerates suddenly multiple times continuously on a certain road section, this index will remain at a high level, indicating that data correction needs to be more cautious in this scenario.

[0047] After the above-mentioned steering section drift risk index and start-stop section drift risk index are obtained, different forms of mapping operations need to be performed on the basic priority parameters. First, for the steering section drift risk index, a non-linear mapping is introduced to generate the weighting coefficient K1. The non-linear mapping can be achieved by setting polynomial functions, Sigmoid functions or other methods, and the basic priority parameters are curve amplified or compressed according to the numerical range of the steering section risk index. If the vehicle turns at high speed for a long time, the steering section drift risk index will rise to a higher range, and the corresponding non-linear mapping output will apply a stronger amplification weight to the basic priority parameters, giving a higher priority to the sharp turn scenario. Subsequently, for the start-stop section drift risk index, a piecewise mapping is introduced to generate the weighting coefficient K2. The piecewise mapping will use different linear or step functions to map the basic priority parameters according to different intervals of the start-stop risk index. If the start-stop section index exceeds a certain key threshold, the weighting multiple will be increased when mapping the basic priority parameters, so that the scenario of frequent start-stop can be more fully corrected and supported. If the start-stop section risk index is at a medium level, the piecewise mapping will give a relatively moderate coefficient, enabling the system to take into account the differences between high and low acceleration and deceleration situations.

[0048] After that, it is necessary to perform a linear mapping on the basic priority parameters according to the steady-state value of the timing drift prediction index to generate the weighting coefficient K3. The steady-state value of the timing drift prediction index usually comes from the statistical characteristics over a long period of time. When the vehicle is driving under a generally stable or not highly fluctuating working condition, this index will remain stable within a lower or medium numerical range. At this time, a simple linear function can be used to multiply or add this steady-state value to the basic priority parameters, so as to obtain a weighting coefficient K3 that is more suitable for the smooth riding scenario. For example, if the vehicle speed is relatively constant on a certain section of the road and there are no obvious spikes in the angular velocity, and the timing drift prediction index remains at a low level for a long time, then the linear mapping in the steady state will keep the weighting coefficient K3 within a small gain range, thus ensuring that the system will not invest too much computational and correction costs in low-risk scenarios.

[0049] Finally, the weighting coefficients K1, K2, and K3 need to be applied to the data corresponding to the respective scenarios for weighted superposition to obtain the final priority parameter. At this time, the non-linear amplification effect promoted by the steering section drift risk index will act on the sharp turn scenario data, the piecewise mapping corresponding to the start-stop section drift risk index will affect the rapid acceleration and deceleration scenario data, and the linear mapping corresponding to the steady state value will affect the smooth riding scenario data. For example, if a vehicle makes multiple sharp turns and several sudden braking operations within a period of time, the data of these two scenarios will be amplified by K1 and K2 respectively, and the priority parameter will increase significantly in these two scenarios, thus guiding the subsequent data fusion and correction processes to interpolate and align these high-risk sections more frequently; if the vehicle maintains a medium-low speed and uniform driving on another section and the timing drift prediction index remains stable, the corresponding linear mapping K3 will not significantly increase the priority parameter value, allowing the system resources to perform timing correction at a lower frequency. In such a global framework, scenarios of sharp turns, rapid start-stop, and smooth riding can be reasonably distinguished, and the data information of each scenario will obtain corresponding processing priorities based on different weighting strategies, so as to more accurately reduce the multi-sensor timing misalignment hidden danger brought by high-dynamic behaviors in limited hardware resources and provide more stable perception and auxiliary decision-making guarantees for safe riding.

[0050] Please continue to refer to Figure 1 , perform time window segmentation on the data sequence to be corrected, generate a timing correlation matrix based on local features within each time window, calculate the optimal timing mapping relationship according to the timing correlation matrix, and obtain the timing correction parameter; In an embodiment of the present invention, the performing time window segmentation on the data sequence to be corrected, generating a timing correlation matrix based on local features within each time window, calculating the optimal timing mapping relationship according to the timing correlation matrix, and obtaining the timing correction parameter includes: Calculate the steering angular acceleration and speed change rate for the data sequence to be corrected, identify sharp turn points and rapid start-stop points according to the steering angular acceleration and the speed change rate, and divide the data sequence to be corrected into a steering section, an acceleration section, a deceleration section, and a uniform speed section; According to the characteristics of various motion segments, adopt a differential window expansion strategy for different segments. Expand the window of the steering section with a step size of T, expand the windows of the acceleration section and the deceleration section with a step size of 2T, and expand the window of the uniform speed section with a step size of 4T, where T is the basic expansion step size parameter, to generate a multi-scale window sequence; Perform wavelet transform on the data within the multi-scale window sequence, extract the energy distribution characteristics of different frequency bands, calculate the timing consistency score between windows according to the energy distribution characteristics, and generate a segmented feature vector; Combine the segmented feature vectors to construct a block time-series correlation matrix, and set different weight coefficients for the block time-series correlation matrix according to the motion characteristics of the electric bicycle to obtain a weighted correlation matrix; Calculate the time-series mapping cost function according to the weighted correlation matrix, and solve the optimal time-series mapping function with the steering continuity and speed smoothness as the constraint conditions; Resample and calculate the data of each motion segment according to the optimal time-series mapping function, and generate time-series correction parameters in combination with the characteristic weights of the corresponding motion segments.

