Two-wheeled vehicle riding safety ai voice warning method and system

Abstract

The present application relates to the technical field of two-wheeled vehicle riding safety assistance and intelligent human-computer interaction, in particular to an AI voice early warning method and system for two-wheeled vehicle riding safety, which faces complex occlusion and lateral clamping scenarios of two-wheeled vehicles, fuses multiple source data to construct a channel collapse degree and an occlusion potential threat index, and identifies an escapable channel collapse state; a risk field and a virtual target probability distribution are introduced at the edge of the occlusion, an invisible risk is explicitly quantified and counted into a risk score and a level; in combination with insufficient action accessibility, dynamic regulation of risk triggering, broadcast frequency and suggestion strategy is realized, early warning accuracy, timeliness and usability are improved, false positives and false negatives are reduced, and riding safety is improved.

Description

Technical Field

[0001] This invention relates to the field of two-wheeled vehicle riding safety assistance and intelligent human-computer interaction technology, and more specifically, to an AI voice warning method and system for two-wheeled vehicle riding safety. Background Technology

[0002] Two-wheeled vehicle riding scenarios are characterized by open road boundaries, diverse types of traffic participants, rapid changes in relative speed, and limited lateral space. Riding safety assistance typically relies on multi-source perception data to achieve target detection, tracking, and collision time calculation, combined with road boundaries or lane lines to estimate passable areas, thereby providing risk warnings and voice alerts to a certain extent. This approach, under conditions of clear visibility, stable visible targets, and ample space, can effectively reflect risks such as forward rear-end collisions or typical lane-changing maneuvers, and possesses relatively mature engineering implementation paths and accumulated interaction experience.

[0003] However, in high-frequency scenarios such as urban side intersections, roadside parking areas, bus stops, construction barriers, and curved roads, dynamic obstructions caused by large vehicles or fixed obstacles often occur, rendering potential cross-traffic targets, suddenly appearing vehicles, or non-motorized vehicles unobservable for a period of time. Simultaneously, two-wheeled vehicles may also be subject to lateral restraints, such as being squeezed by left-side guardrails or curbs, being approached by large vehicles traveling at the same speed on the right, being blocked by oncoming traffic, or the road narrowing, rapidly reducing available lateral escape space. At this point, the key risk issue is no longer whether a collision with a visible target will occur, but whether the escape route is collapsing: even if a real threat has not yet been observed, the rider has already lost sufficient space and time to maneuver. Once a small target suddenly appears within the obstructed area, the consequences of the event will be amplified by the restricted passage, creating a highly serious and irreversible risk situation.

[0004] Many existing early warning methods still rely on visible targets as the core triggering basis, such as collision time thresholds, relative distance thresholds, or braking intentions of forward targets. Some methods attempt to incorporate road boundary and traffic space information, but often treat traffic space as a static geometric constraint or comfort indicator, rarely coupling it with occlusion uncertainty in modeling. This makes it particularly difficult to express potential emergent risks in an interpretable and quantifiable way before occlusion is released. Furthermore, when lateral clamping constraints exist, the cyclist's reachability is significantly limited. Highly directional avoidance suggestions may not be executable under low maneuverability or low adhesion conditions. Without action reachability gating, inconsistencies between prompts and executability may arise, affecting the credibility of the interaction and safety. Meanwhile, voice warnings also face practical constraints such as rapidly escalating risks, abrupt changes, and increased cognitive load: overly frequent or information-heavy announcements may distract attention, and the lack of a preemptive mechanism at critical moments may cause the best intervention window to be missed.

[0005] Therefore, the industry is gradually focusing on an assessment framework that better reflects the actual risk mechanism of cycling: under dynamic occlusion and lateral clamping coupling conditions, the risk is elevated from a single visible target collision problem to an escape channel collapse problem. This requires characterizing the urgency of the passage, such as the effective width, effective length, and exit availability of the passage, as well as expressing potential targets in the occlusion area in a probabilistic form. Furthermore, it requires adaptive adjustment of the suggested output and voice strategy in combination with insufficient action accessibility, so as to identify and provide stable early warning of high-severity emerging risks before the actual threat target is observed. Summary of the Invention

[0006] In view of the shortcomings of existing technologies, the purpose of this invention is to provide an AI voice warning method and system for two-wheeled vehicle riding safety.

[0007] To achieve the above objectives, the present invention provides the following technical solution: The AI-powered voice warning method for two-wheeled vehicle riding safety includes the following steps: Step 1: Obtain multi-source data related to risk warnings during two-wheeled vehicle riding, and add timestamps to the multi-source data; Step 2: Preprocess the multi-source data and output the preprocessed data; Step 3: Based on the preprocessed data, perform scene element identification and state estimation to obtain scene perception results that include at least the state of dynamic occlusion, lateral clamping constraint, and passageway. Step 4: Calculate the channel collapse degree index and the occlusion potential threat intensity index based on the scene perception results, and identify whether the escapeable channel collapse state has been entered based on the channel collapse degree index and the occlusion potential threat intensity index; when the escapeable channel collapse state is identified, establish the occlusion edge risk field, generate multiple virtual targets, and assign probability mass to each virtual target to obtain the virtual target risk quantity corresponding to each virtual target. Step 5: Calculate the risk score based on the scene perception results, virtual target risk quantity, and insufficient action accessibility index, and output the risk level according to the risk score; determine the set of dynamic adjustment level parameters according to the risk level, and dynamically adjust the risk triggering mechanism, voice broadcast strategy, and early warning suggestion output strategy under the constraints of the set of dynamic adjustment level parameters. Step 6: Generate and output AI voice warning information corresponding to the risk level based on the risk level and dynamic control results.

[0008] Furthermore, the multi-source data includes at least inertial measurement data to characterize the motion state of the two-wheeled vehicle and environmental perception data to characterize the state of the surrounding environment.

[0009] Furthermore, data preprocessing includes time synchronization, filtering and denoising, coordinate system calibration, and data validity testing of multi-source data.

[0010] Furthermore, the dynamic occlusion state is used to characterize the degree of occlusion of the observability of potential targets and the changes in the shape of the occlusion edge; the lateral clamping constraint state is used to characterize the available space margin on the left and right sides of the two-wheeled vehicle and its sustainability; and the passageway state is used to characterize the effective width, length and exit availability of the available escape passage for the two-wheeled vehicle.

[0011] Furthermore, the channel collapse index is calculated from the effective width of the passageway, the effective length of the passageway, the availability of the passageway exit, and the lateral clamping constraint state.

[0012] Furthermore, the potential threat intensity index of the occlusion is calculated from the dynamic occlusion state, which includes at least the degree of occlusion, the geometric range of the occlusion edge, the relative motion information of the occlusion edge, and the probability of potential targets appearing within the occlusion area.

[0013] Furthermore, when the system identifies an escape route collapsing, it establishes an occlusion edge risk field and generates multiple virtual targets, including: Extract the occlusion edges of dynamic occluders from the scene perception results, and discretize the occlusion edges into several edge sampling points in the two-wheeled vehicle motion coordinate system; Using each edge sampling point as a risk source, and combining the current speed direction of the two-wheeled vehicle, the status of the passage, and the relative motion information of the obstruction, a risk potential function that decays with distance and is anisotropic along the direction of travel is constructed. Based on the risk potential function, an obstruction edge risk field is formed in the unobservable area outside the obstruction edge. Within the high-risk area of ​​the occlusion edge risk field, several candidate points are determined according to a preset spatial resolution or by searching for local extrema, and each candidate point is mapped to the hypothetical location of the virtual target. Based on the field strength value of the occlusion edge risk field at each candidate point, the relative position of the candidate point and the center line of the passage, and the escape channel margin at the candidate point, the probabilistic quality of each virtual target is determined and normalized, thereby generating multiple virtual targets and their corresponding probabilistic qualities.

[0014] Furthermore, the risk score is calculated based on scene perception results, virtual target risk level, and action accessibility insufficiency indicators, including: The basic risk quantity is calculated from the scene perception results. The basic risk quantity includes at least the channel risk quantity obtained from the passage status and the occlusion risk quantity obtained from the dynamic occlusion status. The virtual target risk quantities corresponding to multiple virtual targets are aggregated according to their probabilistic quality to obtain the virtual target aggregated risk quantity; Normalize the accessibility insufficiency index to obtain the accessibility penalty. The basic risk quantity, the virtual target aggregated risk quantity, and the accessibility penalty quantity are weighted, merged, and normalized with a limit to obtain a risk score. The risk score is updated over time and used to output the risk level.