[0051] Specifically, when calculating the steering angular acceleration and speed change rate for the data sequence to be corrected, it is necessary to first perform a difference operation on the angular velocity of adjacent data points based on the steering angle and speed change process of each data point on the time axis, and then multiply the difference result by the unit time step or normalize it according to the discrete time interval to obtain the steering angular acceleration reflecting the severity of vehicle steering; the speed change rate is obtained by continuously differentiating the speed sequence and then dividing by the difference between adjacent moments. If there is a peak in the steering angular acceleration and a sudden increase in the speed change rate within a certain time period, it can be determined that there is a sharp turning point or a fast start / stop point in this time period. After identifying the sharp turning point, the steering angular acceleration and speed change rate in this area will be significantly higher than the surrounding smooth segments. If there are multiple angular velocity peaks continuously appearing on a certain route along with large fluctuations in speed, such segments can be marked as sharp turning points or fast start / stop points, and the data sequence to be corrected is correspondingly segmented into a steering segment, an acceleration segment, a deceleration segment, and a constant-speed segment. The steering segment usually contains a large number of angular velocity mutations and the speed change does not necessarily continuously increase or decrease; the acceleration segment is characterized by a positive speed change rate and a relatively fast numerical increase; the deceleration segment is characterized by a negative speed change rate and a relatively rapid numerical decrease; the constant-speed segment shows that both the steering angular acceleration and the speed change rate are maintained within a relatively small range.

[0052] After dividing various motion segments, a differential window expansion strategy is adopted according to the characteristics of the turning segment, acceleration segment, deceleration segment, and constant-speed segment. The turning segment usually includes large-angle switches and significant fluctuations in angular acceleration. To more precisely capture the high-speed changes during the turning process, an expansion scale with a step size of T needs to be selected. The acceleration segment and deceleration segment have obvious positive or negative jumps in the rate of speed change, but their changes have a slightly lower spatio-temporal frequency compared to angular acceleration. Therefore, a multiple scale with a step size of 2T is adopted for window expansion in these motion segments. The constant-speed segment is more inclined to be stable, with fewer high-amplitude fluctuations in speed and steering. Therefore, a step size of 4T can be selected for window expansion in the constant-speed segment, which can not only save the processing requirements for high-resolution data but also retain the necessary time-series information for subsequent analysis. After this differential expansion process, a multi-scale window sequence will form several window sets divided according to T, 2T, or 4T resolutions on the same data sequence to be corrected, enabling subsequent processing to perform hierarchical analysis for high-fluctuation and low-fluctuation sections respectively.

[0053] To further extract the energy distribution characteristics in different frequency bands, wavelet transform is performed on the data within the above multi-scale window sequence. Wavelet transform projects the original time-domain data into the frequency-domain space composed of multiple sub-bands, separating the information of high-frequency sub-bands and low-frequency sub-bands, and measuring the energy of the window in the corresponding frequency band by the magnitude of the coefficient strength. If there is a concentration of high-frequency sub-band energy and a significantly large amplitude coefficient in the window of a certain turning segment, it indicates that the vehicle has a short-term sharp turn or a strong change in attitude; if the increase in medium-low frequency sub-band energy is mainly reflected in the window of the acceleration segment or deceleration segment, it reflects that the vehicle's speed progressive adjustment is more significant in this interval. After aggregating and statistically analyzing the wavelet coefficients of each window, the temporal consistency score between windows can be calculated. For example, similarity metrics or correlation indicators are used to measure the energy distribution difference between adjacent windows in the same sub-band. If the energy distribution trends of adjacent windows are similar, it indicates that the vehicle does not experience a drastic mode switch during this section of driving. After combining these temporal consistency scores, a segmented feature vector can be formed to condense the core features of each multi-scale window in the wavelet domain.

[0054] After obtaining the segmented feature vectors, it is necessary to further combine them to construct a block-wise temporal correlation matrix. The block-wise temporal correlation matrix means that the segmented feature vectors are first aggregated according to the types of motion segments (turning segments, acceleration segments, deceleration segments, constant-speed segments), and then the similarity data of different segments are filled in the same matrix according to the window index or time order. If there is a certain similarity pattern between the temporal consistency scores of a turning segment and those of an acceleration segment, a relatively high correlation entry will appear at the corresponding matrix position; if the segmented feature vectors of the deceleration segment and the constant-speed segment are significantly different, a relatively low correlation value will appear in the corresponding matrix. To reflect the motion feature differences of electric bicycles, different weight coefficients also need to be set for this correlation matrix to generate a weighted correlation matrix. When weighting, different weights can be assigned according to the attention to turning continuity within the turning segment or the attention to speed transition within the acceleration segment, so as to strengthen the information correlation degree in the high-dynamic range and make the subsequent temporal mapping pay more attention to the data segments with sharp changes.

[0055] In this solution, the temporal mapping cost function is calculated based on the weighted correlation matrix and solved with the constraints of turning continuity and speed smoothness. This process involves the precise modeling of the time alignment deviation and dynamic changes of different motion segments. First of all, the goal of the temporal mapping cost function is to optimize the time alignment between multiple motion segments (such as turning segments, acceleration segments, deceleration segments, and constant-speed segments). The core idea of this cost function is to evaluate the quality of temporal alignment by quantifying the deviation between different motion segments on the time axis in each window.

[0056] The cost function usually takes into account the time alignment error between windows. Specifically, if there are inconsistent fluctuation peaks in a turning segment in multiple windows, this may imply a time mismatch. The turning segment generally refers to the change in the wheel rotation angle when an electric bicycle turns, and it should present a continuous and smooth curve in the time series. If the fluctuation peaks of each window are different, it means that the time relationship between these data points has not been correctly matched. Similarly, if the acceleration segment or deceleration segment shows an inconsistent time distribution of speed changes in multiple windows, it will also affect the temporal alignment and thus increase the cost. Therefore, the construction of the cost function will comprehensively consider the time alignment error between each window, especially the temporal consistency of the turning segment and the acceleration / deceleration segments.

[0057] Next, through iterative solution or search of an optimization algorithm, the cost function can find the optimal timing mapping function, enabling a higher degree of temporal matching for the angular continuity of the turning segment and the speed smoothness of the acceleration and deceleration segments. For example, in the turning segment, the turning angle of the vehicle should change smoothly to avoid temporal jumps; in the acceleration and deceleration segments, the speed should exhibit a smooth transition to avoid sudden changes. If the value of the cost function is high, it indicates a significant time alignment problem and further optimization is required. When the cost function converges, it means that the timing mapping has reached the optimal state, effectively ensuring turning continuity and speed smoothness.

[0058] Finally, after obtaining the optimal timing mapping function, resampling calculations are performed in combination with the characteristic weights of each motion segment. This process resamples the data of each motion segment according to the dynamic characteristics of different motion segments to achieve timing alignment. For example, in the turning segment, if the weight of this segment is high, more interpolation points need to be invested in this segment during resampling, or the time scale needs to be divided with a finer granularity to ensure the timing alignment accuracy of this segment. For the constant-speed segment, since its change is relatively stable and the weight is low, a more relaxed sampling mapping method can be used during resampling, thereby reducing unnecessary computational overhead and ensuring the real-time performance and computational efficiency of the system.

[0059] Overall, through this series of steps, the timing correction parameters can effectively reflect the dynamic characteristics of the electric bicycle in different motion stages, providing accurate and real-time data support for multi-sensor fusion, thereby enhancing the system's adaptability to complex motion scenarios and prediction accuracy.

[0060] Please continue to refer to Figure 1 , and perform non-linear interpolation and timing remapping on the data sequence to be corrected according to the timing correction parameters to obtain a corrected sensor data sequence; In an embodiment of the present invention, the performing non-linear interpolation and timing remapping on the data sequence to be corrected according to the timing correction parameters to obtain a corrected sensor data sequence includes: Classify the data sequence to be corrected according to the motion characteristics of the electric bicycle, divide the data into turning process data, acceleration and deceleration process data, and constant-speed process data, and perform corresponding grouping on the timing correction parameters to obtain classified correction parameters; Perform non-uniform division on the time axis of the turning process data, where the division density is proportional to the turning angle change rate, and perform variable-density division on the time axis of the acceleration and deceleration process data, where the division density is proportional to the acceleration amplitude, to generate a differential time grid; Interpolating control points are generated according to the differential time grid and the classification correction parameters, and different interpolation constraint conditions are set for different types of data, where the control point interval of the steering process data is t1, the control point interval of the acceleration / deceleration process data is t2, and the control point interval of the constant-speed process data is t3; A piecewise non-linear interpolation function is constructed based on the interpolating control points. Cubic spline interpolation is used for the steering process data, quadratic spline interpolation is used for the acceleration / deceleration process data, and linear interpolation is used for the constant-speed process data to obtain a multi-mode interpolation function; A time series mapping matrix is calculated according to the multi-mode interpolation function, and the classification correction parameters are substituted into the time series mapping matrix to generate a correction remapping function; The correction remapping function is used to resample and correct the timestamps of the data sequence to be corrected, and the corrected sensor data sequence is obtained.

[0061] Specifically, in the implementation process, the data sequence to be corrected needs to be classified according to the motion characteristics of the electric bicycle first. The motion characteristics include the steering process, the acceleration / deceleration process, and the constant-speed process. During the steering process, the angle changes greatly, showing obvious fluctuations, especially when the vehicle turns; the acceleration / deceleration process shows large fluctuations in acceleration or deceleration; while the data in the constant-speed process is relatively stable with small speed changes. The requirements for time series correction of these three different motion processes are different, so corresponding data classification must be carried out. The classified data are used as the steering process data, the acceleration / deceleration process data, and the constant-speed process data for subsequent processing respectively. In this process, the classification basis includes the data collected by the sensors, such as wheel angle, acceleration, speed change and other information. By analyzing these information, different motion processes can be accurately divided.

[0062] The classified data is further used to generate classification correction parameters. These correction parameters set corresponding correction methods for each data segment. The dynamic changes in the steering segment and the acceleration / deceleration segment are more significant, and more refined correction is required to improve the time series accuracy; while the changes in the constant-speed segment are smaller, and the accuracy requirements for correction are relatively loose. Therefore, the setting of the classification correction parameters will directly affect the fineness of the subsequent interpolation and resampling processes.

[0063] The next step involves non-uniform partitioning of the time axis for different motion processes. This step aims to improve the calibration accuracy through differentiated time grids. During the turning process, the change in wheel angle is relatively drastic. Therefore, it is necessary to ensure an accurate description of the turning process by making the partitioning density of the time axis proportional to the change rate of the steering angle. Specifically, when implementing, the change rate of the steering angle can be calculated at each time point, and then the density of the time grid can be dynamically adjusted, so that moments with larger angle changes have a higher sampling density, thereby capturing more details. Similarly, during the acceleration and deceleration processes, due to the large amplitude of acceleration changes, the partitioning density of the time grid needs to be adjusted according to the magnitude of the acceleration amplitude. When the acceleration changes drastically, the density of the time grid is larger, thus more accurately reflecting the changes in the acceleration and deceleration processes. The change in the constant-speed segment is relatively stable, so a time grid with a lower density is adopted to reduce the intervention. In this way, it can be ensured that the data of different motion processes are fully and accurately represented in time series.

[0064] After generating the differentiated time grids, the next step is to generate interpolation control points through these time grids and classification calibration parameters. Interpolation control points play a crucial role in time series calibration, and the setting of the control points affects the accuracy of the subsequent interpolation results. Specifically, for the turning process data, since the angle change is relatively drastic, the interval of the interpolation control points is shorter, set to t1; during the acceleration and deceleration processes, since the acceleration change is relatively significant, the interval of the interpolation control points is t2; while for the constant-speed segment, due to the smaller change, the control point interval is set to t3. By setting the control point intervals, dense sampling can be carried out at moments requiring high precision, and the sampling density can be reduced at moments with smaller changes, thereby optimizing the computational cost.

[0065] After setting the interpolation control points, the next step is to construct a non-linear interpolation function. Cubic spline interpolation is used for the turning process data because cubic splines can ensure the second-order continuity of the interpolation function at each point, avoiding discontinuities when the steering angle changes greatly. Quadratic spline interpolation is used for the acceleration and deceleration process data. Although quadratic spline interpolation is not as smooth as cubic spline interpolation, it can fully reflect the change characteristics of the acceleration and deceleration segments, and its computational complexity is relatively low. Linear interpolation is used for the constant-speed process data. Linear interpolation is suitable for the constant-speed segment because the data change in this segment is small, and linear interpolation can effectively meet its accuracy requirements while avoiding unnecessary computational overhead.