[0015] Furthermore, dynamic regulation includes: When the channel collapse index is greater than the first threshold, the risk triggering mechanism is switched from collision time threshold triggering based on visible targets to passage channel threshold triggering, and the weight gain of virtual target risk quantity and the first threshold are adaptively adjusted according to the risk level. Based on the action accessibility insufficiency index, the early warning suggestions are accessibility-gated. When the action accessibility insufficiency index is greater than the second threshold, the output of directional avoidance suggestions is prohibited and the output of conservative deceleration and preparatory braking suggestions is output. The second threshold and gating intensity are adaptively adjusted according to the risk level. The priority of voice preemption and the frequency of voice broadcasting are dynamically adjusted based on the risk change rate index and the risk jump amplitude index. When the risk change rate index is greater than the third threshold or the risk jump amplitude index is greater than the fourth threshold, a high-priority voice preemption broadcast is triggered. The upper limit of the voice preemption priority and the broadcasting frequency is adaptively adjusted according to the risk level. The amount of information in the speech content is dynamically compressed based on the effective width of the passage and the cognitive load index, and the compression intensity is adaptively adjusted according to the risk level.

[0016] Furthermore, the AI ​​voice warning system for two-wheeled vehicle riding safety includes: The data acquisition module is used to acquire multi-source data related to risk warnings during two-wheeled vehicle riding and add timestamps to the multi-source data; The data preprocessing module is used to preprocess multi-source data and output preprocessed data. The scene perception module is used to identify scene elements and estimate their states based on preprocessed data, and to obtain scene perception results that include at least the states of dynamic occlusion bodies, lateral clamping constraints, and passageways. The coupled disaster state identification module is used to calculate the channel collapse degree index and the occlusion potential threat intensity index based on the scene perception results, and to identify whether the escapeable channel collapse state has been entered based on the channel collapse degree index and the occlusion potential threat intensity index; when the escapeable channel collapse state is identified, an occlusion edge risk field is established, multiple virtual targets are generated, and a probability mass is assigned to each virtual target to obtain the virtual target risk quantity corresponding to each virtual target; The risk assessment and dynamic control module is used to calculate risk scores based on scene perception results, virtual target risk quantity, and insufficient action accessibility indicators, and output risk levels based on risk scores; determine the set of dynamic control level parameters based on risk levels, and dynamically control the risk triggering mechanism, voice broadcast strategy, and early warning suggestion output strategy under the constraints of the dynamic control level parameter set; The voice warning output module is used to generate and output AI voice warning information corresponding to the risk level based on the risk level and dynamic control results.

[0017] Compared with the prior art, the present invention has the following beneficial effects: This invention acquires and preprocesses multi-source data on two-wheeled vehicle riding, completes scene element identification and state estimation, and obtains scene perception results including at least the dynamic occlusion state, lateral clamping constraint state, and passageway state. It then calculates the passage collapse degree index and the occlusion potential threat intensity index, and identifies whether the escapeable passage has collapsed. When this state is entered, the invention establishes an occlusion edge risk field, generates multiple virtual targets, and assigns probability mass to each virtual target to obtain the virtual target risk quantity. Thus, even when the real threat target is still in the occlusion area and has not yet been directly observed, the potential sudden risk can be explicitly introduced into risk assessment in a probabilistic form. This invention elevates the core risk from whether a visible target will collide to whether the escapeway is collapsing and whether there is sudden uncertainty at the occlusion edge, and couples this with the passage urgency. This significantly reduces missed triggers, trigger lags, and trigger jitter during the occlusion release phase, and is particularly suitable for high-frequency and complex scenarios such as side road exits, roadside large vehicle occlusion, and road narrowing guardrail clamping, enabling earlier intervention and more accurate prediction of real risk mechanisms. This invention incorporates an insufficient action accessibility index into risk score calculation. This index, along with scene perception results and virtual target risk quantities, outputs a risk level. Based on the risk level, a set of dynamic adjustment parameters is determined, and the warning suggestion output strategy is dynamically adjusted under the constraints of this parameter set. Since lateral clamping constraints, changes in attachment conditions, and insufficient available maneuverability and braking margins directly limit the achievable lateral displacement and deceleration boundaries of two-wheeled vehicles, this invention uses the insufficient action accessibility index to achieve accessibility gating of the suggestion output: when the degree of insufficient action accessibility is high, it suppresses or prohibits the output of highly directional evasion suggestions that are prone to inducing misoperation, and instead outputs more robust, conservative suggestions such as deceleration, pre-braking, and maintaining vehicle stability; when the degree of insufficient action accessibility is low, it allows the output of more directional evasion suggestions. This mechanism ensures that voice prompts are consistent with the rider's executable action space under specific road geometry, clamping and squeezing conditions, and narrow passages, reducing the mismatch problem of "suggestions that are theoretically reasonable but difficult to execute in reality," and lowering the probability of amplifying risks due to incorrect prompts, thereby improving the engineering usability and safety consistency of warning suggestions. This invention determines a set of dynamic adjustment parameters based on risk levels, and dynamically adjusts the risk triggering mechanism, voice broadcast strategy, and early warning suggestion output strategy under the constraints of this set. Furthermore, it generates and outputs AI-generated voice early warning information corresponding to the risk level based on the risk level and the dynamic adjustment results. By incorporating key parameters such as voice preemption priority, voice broadcast frequency limit, and voice content information compression intensity into the adjustment, this invention can implement high-priority preemption broadcasts when risk rises rapidly or changes abruptly, ensuring that critical alarms reach cyclists within the optimal intervention window. Simultaneously, frequency limit control avoids repeated broadcasts in complex traffic flows that could cause distraction. Furthermore, this invention dynamically compresses voice content based on the effective width of the traffic channel and cognitive load: the narrower the channel and the higher the task load, the shorter the voice, the more direct the wording, and the removal of secondary explanations, thus ensuring clear and understandable information and reducing information overload even under high-risk and high-load conditions. This strategy enables voice alerts to be timely, controlled in frequency, and adaptively compressed, significantly increasing the probability that key alerts are heard, understood, and executed. Furthermore, it can be linked with risk levels and triggering conditions to facilitate subsequent evaluation and optimization. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall process of the AI ​​voice warning method for two-wheeled vehicle riding safety of the present invention; Figure 2 This is a schematic diagram of the framework for generating scene perception results according to the present invention; Figure 3 This is a schematic diagram illustrating the identification of escape channel collapse state and the construction of occlusion edge risk field according to the present invention. Figure 4 This is a schematic diagram illustrating the strategy for generating risk scores, risk levels, and voice warning outputs in this invention. Detailed Implementation