[0066] Through the above steps, a multi-mode interpolation function is obtained. This function can adopt different interpolation methods according to the dynamic characteristics of different types of data to ensure the accuracy of time series. Then, based on these interpolation functions, a time series mapping matrix is calculated. After substituting the classification correction parameters into the matrix, a corrected remapping function is obtained. This function is responsible for mapping the original data sequence to the target time series, minimizing the time series error and making the time alignment between data more accurate.

[0067] Finally, the corrected remapping function is used to resample and correct the timestamps of the data sequence to be corrected. Resampling is achieved by generating new data points on the new time axis to ensure that the data at each time point is accurately reflected in the time series. Timestamp correction is to correct the timestamps of the original data through the interpolation function to make the time series relationship of the data more accurate. Through this series of steps, a corrected sensor data sequence is finally generated. These corrected data have higher time series accuracy and can provide more accurate data support for subsequent multi-sensor fusion.

[0068] In an embodiment of the present invention, the non-uniform division of the time axis of the steering process data, where the division density is proportional to the steering angle change rate, and the variable density division of the time axis of the acceleration and deceleration process data, where the division density is proportional to the acceleration amplitude, to generate a differential time grid, includes: Perform a motion scenario analysis on the steering process data and the acceleration and deceleration process data, divide the steering process data into large-angle turning segments, small-angle turning segments, and transition turning segments, and divide the acceleration and deceleration process data into emergency braking segments, slow deceleration segments, and uniform acceleration segments to obtain a scene segmentation sequence; Calculate the steering angle change rate and the acceleration amplitude according to the scene segmentation sequence, assign a weight coefficient W1 to the steering angle change rate of the large-angle turning segment, assign a weight coefficient W2 to the steering angle change rate of the small-angle turning segment, and assign a weight coefficient W3 to the steering angle change rate of the transition turning segment to obtain a weighted steering feature; Assign a weight coefficient V1 to the acceleration amplitude of the emergency braking segment, assign a weight coefficient V2 to the acceleration amplitude of the slow deceleration segment, and assign a weight coefficient V3 to the acceleration amplitude of the uniform acceleration segment to obtain a weighted acceleration feature; Calculate the division step size of the steering process according to the weighted steering feature. The step size is inversely proportional to the weighted steering feature to generate a steering process time grid; Calculate the division step size of the acceleration and deceleration process according to the weighted acceleration feature. The step size is inversely proportional to the weighted acceleration feature to generate an acceleration and deceleration process time grid; According to the motion continuity constraint of the electric bicycle, perform boundary optimization and merging on the steering process time grid and the acceleration and deceleration process time grid to generate a differential time grid.

[0069] Specifically, during the implementation of timing correction, it is first necessary to perform a motion scenario analysis on the steering process data and acceleration / deceleration process data of the electric bicycle. The purpose of this analysis is to subdivide the data in different motion processes for more accurate subsequent timing division and correction. Specifically, for the steering process data, it needs to be divided into a large-angle turning section, a small-angle turning section, and a transitional turning section according to the change of the steering angle. The large-angle turning section usually indicates that the vehicle makes a sharp turn with a large change in the steering angle; the small-angle turning section is a turn within a smaller steering range with a relatively small angle change; the transitional turning section is in the transitional stage of steering with a relatively gentle and unstable angle change. Therefore, by calculating the change in the steering angle at each time point, the steering data can be accurately classified into the corresponding motion scenarios.

[0070] Similarly, for the acceleration / deceleration process data, relying on the change in the vehicle's acceleration, the data can be divided into an emergency braking section, a slow deceleration section, and a uniform acceleration section. The emergency braking section refers to the process where, during braking or deceleration, the vehicle generates a large negative acceleration and the speed rapidly decreases; the slow deceleration section means that the vehicle decelerates relatively smoothly with a gentle change in acceleration; the uniform acceleration section is the stage of uniform change during the vehicle's acceleration process with a stable acceleration. Therefore, in the process of classifying the data of these two processes, first, based on the acceleration characteristics of the vehicle, the data is gradually allocated to their respective acceleration or deceleration scenarios.

[0071] Once the scenario division is completed, the next step is to perform weighted processing on the data of the steering process and the acceleration / deceleration process according to the characteristics of each type of scenario. Since the large-angle turning section, small-angle turning section, and transitional turning section in the steering process have different amplitudes of angle change, different weight coefficients need to be assigned. Specifically, the large-angle turning section, due to its higher requirement for vehicle control, needs to be assigned a larger weight coefficient W1, so that more attention can be paid to the accuracy of these key turning processes during timing correction; the small-angle turning section, due to its smaller impact on the vehicle, is assigned a lower weight coefficient W2; while the transitional turning section, due to its more complex dynamic changes, is assigned a weight coefficient W3 at an intermediate value. Through this weighted method, appropriate attention can be ensured for different types of steering processes.

[0072] Similarly, in the data of the acceleration and deceleration process, different weight coefficients also need to be assigned to the acceleration amplitudes in the emergency braking section, slow deceleration section, and uniform acceleration section. In the emergency braking section, since it involves a large amount of deceleration and may generate large timing fluctuations, a relatively high weight coefficient V1 needs to be assigned; the slow deceleration section is relatively stable, and the weight coefficient V2 is assigned; the uniform acceleration section requires a more stable time allocation, and the weight coefficient V3 is assigned. Through this weighting method, more attention can be paid to the critical moments and dynamic changes during the acceleration and deceleration process in the subsequent timing division and interpolation process.

[0073] The weighted steering characteristics and acceleration characteristics will directly affect the time division of the steering process and the acceleration and deceleration process. Specifically, the time division step of the steering process is inversely proportional to the weighted steering characteristics. Since the steering angle changes greatly in the large-angle turning section, its step is small, that is, the sampling points in each time period are denser, ensuring the accuracy of steering; while in the small-angle turning section and the transition turning section, due to their smaller changes, the step is relatively large, reducing the sampling points, thereby reducing the computational complexity. In this way, the adjustment of the time step can accurately control the data density during the steering process and ensure that the dynamic characteristics of the steering process are fully reflected.

[0074] Similarly, the division step of the acceleration and deceleration process is also inversely proportional to the weighted acceleration characteristics. In the emergency braking section and the uniform acceleration section, due to their large changes, the step is small, increasing the sampling density to more accurately reflect the change of acceleration; while in the slow deceleration section, due to its slow change, the step is large, thus reducing the unnecessary computational burden. In this way, the time division can be finely adjusted according to the change of acceleration to ensure that the data density of the acceleration and deceleration process matches the dynamic characteristics.

[0075] After obtaining the time grids of the steering process and the acceleration and deceleration process, the next step is to optimize and merge the boundaries of these time grids according to the motion continuity constraint of the electric bicycle. The motion of the electric bicycle has a certain continuity, especially during the transition process of steering and acceleration and deceleration, the time boundaries of the data should not be too abrupt. Therefore, it is necessary to optimize the boundaries of the time grids to ensure a smooth transition between each process and merge adjacent time grids to reduce computational redundancy. By optimizing the boundaries, the timing relationship between each motion process can be effectively ensured not to be disrupted, and the unreasonable gaps between the steering section and the acceleration and deceleration section can be avoided.

[0076] Finally, after boundary optimization and merging, the resulting differential time grid can accurately reflect the dynamic changes in the steering, acceleration, deceleration, and constant-speed processes according to the motion characteristics and timing requirements of the electric bicycle. This differential time grid can not only improve the time alignment accuracy of the data but also provide a more accurate time basis for subsequent interpolation and resampling, ensuring that the data after timing correction achieves the best balance in terms of accuracy and computational efficiency.

[0077] Please continue to refer to Figure 1 , extract target features based on the corrected sensor data sequence, calculate the collision risk level according to the motion feature sequence and the target features, and generate a warning signal.

[0078] In an embodiment of the present invention, the extracting target features based on the corrected sensor data sequence, calculating the collision risk level according to the motion feature sequence and the target features, and generating a warning signal includes: Perform target detection on the corrected sensor data sequence, divide the detection scene into a turning scene, a rapid start-stop scene, and a straight-line scene according to the motion feature sequence, and extract position features and motion features of the target in different scenes to obtain scene-based target features; Classify the target according to the scene-based target features, focus on the laterally approaching target in the turning scene, focus on the longitudinally moving target in the rapid start-stop scene, and focus on the omnidirectional moving target in the straight-line scene to generate a target threat degree index; Perform coupled calculation on the target threat degree index and the motion feature sequence, use the steering angular velocity as a weighting factor in the turning scene, use the acceleration as a weighting factor in the rapid start-stop scene, and use the speed as a weighting factor in the straight-line scene to obtain a scene weight coefficient; Divide the target into an emergency avoidance area, a steering avoidance area, a deceleration avoidance area, and a safe area according to the scene weight coefficient, assign a collision risk level to different areas, and generate a partition risk level sequence; Match the partition risk level sequence with the maneuverability parameter of the electric bicycle, issue an audible and visual warning with an interval of P to the emergency avoidance area, issue an audible and visual warning with an interval of 2P to the steering avoidance area, and issue an audible and visual warning with an interval of 4P to the deceleration avoidance area, where P is a basic warning cycle parameter, to obtain a partition warning signal; Fuse the partition warning signals to generate a warning signal.

[0079] Specifically, during the time series correction and target detection of an electric bicycle, it is first necessary to perform target detection on the corrected sensor data sequence. The purpose of target detection is to identify and classify different motion scenarios based on the motion feature sequence of the electric bicycle, and further analyze the target features in each scenario. The motion scenarios are mainly divided into turning scenarios, rapid start-stop scenarios, and straight-line scenarios. In the turning scenario, the motion of the electric bicycle is mainly manifested as changes in the steering angle; in the rapid start-stop scenario, it is mainly manifested as rapid acceleration and sudden braking; while in the straight-line scenario, the electric bicycle travels at a constant speed along a straight path. According to these different motion characteristics, the detection scenarios are divided into these three categories, and different position features and motion features of the target are extracted according to the different scenarios.

[0080] Specifically, the targets in the turning scenario mainly show lateral approach, and these targets may appear on the side of the rider when the electric bicycle turns. Therefore, in the turning scenario, the system will focus on the targets approaching the electric bicycle from the side, and these targets may pose a greater risk of avoidance. In the rapid start-stop scenario, the speed of the electric bicycle changes rapidly. Therefore, the main focus is on the targets moving longitudinally, especially those that may collide longitudinally with the electric bicycle, such as vehicles or pedestrians braking suddenly ahead. In the straight-line scenario, the motion of the electric bicycle is relatively stable, and the targets are omnidirectional. This means that attention should be paid to the targets that may approach the electric bicycle from all directions, such as those in front, behind, or on the left and right. The threat level of these targets will be evaluated according to their approaching methods and speeds, so as to take appropriate countermeasures in different scenarios.

[0081] After the target extraction and classification are completed, the next step is to calculate the threat level index of the target based on the scenario-based target features. In the turning scenario, the threat level of the target is usually closely related to the lateral approach speed of the target. Therefore, the steering angular velocity is introduced as a weighting factor. The greater the steering angular velocity, the higher the threat level of the target. In the rapid start-stop scenario, due to the large acceleration of the electric bicycle, especially in the case of sudden braking, the acceleration becomes the weighting factor, and the threat level of the target is proportional to the amplitude of the acceleration. In the straight-line scenario, since the speed of the electric bicycle is relatively stable, the speed is used as the weighting factor to calculate the threat level of the target. The greater the speed, the corresponding increase in the threat level of the target. By performing weighted calculations on the targets in each scenario, the threat level index of each target can be obtained, and this index can be further used to judge the potential risk of the target.