[0019] Example 1: Refer to Figures 1 to 4 AI-powered voice warning methods for two-wheeled vehicle riding safety include: Step 1: Acquire multi-source data related to risk warnings during two-wheeled vehicle riding and add timestamps to the multi-source data. The significance of this step is to provide continuous, aligned, and traceable basic input for subsequent risk identification and voice warnings. By acquiring multi-source data related to risk warnings, the method can simultaneously reflect changes in the two-wheeled vehicle's own movement and changes in the surrounding environment. By adding timestamps to the multi-source data, data from different sources and with different sampling frequencies can be synchronously correlated on a unified timeline, avoiding scene misjudgment and risk estimation deviations caused by inconsistent collection timing. This provides a consistent data timing basis for data preprocessing, scene perception, and subsequent indicator calculations. In one specific implementation, step 1 can be implemented as follows, used to generate a highly consistent multi-source data stream that can be used for risk warning and to complete time calibration during two-wheeled vehicle riding: Inertial measurement data is collected to characterize the motion state of the two-wheeled vehicle. The inertial measurement data includes at least angular velocity, acceleration, and attitude change information derived from them. The sampling frequency can be set higher than that of the environmental perception data to ensure the ability to characterize short-term dynamics such as rapid acceleration, rapid deceleration, and rapid steering. Collect environmental perception data to characterize the state of the surrounding environment. The environmental perception data includes at least the perception results of road structure, traffic participants such as vehicles and pedestrians, static obstacles, and occlusion edges. The environmental perception data can be derived from a combination of forward or surround view image data and distance measurement data to enhance the ability to express occlusion areas and passage spaces. For example, when oncoming vehicles meet and parked vehicles on the side of the road form occlusion, the environmental perception data can reflect the position of the occlusion edge and its changes over time. Timestamps are added to both inertial measurement data and environmental perception data. The timestamps are generated using a unified clock reference and record the sampling time, data arrival time, and buffer queuing delay information to distinguish between sampling delay and transmission delay in subsequent processing. The multi-source data after adding timestamps is encapsulated in a queue and output sequentially. Data frames are sorted according to timestamps and cross-source index relationships are established, so that inertial measurement data and environmental perception data within the same time window can be read together. For example, when a two-wheeled vehicle passes through a narrow passage area at about 20 kilometers per hour and there is a large vehicle blocking the right side, the inertial measurement data reflects the vehicle's attitude and yaw changes, while the environmental perception data reflects the state of the blocking edge and the passage space. Both are recorded synchronously under the same timestamp, thus providing a time-consistent, traceable and complementary input basis for subsequent risk warning related calculations. Step 2: Preprocess the multi-source data and output preprocessed data. The significance of this step is to transform the raw multi-source data obtained in Step 1 into a stable input suitable for algorithm processing, reducing the impact of noise, drift, missing data, and latency on subsequent scene perception and risk assessment. By preprocessing the multi-source data, data quality can be improved while maintaining information integrity. The preprocessed data meets the input requirements for subsequent scene element identification and state estimation in terms of time alignment, scale consistency, coordinate consistency, and validity, thereby improving the reliability and consistency of the overall early warning chain. In one specific implementation, step 2 can be implemented as follows to transform the multi-source data obtained in step 1 into preprocessed data that can be used for subsequent scene element identification and state estimation: Multi-source data is synchronized in time. A unified time axis is established based on the timestamps of each data frame. A sliding time window alignment strategy is used to interpolate and resample the inertial measurement data within the time window. The environmental perception data is selected in the same time window or time-compensated, thereby forming a synchronized data pair that can be correlated within the same time window. Multi-source data is filtered and denoised. For inertial measurement data, bandpass-constrained adaptive filtering is used to suppress high-frequency vibration and low-frequency drift. For distance measurement and target detection results in environmental perception data, time-series consistency constraints are used for smoothing to reduce measurement jitter caused by road bumps, sudden changes in lighting, or instantaneous occlusion. For example, on cobblestone roads or speed bump sections, the high-frequency impact component of inertial measurement data is suppressed, and the target boundary jitter of environmental perception data is smoothed. Coordinate system calibration is performed on multi-source data to establish the external parameter relationship between the two-wheeled vehicle motion coordinate system and the coordinate systems of each sensor. The filtered and denoised inertial measurement data and environmental perception data are uniformly transformed into the two-wheeled vehicle motion coordinate system. At the same time, attitude compensation is performed on the environmental perception data in combination with vehicle attitude changes, so that the spatial position, velocity direction and the representation of the passage space are comparable under the same coordinate frame. For example, when turning into a narrow passage, the passage boundary in the environmental perception data remains continuous and stable under the two-wheeled vehicle motion coordinate system. Data validity is checked for multi-source data. The validity of data is determined based on timestamp continuity, numerical range constraints, rate of change constraints, and cross-source consistency constraints. Abnormal data is marked, removed, or output with reduced weight. Short-term prediction is used to complete missing segments to maintain the continuity of preprocessed data. For example, when environmental perception data fails for a short time due to strong backlight, the attitude and velocity changes of inertial measurement data are used to maintain the continuity of the state within the time window. Preprocessed data with validity marks is output to provide a stable, traceable, and interpretable input basis for the subsequent estimation of dynamic occlusion state, lateral clamping constraint state, and passageway state.

[0020] Step 3: Based on the preprocessed data, perform scene element identification and state estimation to obtain scene perception results including at least the dynamic occlusion state, lateral clamping constraint state, and passageway state. The significance of this step lies in elevating information at the "data" level to a structured state expression at the "scene" level, providing interpretable and combinable scene elements for subsequent calculations of passageway collapse indices, occlusion potential threat intensity indices, and risk scores. Through scene element identification and state estimation, on the one hand, a description of the dynamic occlusion state is formed, reflecting the degree of occlusion of the observability of potential targets and changes in the occlusion edge morphology; on the other hand, a lateral clamping constraint state is formed, reflecting the available space margin on the left and right sides of the two-wheeled vehicle and its persistence; simultaneously, a passageway state is formed, reflecting the effective width, length, and exit availability of the two-wheeled vehicle's available escape route. These scene perception results enable the method to continuously update the key state quantities upon which risks depend in the dynamic environment of two-wheeled vehicle riding. In one specific implementation, step 3 can be implemented as follows to generate scene perception results that can be used for risk warning based on the preprocessed data: Based on preprocessed data, scene element recognition is performed. The road boundaries, lane lines, curbs, guardrails, parked vehicles, pedestrians and other traffic participants in the environmental perception data are segmented and targets are detected. The element set that can be used for geometric calculation is output in the two-wheeled vehicle motion coordinate system. At the same time, the attitude change calculated by combining inertial measurement data is used to perform temporal stabilization processing on the element set to ensure that the change of element position between consecutive frames conforms to the motion continuity. Based on the identified occlusion-related elements, dynamic occlusion state estimation is performed. The occlusion contour is extracted from the element set and the geometric range of the occlusion edge is determined. Furthermore, by combining the relative motion information of the occlusion body and the probability of the occurrence of potential targets within the occlusion area, a dynamic occlusion state is obtained to characterize the degree of occlusion and the change in the shape of the occlusion edge, which is used to characterize the observability of potential targets. For example, when a large vehicle on the right side obstructs the view of the intersection and the vehicle moves slowly forward, the dynamic occlusion state can simultaneously reflect the increase in the degree of occlusion and the extrapolation of the occlusion edge. Lateral clamping constraint state estimation is performed based on the identified traffic space boundary elements. The available space margin on the left and right sides of the two-wheeled vehicle is calculated and its persistence is evaluated in a time series. The available space margin is determined by the minimum safe gap between the outline of the two-wheeled vehicle and the left and right side boundary elements. The persistence is determined by the duration and trend of the margin being lower than the preset safety margin within a continuous time window. For example, when the left isolation guardrail and the parked vehicle on the right side form a clamping, the lateral clamping constraint state can output the constraint characteristics of the left and right side spaces tightening at the same time and the duration increasing. Based on the lateral clamping constraint state and road structure elements, the passageway state is estimated, and a drivable area corridor along the current speed direction of the two-wheeled vehicle is constructed. The effective width, effective length, and exit availability of the passageway are calculated. The exit availability is jointly determined by the connectivity of the drivable area ahead of the corridor and the obstruction status at the exit. For example, in the case of oncoming traffic on a narrow road and the exit ahead being occupied by an oncoming vehicle, the passageway state can provide the conclusion that the effective length is shortened and the exit availability is reduced. This results in a scene perception result that includes at least the dynamic occlusion state, the lateral clamping constraint state, and the passageway state, providing structured input for subsequent calculations of passageway collapse index, occlusion potential threat intensity index, and risk score.

[0021] The setting of minimum safety clearance and preset safety margin can be based on a comprehensive consideration of factors such as the overall dimensions of the two-wheeled vehicle, the lateral sway of the riding posture, the increase in lateral space caused by tire slippage and suspension compression, the trajectory uncertainty during braking or steering, changes in road adhesion coefficient, and sensor errors and coordinate system calibration errors. The minimum safety clearance can serve as a hard lower limit under instantaneous geometric constraints, while the preset safety margin can serve as a comfort and controllable redundancy within a time window. For example, taking a common electric two-wheeled vehicle as an example, the widest part of the handlebars is approximately 0.65 meters. Considering the additional 0.05 to 0.10 meters of lateral sway of the rider and vehicle body during riding, as well as slight serpentine corrections, and the lateral uncertainty of 0.05 meters due to sensor and calibration errors, and considering the further increase in lateral trajectory uncertainty on slippery surfaces or during sudden braking, the minimum safety clearance on one side can be set to approximately 0.30 meters to ensure no collision occurs. Simultaneously, the preset safety margin can be set to be 0.40 to 0.50 meters greater than the minimum safety clearance. This system is used to assess whether the available space on the left and right sides is under prolonged congestion within a continuous time window. Based on this, a set of examples can be provided to understand the three types of states: When a two-wheeled vehicle is traveling along a parked truck on the right side and the truck obstructs the view of the intersection, the dynamic obstruction state is used to characterize the gradual increase in the degree of obstruction, the forward and outward expansion of the geometric range of the obstruction edge, and the changes in relative motion information, thereby reflecting the decrease in the observability of potential targets and the changes in the shape of the obstruction edge; When there is a guardrail on the left, a parked vehicle on the right, and the distance between the two sides is narrowing, the lateral clamping constraint state is used to characterize the available space margin on the left and right sides of the two-wheeled vehicle being close to or lower than the preset safety margin and persisting within a few seconds, reflecting the strength and persistence of the clamping constraint; At the same time, when there is an oncoming vehicle or an obstacle occupying the exit ahead, the passageway state is used to characterize the effective width of the two-wheeled vehicle's available escapeway shrinking with clamping, the effective length shortening due to the obstruction ahead, and the reduced availability of the exit, thereby providing an intuitive and quantifiable scenario basis for subsequent risk assessment and voice warning; For example, on a narrow two-way road, a two-wheeled vehicle is traveling at approximately 20 km / h along the right side. A truck is parked on the right, and its front end is slowly moving forward, obstructing the exit of a side road ahead. There is a continuous guardrail on the left. Based on the two-wheeled vehicle's external dimensions and lateral sway during riding, a minimum safety clearance of 0.30 meters and a preset safety margin of 0.45 meters are set for each side. After synchronizing environmental perception data and inertial measurement data, the following conclusions can be drawn: First, the dynamic obstruction status shows that the obstruction caused by the truck gradually increases, with the geometric range of the obstruction edge pushing forward and outward, accompanied by relative motion changes, leading to a decrease in the observability of potential targets such as electric bicycles or pedestrians that may exit the side road. Second, the lateral clamping constraint state... The display shows that the left guardrail and the right truck are jointly squeezing the available space. The available space margin on the right side of the two-wheeled vehicle drops to about 0.32 meters and approaches the minimum safety gap multiple times. The available space margin on the left side is about 0.38 meters and is lower than the preset safety margin for several consecutive seconds, indicating that the clamping constraint is not only strong but also continuous. Thirdly, the passage status shows that the effective width of the available escape passage along the current speed direction approaches the threshold as the clamping shrinks. The effective length is shortened due to the obstruction area in front and the occupation by oncoming vehicles, and the availability of the exit is reduced. This is manifested as poor connectivity in front and insufficient detour space, thus providing an intuitive and quantifiable scenario input for the subsequent calculation of the passage collapse degree index, the obstruction potential threat intensity index, and the risk score.