[0082] Next, according to the obtained threat degree index, the target will be divided into different avoidance areas. According to the characteristics of different scenarios and the threat degree of the target, the target is divided into an emergency avoidance area, a steering avoidance area, a deceleration avoidance area, and a safety area. In the emergency avoidance area, due to the high threat degree of the target, which may pose a greater threat to the electric bicycle, immediate emergency avoidance measures need to be taken. In the steering avoidance area, the electric bicycle may need to make a certain degree of steering to avoid the target. Therefore, the threat degree in this area is slightly lower than that in the emergency avoidance area, but a relatively rapid response is still required. In the deceleration avoidance area, the electric bicycle can avoid colliding with the target by decelerating. The threat degree in this area is relatively low, and the avoidance measures can be appropriately relaxed. In the safety area, the threat degree of the target is low, and no avoidance is required, and the vehicle can continue to travel along the normal trajectory.

[0083] Each avoidance area will be assigned different collision risk levels according to its threat degree. Areas with higher collision risk levels will require shorter warning interval times, while areas with lower collision risk levels can set longer warning intervals. Specifically, the emergency avoidance area will issue an audible and visual warning with an interval time of P, which means that in this area, the system needs to give a higher frequency of warnings to prompt the driver to make a quick response; the steering avoidance area will issue a warning with an interval time of 2P, indicating that the warning needs to be issued at a slightly longer time interval; and the deceleration avoidance area will issue a warning with an interval time of 4P, and the time interval is further extended because the avoidance requirements in this area are relatively low. Here, P is the basic warning cycle parameter, and the specific value of P is set according to the motion state of the electric bicycle and the threat degree of the target. In this way, the threat degrees of different areas and the corresponding warning intervals are precisely matched.

[0084] Finally, the warning signals of all partitions will be fused to generate the final warning signal. This warning signal is a comprehensive output that can dynamically adjust the intensity and frequency of the warning according to the threat degree of the target, the characteristics of the scenario, and the current motion state of the electric bicycle, ensuring that the driver can receive the most effective avoidance information in a timely manner in any situation. Through this refined target detection and threat assessment, the electric bicycle can make more intelligent avoidance decisions in a complex traffic environment, thereby improving the driving safety and real-time response ability.

[0085] The method for preventing collisions of an electric bicycle and the electric bicycle in the embodiments of the present invention have been described above. Next, the AA device in the embodiments of the present invention will be described. Please refer to Figure 2 , an embodiment of the AA device in the embodiments of the present invention includes: A data acquisition module 101, configured to obtain an acceleration sequence and an angular velocity sequence of an electric bicycle, generate a motion feature sequence through frequency domain analysis according to the acceleration sequence and the angular velocity sequence, and calculate a time series drift prediction index according to the motion feature sequence; A hierarchical processing module 102, configured to perform hierarchical processing on multi-sensor sampling data according to the time series drift prediction index to generate a data sequence to be corrected, where the hierarchical processing includes using a ratio of a difference between sampling frequencies to a preset frequency threshold as a priority parameter, and compensating the data according to the priority parameter; A data segmentation module 103, configured to perform time window segmentation on the data sequence to be corrected, generate a time series correlation matrix based on local features in each time window, calculate an optimal time series mapping relationship according to the time series correlation matrix, and obtain a time series correction parameter; A data correction module 104, configured to perform non-linear interpolation and time series remapping on the data sequence to be corrected according to the time series correction parameter to obtain a corrected sensor data sequence; A risk warning module 105, configured to extract target features based on the corrected sensor data sequence, calculate a collision risk level according to the motion feature sequence and the target features, and generate a warning signal.

[0086] The above are only the preferred embodiments of the present invention, and do not limit the patent scope of the present invention. Any equivalent structural transformation made by using the description and drawings of the present invention under the inventive concept of the present invention, or direct / indirect application in other related technical fields, is included in the patent protection scope of the present invention.

Claims

1. A method for preventing collision of an electric bicycle, characterized in that: include: Obtaining an acceleration sequence and an angular velocity sequence of the electric bicycle, generating a motion feature sequence through frequency domain analysis according to the acceleration sequence and the angular velocity sequence, and calculating a timing drift prediction index according to the motion feature sequence; Performing hierarchical processing on the multi-sensor sampling data according to the timing drift prediction index to generate a data sequence to be corrected, wherein the hierarchical processing includes taking the ratio of the difference of the sampling frequency to the preset frequency threshold as a priority parameter, and compensating the data according to the priority parameter; The data sequence to be corrected is divided into time windows, a time series correlation matrix is ​​generated based on local features in each time window, an optimal time series mapping relationship is calculated according to the time series correlation matrix, and a time series correction parameter is obtained; Performing nonlinear interpolation and time sequence remapping on the to-be-corrected data sequence according to the time sequence correction parameters to obtain a corrected sensor data sequence; Target features are extracted based on the corrected sensor data sequence, a collision risk level is calculated according to the motion feature sequence and the target features, and a warning signal is generated.

2. The electric bicycle collision prevention method according to claim 1, characterized in that: The obtaining of the acceleration sequence and angular velocity sequence of the electric bicycle, generating a motion feature sequence through frequency domain analysis according to the acceleration sequence and the angular velocity sequence, and calculating the timing drift prediction index according to the motion feature sequence, comprises: Obtaining an acceleration sequence and an angular velocity sequence, and performing dual-frequency sampling on the acceleration sequence and the angular velocity sequence according to a preset sampling rule to obtain a high-frequency sequence segment and a low-frequency sequence segment; Performing Fourier transformation on the high-frequency sequence fragments and the low-frequency sequence fragments to obtain frequency domain feature data, and separating a high-frequency component representing a sharp turn feature and a low-frequency component representing a speed change from the frequency domain feature data; Calculate a sudden steering change parameter according to the high-frequency component, calculate an acceleration / deceleration parameter according to the low-frequency component, and combine the sudden steering change parameter and the acceleration / deceleration parameter to obtain a motion state feature; Performing time series correlation analysis on the motion state characteristics, calculating the rate of change of characteristic values ​​in adjacent time windows, and obtaining a state transition characteristic sequence; Identify the motion mode of the electric bicycle according to the state transition feature sequence, match the identified motion mode based on a preset motion mode feature library, and generate a motion feature sequence; Weight coefficients are set for the sudden turn parameters and acceleration / deceleration parameters in the motion feature sequence respectively, and weighted calculation is performed according to the weight coefficients and a preset drift risk threshold to obtain a time series drift prediction index.

3. The electric bicycle collision prevention method according to claim 2, characterized in that: The step of identifying the motion mode of the electric bicycle according to the state transition feature sequence, matching the identified motion mode based on a preset motion mode feature library, and generating a motion feature sequence includes: Dividing the state transition feature sequence into time windows, calculating a fluctuation parameter according to fluctuations of feature values ​​within the time window, and dividing the state transition feature sequence into a plurality of feature segments according to the fluctuation parameter; The characteristic segments are hierarchically clustered according to their amplitudes to obtain hierarchical characteristic groups, and characteristic templates are generated according to the temporal distribution rules of the hierarchical characteristic groups; Performing similarity matching between the feature template and a reference feature template in the preset motion pattern feature library to obtain a pattern matching score; Classifying the sports scenes according to the pattern matching scores, and dividing the sports scenes into straight-line riding scenes, turning riding scenes, accelerating riding scenes, and decelerating riding scenes; Calculating characteristic statistical parameters for the state transition characteristic sequences in different riding scenarios respectively, and generating a scene characteristic vector according to the time series variation trend of the characteristic statistical parameters; The scene feature vectors are concatenated according to a temporal relationship to generate a motion feature sequence.

4. The electric bicycle collision prevention method according to claim 1, characterized in that: The step of performing hierarchical processing on the multi-sensor sampling data according to the timing drift prediction index to generate a data sequence to be corrected, wherein the hierarchical processing includes taking the ratio of the difference of the sampling frequency to the preset frequency threshold as a priority parameter, and compensating the data according to the priority parameter, including: Obtaining a sampling time series of multi-sensor sampling data, performing correlation analysis on the sampling time series and the time series drift prediction index, and dividing the data into sharp turn scenario data, fast start-stop scenario data, and smooth riding scenario data according to the correlation degree; Calculating sampling frequencies for the sharp turn scenario data, the fast start-stop scenario data, and the steady riding scenario data respectively to obtain a scenario-specific sampling frequency matrix; Calculate the sampling frequency difference according to the sub-scenario sampling frequency matrix, perform a ratio operation on the sampling frequency difference and the preset frequency threshold to obtain a basic priority parameter; The basic priority parameter is weighted and adjusted according to the timing drift prediction index, a weight coefficient K1 is assigned to the sharp turn scene data, a weight coefficient K2 is assigned to the fast start and stop scene data, and a weight coefficient K3 is assigned to the smooth riding scene data, to obtain a priority parameter; The multi-sensor sampling data is compensated in sections according to the priority parameter, wherein compensation is performed every N data points in the sharp turn scene data, compensation is performed every 2N data points in the fast start-stop scene data, and compensation is performed every 4N data points in the smooth riding scene data, where N is a basic compensation interval parameter, to obtain compensated data; The compensated data are reorganized in time sequence to generate a data sequence to be corrected.

5. The method for preventing collision of an electric bicycle according to claim 4, characterized in that: The basic priority parameter is weightedly adjusted according to the timing drift prediction index, a weight coefficient K1 is assigned to the sharp turn scene data, a weight coefficient K2 is assigned to the fast start and stop scene data, and a weight coefficient K3 is assigned to the smooth riding scene data, to obtain the priority parameters, including: The steering process is segmented according to the change rate of the time series drift prediction index, and the correlation coefficient between the mean angular velocity of each steering segment and the drift prediction index is calculated to obtain the steering segment drift risk index; According to the cross-correlation value of the time series drift prediction index and the acceleration sequence, the fluctuation characteristics of the drift prediction index during acceleration and deceleration are calculated to obtain the start-stop drift risk index; Based on the turning section drift risk index, nonlinear mapping is performed on the basic priority parameter to generate the weighting coefficient K1; Based on the start-stop segment drift risk index, segmentally map the basic priority parameter to generate the weighting coefficient K2; Performing linear mapping on the basic priority parameter according to the steady-state value of the timing drift prediction index to generate the weighting coefficient K3; The weighting coefficient K1, the weighting coefficient K2 and the weighting coefficient K3 are weighted and superimposed on the data of the corresponding scene to obtain the priority parameter.

6. The electric bicycle collision prevention method according to claim 1, characterized in that: The step of dividing the data sequence to be corrected into time windows, generating a time series correlation matrix based on local features in each time window, and calculating an optimal time series mapping relationship according to the time series correlation matrix to obtain a time series correction parameter includes: Calculating the steering angular acceleration and the speed change rate for the data sequence to be corrected, identifying the sharp turning point and the rapid start-stop point according to the steering angular acceleration and the speed change rate, and dividing the data sequence to be corrected into a turning section, an acceleration section, a deceleration section and a constant speed section; According to the characteristics of various motion segments, differentiated window expansion strategies are adopted for different segments. The window of the turning segment is expanded with a step length of T, the windows of the acceleration segment and the deceleration segment are expanded with a step length of 2T, and the window of the uniform speed segment is expanded with a step length of 4T, where T is the basic expansion step length parameter, to generate a multi-scale window sequence; Performing wavelet transform on the data in the multi-scale window sequence, extracting energy distribution features of different frequency bands, calculating the temporal consistency scores between windows according to the energy distribution features, and generating a segmented feature vector; The segmented feature vectors are combined to construct a block time series correlation matrix, and different weight coefficients are set for the block time series correlation matrix according to the motion characteristics of the electric bicycle to obtain a weighted correlation matrix; Calculating the timing mapping cost function according to the weighted correlation matrix, solving the optimal timing mapping function with steering continuity and speed smoothness as constraints; The data of each motion segment is resampled and calculated according to the optimal timing mapping function, and the timing correction parameters are generated in combination with the feature weights of the corresponding motion segments.