[0022] Step 4: Calculate the channel collapse degree index and the occlusion potential threat intensity index based on the scene perception results, and identify whether the escapeable channel collapse state has been entered based on these indices. When the escapeable channel collapse state is identified, an occlusion edge risk field is established, multiple virtual targets are generated, and a probability mass is assigned to each virtual target to obtain the virtual target risk quantity corresponding to each virtual target. The significance of this step is that in complex cycling scenarios, not only is the direct risk of "visible targets" considered, but also the coupled risks caused by "escapeable channel collapse" and "potential targets after occlusion" can be modeled and identified in advance. First, the channel collapse degree index is calculated based on the scene perception results to quantify the degree of contraction of the available escape channel as reflected in the passage channel state and the lateral clamping constraint state; at the same time, the occlusion potential threat intensity index is calculated to quantify the uncertainty of the appearance of potential targets in the occlusion area reflected in the dynamic occlusion state and its potential risk contribution. Then, based on the channel collapse degree index and the occlusion potential threat intensity index, it is identified whether the escapeable channel has entered a collapsed state. This allows the method to trigger more targeted risk modeling under the combined conditions of "decreased channel escapeability and increased occlusion risk." When the escapeable channel collapse state is identified, an occlusion edge risk field is established to express the risk of the unobservable area near the occlusion edge in the form of a field. This allows the risk to be continuously distributed in space and reflects the coupled influence of the occlusion edge, the direction of travel, and the passageway. By generating multiple virtual targets, potential targets that cannot be directly observed but may appear from the occlusion area are represented as hypothetical targets, facilitating their unified inclusion in subsequent risk assessments. By assigning probability mass to each virtual target, different virtual targets have distinguishable "probability of occurrence" at the uncertainty level. Finally, the virtual target risk quantity corresponding to each virtual target is obtained, allowing "potential target risk" to be quantitatively included in subsequent risk score calculations, thereby improving the foresight of early warnings in the case of sudden occlusion and channel restriction.

[0023] To facilitate implementation by those skilled in the art, an example calculation method for the channel collapse index is provided, along with a set of optional dimensionless mapping and coupled fusion calculation methods. Let the effective width of the passageway within the same time window be... (m), the effective length of the passageway is (m), the availability of the channel exit is ( (Indicates fully usable), lateral clamping constraint degree is (A larger value indicates stronger clamping). Let the width threshold corresponding to the preset safety margin be... The minimum width corresponding to the minimum safety clearance is The length safety threshold is the greater of the braking safety distance and the steering reachable distance. The minimum length is .in, Can be determined by the current speed With minimum achievable deceleration boundary Estimate (e.g.) ), It can be estimated from the achievable lateral displacement boundary and the available handling margin; the above parameters can be set by those skilled in the art in combination with vehicle type and road conditions.

[0024] (1) Width urgency: Define width urgency ( It indicates that it is not urgent. (indicating extreme urgency) ;in Indicates will Limit to Interval.

[0025] (2) Length urgency: Define length urgency for ; (3) Exit blocking degree: Define the exit blocking degree for ; (4) Clamping constraint: Let the available space margins on the left and right sides be respectively (m), time window is Within this time window, the margins on both sides are lower than the preset safety margin. The duration is (s), the minimum safe clearance on one side is (m), then we can take ;in This is a weighting coefficient used to balance the contributions of "clamping continuity" and "approaching the minimum safe gap".

[0026] (5) Coupling and integration: To reflect "exit blockage" To address the nonlinear amplification effect when "clamping" persists, a multiplicative penalty term is introduced, defining a channel collapse index. for ; in , When it is necessary to increase sensitivity to "narrow passage + occupied exit", the sensitivity can be appropriately increased. and .

[0027] For example, in a narrow oncoming traffic scenario on a city branch road, a two-wheeled vehicle is traveling at approximately 20 km / h along the right side. There is a continuous guardrail on the left and a parked truck on the right. An oncoming vehicle appears about 12 meters ahead, occupying part of the passageway. Simultaneously, the truck's front end is slowly moving forward, reducing connectivity at the exit. Within the same time window, the effective width of the passageway is taken as the minimum usable width of the transverse cross-section of the passageway, which is 0.78 meters. The effective length of the passageway is taken as the drivable distance along the current speed direction, which is 9.5 meters. The availability of the passageway exit is zero. The continuous quantity is represented as 0.35. In the lateral clamping constraint state, the available space margin on the left is 0.37 meters, the available space margin on the right is 0.31 meters, and the duration of both margins being lower than the preset safety margin of 0.45 meters has reached 2.0 seconds and continues to increase. Moreover, the margin on the right repeatedly approaches the minimum safety clearance of 0.30 meters. Subsequently, when performing dimensionless normalization, the segment boundary of 0.78 meters relative to 0.45 meters and 0.30 meters is mapped to a higher width urgency, and 9.5 meters is relative to the braking safety distance. The larger of the turning reachable distances, such as 15 meters, is mapped to a higher length urgency. The channel exit availability of 0.35 is directly mapped to a higher exit blockage. The clamping performance lasting 2.0 seconds and approaching the minimum safety gap is mapped to a higher clamping constraint. In the coupling and fusion stage, the width urgency, length urgency, exit blockage, and clamping constraint are weighted and superimposed according to risk correlation. At the same time, since the exit blockage is high and the clamping constraint persists within the time window, a multiplicative penalty term is introduced to nonlinearly amplify the fusion result, resulting in a significant increase in the channel collapse index, which continues to rise in subsequent time windows. In the stabilization stage, although the effective width of the passageway briefly rebounds to 0.82 meters due to detection jitter in some frames, the rising edge-priority hysteresis smoothing strategy preserves the trend of intensified collapse. The channel collapse index will not frequently jump, but will remain at a high level in a continuous and traceable manner and remain sensitive to rapid deterioration, thus intuitively reflecting that the available escape channel for two-wheeled vehicles is rapidly shrinking in this oncoming clamping scenario.

[0028] Example calculation method for the potential threat intensity index of obstruction: To facilitate implementation by those skilled in the art, a set of optional methods for calculating the potential threat intensity index of occlusion is provided. Let the degree of occlusion within the same time window be denoted as... The minimum projected distance from the critical segment of the occluded edge to the outer contour of the two-wheeled vehicle is (m); The intersection angle between the key segment and the predicted trajectory of the two-wheeled vehicle is (rad); the occlusion edge intrusion degree is Intrusion rate is The probability of a potential target appearing within the obscured area is: Among them, the degree of invasiveness Used to characterize the strength of the intrusion tendency of the occlusion edge towards the passageway / predicted trajectory direction, intrusion rate Used to characterize the rate of change of this intrusion trend; Information can be obtained from intersection type, traffic flow density clues, historical statistics, or V2X / map priors.

[0029] (1) Observability loss: Define observability loss for ; (2) Geometric urgency: Define geometric urgency (The closer the distance and the stronger the intersection, the greater the value) ; in For reference distance (e.g.) ), It is the minimum distance lower limit (e.g., on the same order of magnitude as the minimum safety clearance on one side). upper limit angle (e.g.) ), These are the weighting coefficients.