7. The method for preventing collision of an electric bicycle according to claim 1, characterized in that: The step of performing nonlinear interpolation and time sequence remapping on the to-be-corrected data sequence according to the time sequence correction parameters to obtain a corrected sensor data sequence includes: Classifying the data sequence to be corrected according to the motion characteristics of the electric bicycle, dividing the data into steering process data, acceleration and deceleration process data and uniform speed process data, and grouping the timing correction parameters accordingly to obtain the classified correction parameters; The time axis of the steering process data is non-uniformly divided, and the division density is proportional to the steering angle change rate; the time axis of the acceleration and deceleration process data is variable-density divided, and the division density is proportional to the acceleration amplitude, to generate a differentiated time grid; Generate interpolation control points according to the differentiated time grid and the classification correction parameters, and set different interpolation constraints for different types of data, wherein the control point interval of the steering process data is t1, the control point interval of the acceleration and deceleration process data is t2, and the control point interval of the uniform speed process data is t3; A piecewise nonlinear interpolation function is constructed based on the interpolation control points, cubic spline interpolation is used for the steering process data, quadratic spline interpolation is used for the acceleration and deceleration process data, and linear interpolation is used for the uniform speed process data to obtain a multi-mode interpolation function; Calculating a time series mapping matrix according to the multi-mode interpolation function, substituting the classification correction parameters into the time series mapping matrix, and generating a correction remapping function; The correction remapping function is used to resample and timestamp the data sequence to be corrected to obtain the corrected sensor data sequence.

8. The method for preventing collision of an electric bicycle according to claim 7, characterized in that: The time axis of the steering process data is non-uniformly divided, the division density is proportional to the steering angle change rate, the time axis of the acceleration and deceleration process data is variable density divided, the division density is proportional to the acceleration amplitude, and a differentiated time grid is generated, including: Performing motion scene analysis on the steering process data and the acceleration / deceleration process data, dividing the steering process data into a large-angle turning segment, a small-angle turning segment, and a transition turning segment, and dividing the acceleration / deceleration process data into an emergency braking segment, a slow deceleration segment, and a uniform acceleration segment, to obtain a scene segment sequence; Calculating the steering angle change rate and the acceleration amplitude according to the scene segment sequence, assigning a weight coefficient W1 to the steering angle change rate of the large-angle turning segment, assigning a weight coefficient W2 to the steering angle change rate of the small-angle turning segment, and assigning a weight coefficient W3 to the steering angle change rate of the transition turning segment, to obtain a weighted steering feature; A weight coefficient V1 is assigned to the acceleration amplitude of the emergency braking section, a weight coefficient V2 is assigned to the acceleration amplitude of the slow deceleration section, and a weight coefficient V3 is assigned to the acceleration amplitude of the uniform acceleration section, to obtain a weighted acceleration feature; Calculating the division step length of the steering process according to the weighted steering feature, wherein the step length is inversely proportional to the weighted steering feature, and generating a steering process time grid; Calculating the division step length of the acceleration and deceleration process according to the weighted acceleration feature, wherein the step length is inversely proportional to the weighted acceleration feature, and generating a time grid of the acceleration and deceleration process; According to the motion continuity constraint of the electric bicycle, the steering process time grid and the acceleration / deceleration process time grid are boundary optimized and merged to generate a differentiated time grid.

9. The electric bicycle collision prevention method according to claim 1, characterized in that: The step of extracting target features based on the corrected sensor data sequence, calculating a collision risk level according to the motion feature sequence and the target features, and generating a warning signal includes: Performing target detection on the corrected sensor data sequence, dividing the detection scene into a turning scene, a fast start-stop scene, and a straight-ahead scene according to the motion feature sequence, extracting position features and motion features of the targets in different scenes to obtain scenario-based target features; Classifying the targets according to the scenario-based target features, focusing on lateral approaching targets in turning scenarios, focusing on longitudinal moving targets in fast start-stop scenarios, and focusing on omnidirectional moving targets in straight-ahead scenarios, and generating a target threat index; The target threat index and the motion feature sequence are coupled and calculated, the steering angular velocity is used as a weighting factor in a turning scenario, the acceleration is used as a weighting factor in a fast start-stop scenario, and the speed is used as a weighting factor in a straight-ahead scenario, to obtain a scenario weight coefficient; Dividing the target into an emergency avoidance zone, a steering avoidance zone, a deceleration avoidance zone and a safety zone according to the scenario weight coefficient, assigning collision risk levels to different zones, and generating a zone risk level sequence; The partition risk level sequence is matched with the maneuverability parameter of the electric bicycle, and an acoustic and light warning with an interval time of P is issued for the emergency avoidance zone, an acoustic and light warning with an interval time of 2P is issued for the turning avoidance zone, and an acoustic and light warning with an interval time of 4P is issued for the deceleration avoidance zone, where P is a basic warning cycle parameter, to obtain a partition warning signal; The partition warning signals are fused to generate a warning signal.

10. An electric bicycle, characterized in that: The electric bicycle comprises: A data acquisition module, used to obtain an acceleration sequence and an angular velocity sequence of the electric bicycle, generate a motion feature sequence through frequency domain analysis according to the acceleration sequence and the angular velocity sequence, and calculate a timing drift prediction index according to the motion feature sequence; A hierarchical processing module, used for hierarchically processing the multi-sensor sampling data according to the timing drift prediction index to generate a data sequence to be corrected, wherein the hierarchical processing includes taking the ratio of the difference of the sampling frequency to the preset frequency threshold as a priority parameter, and compensating the data according to the priority parameter; A data segmentation module is used to segment the data sequence to be corrected into time windows, generate a time series correlation matrix based on local features in each time window, calculate the optimal time series mapping relationship according to the time series correlation matrix, and obtain a time series correction parameter; A data correction module, used for performing nonlinear interpolation and time sequence remapping on the to-be-corrected data sequence according to the time sequence correction parameters to obtain a corrected sensor data sequence; The risk warning module is used to extract target features based on the corrected sensor data sequence, calculate the collision risk level according to the motion feature sequence and the target features, and generate a warning signal.

Images