[0030] (3) Conditional Amplification Fusion: When "geometric urgency is high and there is an intrusion trend", the probability term is amplified to reflect the increased risk of the target suddenly appearing in the occlusion area and directly entering the predicted trajectory. Define the conditional amplification coefficient. for ;in This is the amplification factor. Further definition of the potential threat intensity index caused by obstruction. for ;in , When the occlusion edge shows a continuous intrusive trend (e.g. When the price rises monotonically over multiple consecutive time windows (or exceeds a preset threshold and remains there), it can improve... or improve To enhance sensitivity to "emergence probability"; when the occlusion is mainly due to static occlusion but the edge intrusion tendency is weak, the sensitivity can be appropriately increased. To enhance the portrayal of "lack of field of vision + geometric urgency".

[0031] In one specific implementation, the channel collapse degree index and the occlusion potential threat intensity index are calculated based on the scene perception results, and the status of whether an escapeable channel collapse state has been entered is identified accordingly. This can be performed by the following steps: The scene perception results within the same time window are acquired and a set of basic quantities for index calculation is formed. The set of basic quantities includes at least the effective width of the passage, the effective length of the passage, the availability of the passage exit, the lateral clamping constraint state, and the dynamic occlusion state. The lateral clamping constraint state is used to characterize the available space margin on the left and right sides and its persistence. The dynamic occlusion state includes at least the degree of occlusion, the geometric range of the occlusion edge, the relative motion information of the occlusion edge, and the probability of potential targets appearing within the occlusion area. Dimensionless normalization and element extraction are performed on the basic quantity set. The effective width of the passage, the effective length of the passage, the availability of the passage exit, and the lateral clamping constraint state are mapped to width urgency, length urgency, exit blockage, and clamping constraint, respectively. The passage collapse index is calculated according to the coupling fusion rule. At the same time, the degree of occlusion is mapped to the observability loss. The edge segments intersecting with the predicted trajectory of the two-wheeled vehicle in the geometric range of the occlusion edge are extracted as key edge segments, and the minimum projection distance and intersection angle are calculated to form the geometric urgency. The relative motion information of the occlusion edge is mapped to the edge intrusion degree and intrusion rate, and fused with the probability of potential target occurrence according to the condition amplification rule to obtain the occlusion potential threat intensity index. Based on the channel collapse degree index and the occlusion potential threat intensity index, an identification criterion for the escapeable channel collapse state is constructed. The time window consistency constraint method is used to determine whether the channel collapse degree index exceeds the channel collapse threshold and remains at least for a preset duration, and whether the occlusion potential threat intensity index exceeds the occlusion threat threshold and shows an upward trend. When both are satisfied within the same time window, the escapeable channel collapse state is determined. At the same time, complementary triggering conditions are introduced to cover extreme scenarios. That is, when the channel collapse degree index is significantly higher than the channel collapse threshold, even if the occlusion potential threat intensity index is at a medium level, the escapeable channel collapse state can be determined. Or, when the occlusion potential threat intensity index is significantly higher than the occlusion threat threshold and the occlusion edge intrusion continues to increase, even if the channel collapse degree index is at a critical level, the escapeable channel collapse state can be determined. This improves the sensitivity to sudden collapse. The system outputs the recognition results and performs state hysteresis and de-jitter processing. Different thresholds or durations are set for entry and exit judgments to form hysteresis intervals, avoiding frequent state switching caused by short-term measurement jitter. For example, in a scenario where a two-wheeled vehicle passes through a narrow road at approximately 20 km / h, the left guardrail and the truck parked on the right reduce the effective width of the passage to 0.75 meters, the effective length of the passage is shortened to 9 meters due to oncoming traffic, and the availability of the passage exit drops to 0.30. This causes the passage collapse index to exceed the passage collapse threshold within two consecutive time windows. At the same time, the truck blocks the exit of the side road and moves forward slowly, causing the occlusion degree to rise to 0.70. The overlap between the key edge segment and the predicted trajectory is enhanced, and an edge intrusion trend appears. The probability of potential target appearance is set to 0.55, causing the occlusion potential threat intensity index to quickly exceed the occlusion threat threshold and continue to rise. In the third step, it can be determined that the vehicle has entered an escapeable passage collapse state. Hysteresis and de-jitter processing ensure that this state maintains a stable output until the risk is truly eliminated, thus providing a continuous and traceable basis for subsequent safety strategy selection.

[0032] Example calculation method for occlusion edge risk field and virtual target risk quantity: To facilitate implementation by those skilled in the art, a set of optional methods for constructing the occlusion edge risk field, generating virtual targets, and calculating virtual target risk quantities are presented in this section. Denotes the Euclidean norm. Indicates will Limit to interval, This indicates that the input quantity is linearly mapped to a preset upper and lower bound. The range (upper and lower boundaries can be set by historical statistics or engineering experience).

[0033] (1) Edge sampling points and risk sources: The occlusion edge of the dynamic occlusion body is discretized into a set of edge sampling points in the two-wheeled vehicle motion coordinate system. ,in or This represents the coordinates of the sampling point. A local normal direction is assigned to each sampling point. (pointing to the unobservable area), and the normal direction relative to the speed direction of the two-wheeled vehicle. The included angle (Or calculated from the relationship between the velocity direction and the point vector direction), used to characterize the anisotropy of risk propagation along the direction of travel.

[0034] (2) Anisotropic distance decay risk potential function (single source): for any point in the unobservable region sampling points As a single source of risk, define an anisotropic distance decay risk potential function. for ;in Indicates from point to Direction and speed direction of the two-wheeled vehicle The included angle, For attenuation scale, Used to control stronger anisotropy in the forward direction (e.g.) The larger the value, the stronger the potential energy along the velocity direction.

[0035] When the effective width of the passageway Smaller, export availability Lower or obscured edge intrusion When the potential function gain is high, it can be amplified by making... ;in Based on the gain, This is the gain coefficient; This is the channel collapse index (which can be calculated from the previous section). For export availability, This represents the degree of edge intrusion.

[0036] (3) Risk field superposition: The risk field at the occlusion edge is obtained by superimposing the potential functions of each risk source. : ; (4) Candidate point / virtual target generation: Under the condition of satisfying the risk field Candidate point search is performed within the region, where The preset field strength threshold is used. Candidate points can be obtained in any of the following ways: First, according to the preset spatial resolution. First, perform a grid scan and select local peak points where the field strength exceeds a threshold; second, use local extremum search to obtain a set of candidate points. To avoid duplicate assumptions due to overly dense candidate points, a minimum spacing constraint is applied. ,in Set the minimum spacing. For each candidate point... Mapped to a virtual target The assumed location.

[0037] (5) Probability mass assignment (softmax example): Let the field strength at the candidate point be... The lateral offset of the candidate point from the centerline of the passageway is The escape route margin at the candidate point is (The smaller the value, the more dangerous it is). Define candidate point scores. for ,in These are the weighting coefficients. The virtual target is obtained using softmax. probability mass : Thus satisfying In an alternative implementation, it is also possible to... Set minimum quality limit Then normalize it to avoid numerical instability caused by extremely low quality when there are a large number of candidate points.

[0038] (6) Virtual target risk quantity (coupled with "lower bound of shortest arrival time"): Let the speed of the two-wheeled vehicle be... The shortest distance from the candidate point to the outer contour of the two-wheeled vehicle is Then define the lower bound of the shortest arrival time. ;in This is a very small quantity, used to avoid division by zero. Define the virtual target risk quantity. for Thus, the "probability of occurrence (probability quality)," "local risk field strength," and "time urgency (lower bound of arrival time)" jointly determine the virtual target risk quantity.

[0039] (7) Enhanced tail risk: To improve sensitivity to "low probability, high consequence" events, when there is virtual targets (where When setting a high-risk threshold, the probability quality or risk weight can be non-linearly enhanced. For example, the probability quality can be power-transformed and then normalized. ; and in the formula Replace with To obtain the enhanced ; or equivalent to A multiplicative enhancement factor is applied and amplitude is limited. The above tail risk enhancement is an optional implementation and does not restrict other equivalent nonlinear enhancement methods.

[0040] Step 5: Calculate the risk score based on the scene perception results, virtual target risk quantity, and insufficient action accessibility index, and output the risk level according to the risk score; determine the dynamic adjustment gear parameter set according to the risk level, and dynamically adjust the risk triggering mechanism, voice broadcast strategy, and early warning suggestion output strategy under the constraints of the dynamic adjustment gear parameter set; the significance of this step is to unify and integrate "scene status", "potential target risk" and "current executable capability of two-wheeled vehicle" to form a single decision quantity for voice early warning, and enable the early warning behavior to dynamically adapt to the risk level. By calculating risk scores based on scene perception results and virtual target risk quantities, the risk score simultaneously reflects the basic risks corresponding to the passageway status and dynamic obstruction status, as well as the potential obstruction risks represented by the virtual target risk quantity. By introducing an action accessibility indicative indicator, the risk score reflects not only "external danger" but also the degree of limitation of available avoidance actions in the current state, thus avoiding providing mismatched warning strength or suggestions when actions are difficult to achieve. Based on the risk score, a risk level is output, transforming voice warnings from continuous quantities into configurable and tiered triggering criteria, facilitating different intensities, information contents, and preemption strategies for voice output at different risk levels. A dynamic adjustment level parameter set is determined based on the risk level, enabling the system to switch adjustment levels according to the risk level. Under the constraints of this parameter set, the risk triggering mechanism, voice broadcasting strategy, and warning suggestion output strategy are dynamically adjusted, allowing warning triggering conditions, broadcasting frequency, and content organization to adaptively adjust with changes in risk level. This ensures timeliness while reducing unnecessary interference and false triggers, improving usability and safety during cycling.

[0041] In one specific implementation, the risk score is calculated based on scene perception results, virtual target risk quantity, and action accessibility insufficiency index, and is used to update the output risk level over time. This can be performed as follows: The scene perception results within the same time window are acquired and the basic risk quantity is calculated. The basic risk quantity includes at least the channel risk quantity obtained from the passage channel state and the occlusion risk quantity obtained from the dynamic occlusion state. The channel risk quantity is formed by dimensionless mapping of the effective width of the passage channel, the effective length of the passage channel, the availability of the channel exit, and the lateral clamping constraint state, and is used to characterize the degree of contraction of the available escape space. The occlusion risk quantity is formed by conditional amplification and fusion of the occlusion degree, the geometric urgency of the occlusion edge, the occlusion edge intrusion degree, the intrusion rate, and the probability of the occurrence of potential targets in the occlusion area, and is used to characterize the degree of increase in the suddenness risk caused by occlusion. The risk quantities of multiple virtual targets are obtained and probabilistically aggregated. For each virtual target, its probability quality and virtual target risk quantity are read. Aggregation is performed using either probability quality-weighted summation or probability quality-weighted expected risk form to obtain the aggregated risk quantity of the virtual target. At the same time, a tail risk enhancement strategy can be introduced to improve the sensitivity to high-risk, low-probability virtual targets. That is, when the risk quantity of a virtual target exceeds the preset high-risk threshold, its weight is non-linearly increased and then normalized, so that the aggregation result takes into account both the overall risk and the extreme emergence risk. The accessibility insufficiency index is obtained and normalized to obtain the accessibility penalty amount. The accessibility insufficiency index is normalized relative to the current speed of the two-wheeled vehicle, road adhesion conditions, available handling margin and available braking margin, and mapped to the accessibility penalty amount of a continuous quantity from zero to one. When the available handling margin and available braking margin are simultaneously limited or the accessibility insufficiency is continuously increasing, a rising edge priority hysteresis smoothing is introduced to make the penalty amount sensitive to rapid deterioration and insensitive to instantaneous jitter. The risk score is obtained by weighted fusion of the basic risk quantity, the virtual target aggregated risk quantity, and the accessibility penalty quantity, followed by amplitude limiting and normalization. The risk level is then output based on the risk score. The weighted fusion includes at least a linear weighting term and a coupling amplification term. When the channel risk quantity is high and the accessibility penalty quantity is high, a multiplicative amplification is applied to the virtual target aggregated risk quantity to reflect the amplified risk of the consequences of sudden targets under conditions of insufficient escape space. Subsequently, the fusion result is amplitude limited and mapped to a risk score range with a unified dimension. At the same time, a consistency constraint is applied to the risk score within the time window to avoid frequent level jumps.

[0042] For example, in a scenario where a two-wheeled vehicle is merging at a city side road exit and the lane narrows, it is traveling at approximately 21 km / h along the right side of the main road. A large van is temporarily parked ahead on the right, blocking the side road exit, while the guardrail on the left restricts lateral space. Within the same time window, the scene perception results indicate an effective passage width of approximately 0.82 meters, an effective passage length of approximately 11 meters, an exit availability of approximately 0.38, and a continuous lateral clamping constraint. Based on the dimensionless mapping and coupling fusion rules, the passage risk is calculated to be 0.72. The dynamic occlusion state indicates an occlusion degree of 0.68, with the key segment of the occlusion edge showing an overlap trend with the predicted trajectory of the two-wheeled vehicle. The key segment extends beyond the two-wheeled vehicle's path. The minimum projection distance of the contour is approximately 0.65 meters, and the intersection angle increases. The relative motion of the occlusion edge shows that the edge intrudes into the passageway direction at approximately 0.18 meters per second, and this intensifies. The probability of a potential target appearing within the occlusion area is taken as 0.52. After conditional amplification and fusion, the occlusion risk is taken as 0.66. Therefore, the basic risk can be taken as a weighted combination of the channel risk and the occlusion risk, for example, 0.69. Furthermore, the occlusion edge risk field generates three virtual targets V1, V2, and V3, with probability masses of 0.55, 0.30, and 0.15, respectively, corresponding to virtual target risk masses of 0.90, 0.60, and 0.40, respectively. Then, the virtual targets are obtained by weighted aggregation based on probability mass. The aggregated risk is 0.55×0.90+0.30×0.60+0.15×0.40=0.735. Since the V1 risk exceeds the preset high-risk threshold of 0.85, the V1 weight can be nonlinearly boosted and then normalized to slightly increase the aggregated result, for example, to 0.76, to enhance the sensitivity to high-risk, low-probability emergent events. Meanwhile, the action accessibility indices are calculated from the current speed, available maneuverability margin, and available braking margin, resulting in a value of 0.78. Mapping this to the preset normalized boundary yields an accessibility penalty of 0.78. Due to the persistent clamping constraint and the significant rising edge of the penalty, a rising edge-priority hysteresis smoothing method is used to keep the penalty at a high level within the continuous time window. Finally, the base risk (0.69), the virtual target aggregated risk (0.76), and the accessibility penalty (0.78) are weighted and fused. For example, a linear weighting is first performed to obtain 0.40×0.69+0.35×0.76+0.25×0.78=0.737. Then, because the channel risk (0.72) is high and the accessibility penalty (0.78) is relatively high, a multiplicative amplification term is applied to the virtual target aggregated risk, for example, (1+0.30×0.72×0.78)≈1.168, which raises the fusion result to approximately 0.80. The result is then normalized to obtain a risk score of 0.80. The risk score is updated over time. If it remains above 0 for several consecutive time windows...If the fluctuation satisfies the consistency constraint, a high-risk level is output. If the effective width of the passageway subsequently increases, the intrusion trend of the obstruction edge weakens, and the penalty for action accessibility decreases, the risk score smoothly decreases under the hysteresis rule, and a corresponding lower risk level is output. This achieves continuous and traceable classification of the coupled risks of passageway contraction, obstruction emergence, and difficulty in achieving actions.

[0043] In one specific implementation, a set of dynamic adjustment parameters is determined based on the risk level, and the risk triggering mechanism, voice broadcast strategy, and early warning suggestion output strategy are dynamically adjusted under the constraints of the dynamic adjustment parameter set. This can be performed as follows: A mapping relationship is established between risk levels and a set of dynamic control parameters, and threshold adaptive setting is completed. At least low-risk, medium-risk, and high-risk levels are determined based on the segmented intervals of risk scores, and different sets of dynamic control parameters are configured for different risk levels. These sets of dynamic control parameters include at least a first threshold, a second threshold, a third threshold, and a fourth threshold, as well as corresponding gain coefficients, gating strength, upper limit of voice broadcast frequency, and information compression strength. The first threshold is set based on the coupling critical point between the channel collapse index and the two-wheeled vehicle outline safety clearance, escape channel margin, and lower limit of usable braking distance, so that the first threshold corresponds to the state where the passageway transitions from usable to a state where it is difficult to rely on the collision time of a visible target for triggering. The second threshold is set based on the insufficient mobility index and the usable maneuvering margin, usable braking margin, and achievable lateral displacement and achievable deceleration boundaries under road adhesion conditions, so that the second threshold corresponds to the state where the passageway transitions from usable to a state where it is difficult to rely on the collision time of a visible target for triggering. The first threshold is the critical point at which highly directional avoidance actions are difficult to achieve stably within a preset reaction time. The second threshold is set based on the risk change rate index and the risk growth rate perceptible by human-computer interaction, the voice prompt arrival delay, and the driver's reaction time. This makes the third threshold correspond to the critical slope where the risk rises rapidly within a short time window and requires preemptive broadcasting for early intervention. The third threshold is set based on the risk jump amplitude index and the amplitude threshold point at which the risk score jumps across levels due to sudden occupancy, a sharp drop in the availability of the channel exit, or a sudden change in opposite occupancy. This makes the fourth threshold correspond to the critical amplitude at which the risk undergoes a structural change within a single time window and must be broadcast immediately. Furthermore, all four thresholds are adaptively adjusted according to the risk level. That is, the higher the risk level, the more appropriately the first and second thresholds are lowered to improve sensitivity, and the more appropriately the third and fourth thresholds are lowered to accelerate preemptive triggering. At the same time, the weight gain of the virtual target risk quantity is increased, the upper limit of the gating strength is increased, and the redundancy of voice information is reduced. Under the constraint of the first threshold, the risk triggering mechanism is switched and the gain is adjusted. When the channel collapse index is greater than the first threshold, the risk triggering mechanism is switched from collision time threshold triggering based on visible targets to threshold triggering based on passage channels. The weight gain of virtual target risk quantity and the first threshold are adaptively adjusted according to the risk level. The weight gain of virtual target risk quantity is used to increase the contribution ratio of virtual target aggregate risk to overall risk when the uncertainty of occlusion is significant. This makes the triggering criterion shift from relying on the collision time of visible targets to relying on the urgency of passage channels and potential occlusion threats, thereby avoiding missed triggering when visible targets are missing or severely occluded. Under the constraint of the second threshold, the output strategy of the warning suggestion is accessibility gated. The warning suggestion is accessibility gated according to the action accessibility deficiency index. When the action accessibility deficiency index is greater than the second threshold, the output of strong directional avoidance suggestions is prohibited and conservative deceleration and pre-braking suggestions are output. The second threshold and the gate intensity are adaptively adjusted according to the risk level. The gate intensity is used to control the continuous transition from complete prohibition to partial weakening. For example, in the medium risk level, the directional suggestion can be changed from mandatory pointing to optional prompt. In the high risk level, only conservative suggestions such as deceleration, pre-braking, and maintaining vehicle stability are retained to reduce the probability of guiding wrong actions when the maneuvering space is limited. Under the constraints of the third and fourth thresholds, the voice broadcasting strategy is jointly regulated by preemption, frequency, and information compression. The voice preemption priority and voice broadcasting frequency are dynamically adjusted based on the risk change rate and risk jump magnitude indicators. When the risk change rate exceeds the third threshold or the risk jump magnitude exceeds the fourth threshold, a high-priority voice preemption broadcast is triggered. The voice preemption priority and broadcasting frequency upper limit are adaptively adjusted according to the risk level to ensure shorter preemption waiting times and higher maximum broadcasting frequencies at high risk levels. Simultaneously, the effective width of the passageway and... The cognitive load index dynamically compresses the information content of speech, with the compression intensity adaptively adjusted according to the risk level. This ensures that the narrower the effective width of the passageway and the higher the cognitive load, the shorter the speech content, the more direct the instructions, and the removal of secondary explanatory information, thus avoiding information overload. For example, if a two-wheeled vehicle travels at approximately 22 km / h through a narrow road sandwiched between a parked truck on the right and the left guardrail, approaching a side road exit, the risk level rapidly rises from medium to high risk. The passage collapse index rises to 0.78, exceeding the first threshold of 0.70 after being adjusted for high risk. The risk triggering mechanism immediately switches from collision time threshold based on visible targets to passageway threshold based, while simultaneously increasing the virtual target risk weight gain from 1.0 to 1.4 to amplify the risk of sudden occlusion. At this point, the insufficient mobility index rises to 0.82, exceeding the second threshold of 0.75 after adjustment for high-risk levels. The warning suggestion output strategy implements strong gating for directional avoidance suggestions, prohibiting the output of strong left or right directional suggestions, and instead outputs a conservative suggestion to decelerate, prepare to brake, and maintain stable straight-line movement. The risk change rate index within the same time window... If the risk level exceeds the third threshold after being downgraded to a high-risk level and the risk jump magnitude index exceeds the fourth threshold, a high-priority voice preemptive broadcast is triggered, and the upper limit of the broadcast frequency is increased, so that the voice can immediately cover non-critical prompts. At the same time, the effective width of the passage is only 0.80 meters and the cognitive load index is high, so the amount of information in the voice content is compressed into short sentences, such as "Please slow down and prepare to brake as the passage ahead narrows." Thus, under high-risk, narrow passage and high load conditions, the triggering mechanism, early warning suggestions and voice broadcasts are dynamically and collaboratively controlled to ensure timely alarms, reachable suggestions and no information overload. Example calculation method for cognitive load index: To avoid limiting the "cognitive load index" to rely solely on physiological sensors, this invention provides at least one optional implementation. Let a unit time window be defined. The number of traffic participants detected inside was The target interaction count (e.g., the count of events such as cutting in, crossing, and sudden deceleration) is: The rate of change of risk score The effective width of the passageway is The speed of the vehicle is Define cognitive load indicators. for ;in and . This indicates that the input quantity is mapped to a preset upper and lower bound. The interval (upper and lower bounds can be obtained from engineering experience or statistical analysis using a sliding time window, for example, taking the most recent one) (Upper and lower bounds of quantiles within seconds) Indicates will Limit to Interval. The formula introduces... This is used to illustrate the trend that "the narrower the passage, the less reaction space the driver / cyclist has, and the higher the cognitive load."

[0044] Step 6: Generate and output AI-generated voice warning information corresponding to the risk level based on the risk level and dynamic adjustment results. The significance of this step is to translate the risk level and dynamic adjustment results obtained in Step 5 into voice output that cyclists can understand and respond to immediately. By generating and outputting AI-generated voice warning information corresponding to the risk level, different levels of risk can be matched with different levels of prompt intensity and voice expression. Simultaneously, combined with the dynamic adjustment results, the timing, frequency, information content, and structure of the voice warning information are kept consistent with the current risk status, thereby achieving higher priority and more perceptible reminders in high-risk situations. In one specific implementation, step 6, which generates and outputs AI-generated voice warning information corresponding to the risk level based on the risk level and dynamic control results, can be performed as follows: The system acquires risk level and dynamic control results and determines voice output constraints. It reads parameters such as the risk level within the current time window, the risk triggering mechanism status generated by dynamic control, voice preemption priority, upper limit of voice broadcast frequency, compression intensity of voice content information, and gating status of early warning suggestions to form a set of constraints for this voice output. The constraints, together with road scene type, effective width of passageway, urgency of obstruction threat, and degree of insufficient action accessibility, are used as context input for voice generation. Construct voice intents and sentence skeletons corresponding to risk levels. Following the principle of prioritizing prompts and guidance for low-risk levels, explicit warnings and action suggestions for medium-risk levels, and strong warnings and conservative action commands for high-risk levels, generate voice intents and select corresponding sentence skeletons. The sentence skeletons should contain at least three types of slots: hazard source indication, urgency indication, and action suggestion. Hazard source indications should prioritize interpretable elements such as narrowing passageways, sudden risks at obstructed edges, and opposing occupancy. Urgency indications should be determined based on the risk change rate and the magnitude of risk jumps to determine whether to add reinforcement words such as "immediate" and "attention". Action suggestions should be selected based on gating results, choosing conservative deceleration and preparatory braking or permissible directional avoidance suggestions. The final speech text is generated based on dynamic compression intensity and frequency strategy and then undergoes consistency processing. The sentence skeleton is trimmed and rearranged according to the compression intensity of the speech content information. When the effective width of the passage is small or the cognitive load is high, the hazard source indication and the most critical action suggestion are retained first and the explanatory components are deleted. At the same time, the speech text of continuous time windows is processed to avoid repeated broadcasting of the same content under the same risk level, which would cause interference. The output is determined based on the upper limit of the speech broadcast frequency and the preemption priority. Output voice warnings and record the triggering reasons for traceability. After satisfying the preemption and frequency constraints, output AI voice warning information corresponding to the risk level. At the same time, record the risk level, triggering conditions and dynamic control parameters associated with this output for subsequent evaluation and optimization. For example, if a two-wheeled vehicle jumps from a medium-risk level to a high-risk level in a narrow road scenario sandwiched between a guardrail and a parked truck, and the dynamic control result shows that the directional avoidance suggestion is gated and the voice content needs to be strongly compressed and allowed for high-priority preemption, then generate and immediately output a short, strong warning voice: "The passage ahead is narrowing and there is a sudden risk of obstruction. Please slow down immediately and prepare to brake." If the risk level subsequently drops back to the medium-risk level and the frequency limit restricts repeated broadcasts, then change to an intermittent output prompt: "Keep low speed and be aware of the obstructed exit on the right."

[0045] Example 2: An AI voice warning system for two-wheeled vehicle riding safety, comprising: The data acquisition module is used to acquire multi-source data related to risk warnings during two-wheeled vehicle riding and add timestamps to the multi-source data; The data preprocessing module is used to preprocess multi-source data and output preprocessed data. The scene perception module is used to identify scene elements and estimate their states based on preprocessed data, and to obtain scene perception results that include at least the states of dynamic occlusion bodies, lateral clamping constraints, and passageways. The coupled disaster state identification module is used to calculate the channel collapse degree index and the occlusion potential threat intensity index based on the scene perception results, and to identify whether the escapeable channel collapse state has been entered based on the channel collapse degree index and the occlusion potential threat intensity index; when the escapeable channel collapse state is identified, an occlusion edge risk field is established, multiple virtual targets are generated, and a probability mass is assigned to each virtual target to obtain the virtual target risk quantity corresponding to each virtual target; The risk assessment and dynamic control module is used to calculate risk scores based on scene perception results, virtual target risk quantity, and insufficient action accessibility indicators, and output risk levels based on risk scores; determine the set of dynamic control level parameters based on risk levels, and dynamically control the risk triggering mechanism, voice broadcast strategy, and early warning suggestion output strategy under the constraints of the dynamic control level parameter set; The voice warning output module is used to generate and output AI voice warning information corresponding to the risk level based on the risk level and dynamic control results.

[0046] The above formulas are all dimensionless calculations, and the preset parameters in the formulas should be set by those skilled in the art according to the actual situation.

Claims

1. An AI voice warning method for two-wheeled vehicle riding safety, characterized in that, Includes the following steps: Step 1: Obtain multi-source data related to risk warnings during two-wheeled vehicle riding, and add timestamps to the multi-source data; Step 2: Preprocess the multi-source data and output the preprocessed data; Step 3: Based on the preprocessed data, perform scene element identification and state estimation to obtain scene perception results that include at least the state of dynamic occlusion, lateral clamping constraint, and passageway. Step 4: Calculate the channel collapse degree index and the occlusion potential threat intensity index based on the scene perception results, and identify whether the escapeable channel collapse state has been entered based on the channel collapse degree index and the occlusion potential threat intensity index; when the escapeable channel collapse state is identified, establish the occlusion edge risk field, generate multiple virtual targets, and assign probability mass to each virtual target to obtain the virtual target risk quantity corresponding to each virtual target. Step 5: Calculate the risk score based on the scene perception results, virtual target risk quantity, and action accessibility insufficiency index, and output the risk level according to the risk score; The set of dynamic control parameters is determined based on the risk level, and the risk triggering mechanism, voice broadcasting strategy, and early warning suggestion output strategy are dynamically controlled under the constraints of the set of dynamic control parameters. Step 6: Generate and output AI voice warning information corresponding to the risk level based on the risk level and dynamic control results.

2. The AI ​​voice warning method for two-wheeled vehicle riding safety according to claim 1, characterized in that, Multi-source data includes at least inertial measurement data used to characterize the motion state of the two-wheeled vehicle and environmental perception data used to characterize the state of the surrounding environment.

3. The AI ​​voice warning method for two-wheeled vehicle riding safety according to claim 1, characterized in that, Data preprocessing includes time synchronization, filtering and noise reduction, coordinate system calibration, and data validity testing of multi-source data.

4. The AI ​​voice warning method for two-wheeled vehicle riding safety according to claim 1, characterized in that, The dynamic occlusion state is used to characterize the degree of occlusion of the observability of potential targets and the changes in the shape of the occlusion edge. The lateral clamping constraint state is used to characterize the available space margin on the left and right sides of the two-wheeled vehicle and its continuity. The passageway state is used to characterize the effective width, length and exit availability of the available escape passage for the two-wheeled vehicle.

5. The AI ​​voice warning method for two-wheeled vehicle riding safety according to claim 1, characterized in that, The channel collapse index is calculated from the effective width of the passage, the effective length of the passage, the availability of the passage exit, and the lateral clamping constraint state.

6. The AI ​​voice warning method for two-wheeled vehicle riding safety according to claim 1, characterized in that, The potential threat intensity index of occlusion is calculated from the dynamic occlusion state, which includes at least the degree of occlusion, the geometric range of the occlusion edge, the relative motion information of the occlusion edge, and the probability of potential targets appearing within the occlusion area.

7. The AI ​​voice warning method for two-wheeled vehicle riding safety according to claim 1, characterized in that, When the system detects the collapse of an escape route, it establishes an occlusion edge risk field and generates multiple virtual targets, including: Extract the occlusion edges of dynamic occluders from the scene perception results, and discretize the occlusion edges into several edge sampling points in the two-wheeled vehicle motion coordinate system; Using each edge sampling point as a risk source, and combining the current speed direction of the two-wheeled vehicle, the status of the passage, and the relative motion information of the obstruction, a risk potential function that decays with distance and is anisotropic along the direction of travel is constructed. Based on the risk potential function, an obstruction edge risk field is formed in the unobservable area outside the obstruction edge. Within the high-risk area of ​​the occlusion edge risk field, several candidate points are determined according to a preset spatial resolution or by searching for local extrema, and each candidate point is mapped to the hypothetical location of the virtual target. Based on the field strength value of the occlusion edge risk field at each candidate point, the relative position of the candidate point and the center line of the passage, and the escape channel margin at the candidate point, the probabilistic quality of each virtual target is determined and normalized, thereby generating multiple virtual targets and their corresponding probabilistic qualities.

8. The AI ​​voice warning method for two-wheeled vehicle riding safety according to claim 1, characterized in that, The risk score is calculated based on scene perception results, virtual target risk level, and action accessibility insufficiency indicators, including: The basic risk quantity is calculated from the scene perception results. The basic risk quantity includes at least the channel risk quantity obtained from the passage status and the occlusion risk quantity obtained from the dynamic occlusion status. The virtual target risk quantities corresponding to multiple virtual targets are aggregated according to their probabilistic quality to obtain the virtual target aggregated risk quantity; Normalize the accessibility insufficiency index to obtain the accessibility penalty. The basic risk quantity, the virtual target aggregated risk quantity, and the accessibility penalty quantity are weighted, merged, and normalized with a limit to obtain a risk score. The risk score is updated over time and used to output the risk level.

9. The AI ​​voice warning method for two-wheeled vehicle riding safety according to claim 1, characterized in that, Dynamic regulation includes: When the channel collapse index is greater than the first threshold, the risk triggering mechanism is switched from collision time threshold triggering based on visible targets to passage channel threshold triggering, and the weight gain of virtual target risk quantity and the first threshold are adaptively adjusted according to the risk level. Based on the action accessibility insufficiency index, the early warning suggestions are accessibility-gated. When the action accessibility insufficiency index is greater than the second threshold, the output of directional avoidance suggestions is prohibited and the output of conservative deceleration and preparatory braking suggestions is output. The second threshold and gating intensity are adaptively adjusted according to the risk level. The priority of voice preemption and the frequency of voice broadcasting are dynamically adjusted based on the risk change rate index and the risk jump amplitude index. When the risk change rate index is greater than the third threshold or the risk jump amplitude index is greater than the fourth threshold, a high-priority voice preemption broadcast is triggered. The upper limit of the voice preemption priority and the broadcasting frequency is adaptively adjusted according to the risk level. The amount of information in the speech content is dynamically compressed based on the effective width of the passage and the cognitive load index, and the compression intensity is adaptively adjusted according to the risk level.

10. An AI voice warning system for two-wheeled vehicle riding safety, applied to the AI ​​voice warning method for two-wheeled vehicle riding safety as described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to acquire multi-source data related to risk warnings during two-wheeled vehicle riding and add timestamps to the multi-source data; The data preprocessing module is used to preprocess multi-source data and output preprocessed data. The scene perception module is used to identify scene elements and estimate their states based on preprocessed data, and to obtain scene perception results that include at least the states of dynamic occlusion bodies, lateral clamping constraints, and passageways. The coupled disaster state identification module is used to calculate the channel collapse degree index and the occlusion potential threat intensity index based on the scene perception results, and to identify whether the escapeable channel collapse state has been entered based on the channel collapse degree index and the occlusion potential threat intensity index; when the escapeable channel collapse state is identified, an occlusion edge risk field is established, multiple virtual targets are generated, and a probability mass is assigned to each virtual target to obtain the virtual target risk quantity corresponding to each virtual target; The risk assessment and dynamic control module is used to calculate risk scores based on scene perception results, virtual target risk quantity, and action accessibility insufficiency indicators, and output risk levels based on the risk scores. The set of dynamic control parameters is determined based on the risk level, and the risk triggering mechanism, voice broadcasting strategy, and early warning suggestion output strategy are dynamically controlled under the constraints of the set of dynamic control parameters. The voice warning output module is used to generate and output AI voice warning information corresponding to the risk level based on the risk level and dynamic control results.

Images