Multi-sensor fusion riding collision detection method, device, equipment and medium

Abstract

The invention provides a multi-sensor fusion riding collision detection method, device and equipment and a medium, and the method comprises the steps: obtaining multi-source original data in a riding process, carrying out the time-space alignment and semantic fusion processing of the multi-source original data, generating a time-space unified sensing feature, and carrying out the semantic fusion processing of the multi-source original data; wherein the multi-source original data comprises kinematics sensing data, environment distance sensing data and equipment state data; collaborative feature extraction is carried out on the space-time unified sensing features, and a multi-modal dynamic risk vector is generated; based on the multi-modal dynamic risk vector, environment self-adaptive decision making is carried out, and a risk decision making instruction is output and used for executing hierarchical safety response operation. By adopting the method, real collision and interference events can be accurately distinguished in a complex environment, and an all-weather and low-power-consumption safety protection closed loop is constructed for a rider.

Description

Technical Field

[0001] The present invention belongs to the technical field of collision detection, and in particular relates to a multi-sensor fusion riding collision detection method, device, equipment and medium. Background Art

[0002] With the increasing popularity of smart cycling equipment, collision detection technology, a core component of cyclist safety, is evolving from traditional passive protection to active safety measures. Recently emerging intelligent collision detection technology uses a built-in inertial measurement unit (IMU) to monitor motion changes in real time, automatically triggering an alarm when an abnormal impact is detected. This technology primarily relies on a single acceleration sensor for threshold determination, identifying a collision as an event when the acceleration exceeds a preset critical point.

[0003] However, existing detection methods have obvious limitations: bumps and vibrations under complex road conditions can easily be misjudged as collisions, while real accidents such as low-speed skidding may be missed; at the same time, due to the lack of real-time monitoring of the helmet wearing status, when the user does not wear it correctly, it cannot provide effective protection even in the event of a collision. In addition, the continuously running sensor module causes the device's battery life to be significantly shortened, making it difficult to meet the long-term operation needs of food delivery riders and others, and it is difficult to ensure the personal safety of cyclists. Summary of the Invention

[0004] Based on this, it is necessary to provide a multi-sensor fusion riding collision detection method, device, equipment and medium to address the above technical problems. It can accurately distinguish between real collisions and interference events in complex environments through collaborative analysis and dynamic decision-making mechanism of multi-dimensional perception data, and build an all-weather, low-power safety protection closed loop for cyclists.

[0005] In a first aspect, the present application provides a multi-sensor fusion riding collision detection method, comprising:

[0006] Acquire multi-source raw data during the ride, perform spatiotemporal alignment and semantic fusion processing on the multi-source raw data, and generate spatiotemporal unified sensing features. The multi-source raw data includes kinematic perception data, environmental distance perception data, and equipment status data.

[0007] Perform collaborative feature extraction on spatiotemporal unified sensing features to generate multimodal dynamic risk vectors;

[0008] Based on the multimodal dynamic risk vector, environmental adaptive decision-making is carried out and risk decision instructions are output. The risk decision instructions are used to execute graded security response operations.

[0009] In one embodiment, multi-source raw data is obtained during a ride, and spatiotemporal alignment and semantic fusion processing are performed on the multi-source raw data to generate spatiotemporal unified sensing features, including:

[0010] Obtain acceleration sequence and angular velocity sequence in kinematic perception data;

[0011] Obtain target distance sequence and relative speed sequence in environmental distance perception data;

[0012] Performing time interpolation processing on the target distance sequence and relative velocity sequence to generate time-synchronized distance dynamic vectors;

[0013] Decompose the acceleration sequence into gravity components to separate the vertical motion acceleration component and the horizontal interference component;

[0014] The time-synchronized distance dynamic vector, vertical motion acceleration component and equipment status identification are integrated to generate a unified spatiotemporal sensing feature.

[0015] In one embodiment, collaborative feature extraction is performed on the spatiotemporal unified sensing features to generate a multimodal dynamic risk vector, including:

[0016] Performing time window fluctuation analysis on the vertical motion acceleration component to generate a translation feature quantity representing the sudden impact intensity, wherein the time window is a preset time window;

[0017] The angular velocity sequence is subjected to band-pass filtering to extract the rotational instability characteristic, where the frequency band of the band-pass filtering is the sideslip sensitive frequency band;

[0018] Perform multi-target trajectory analysis processing on the time-synchronized range dynamic vectors to generate trajectory conflict threat coefficients;

[0019] The translation feature, rotational instability feature and trajectory conflict threat coefficient are weightedly fused to generate a multimodal dynamic risk vector.

[0020] In one embodiment, performing environmental adaptive decision-making based on a multimodal dynamic risk vector and outputting a risk decision instruction include:

[0021] Get the current GPS location information and query the regional risk level through the geo-fence database;

[0022] Query the preset sensitivity mapping table based on the regional risk level and generate a dynamic sensitivity coefficient;

[0023] The multimodal dynamic risk vector is input into the random forest decision model for processing, and the basic collision risk value is output;

[0024] Performing weighted correction processing on the basic collision risk value based on the dynamic sensitivity coefficient to generate a geographically weighted risk value;

[0025] When the geographically weighted risk value exceeds the dynamic decision threshold, a risk decision instruction is output.

[0026] In one embodiment, a preset sensitivity mapping table is queried based on the regional risk level to generate a dynamic sensitivity coefficient, including:

[0027] The road topology features stored in the geographic fence database are parsed;

[0028] Based on the road topology features, spatial analysis is performed to identify high-risk area types;

[0029] Based on the high-risk area type, a preset sensitivity improvement rule is matched to obtain a dynamic sensitivity coefficient.

[0030] In one embodiment, the method for constructing the geographic fence database includes:

[0031] Obtain historical accident heat map data and extract accident density over-limit area coordinates;

[0032] Obtain real-time traffic flow data and weather feature data;

[0033] Perform linear weighting processing on the accident density over-limit area coordinates, traffic flow data, and weather feature data to generate a regional danger index;

[0034] Perform spatial clustering on the regional danger index to generate a hierarchical risk geographic fence map and load it into the geographic fence database, and the hierarchical risk geographic fence map is used to represent the accident-prone degree of regional intersections.

[0035] In one embodiment, linear weighting processing is performed on the accident density over-limit area coordinates, traffic flow data, and weather feature data to generate a regional danger index, including:

[0036] The regional danger index is calculated by the following formula:

[0037]

[0038] Wherein, R risk is the regional danger index, α is the accident density weight coefficient, is the accident point density normalized value, β is the flow sensitivity coefficient, F traffic is the number of vehicles per unit time, γ is the weather correction coefficient, W weather is the weather feature quantization value.

[0039] In a second aspect, the application also provides a multi-sensor fusion riding collision detection device, including:

[0040] The multi-source data fusion module is configured to acquire multi-source original data in the cycling process, perform spatio-temporal alignment and semantic fusion processing on the multi-source original data, and generate spatio-temporal unified sensing features, wherein the multi-source original data includes kinematic perception data, environmental distance perception data, and equipment state data.

[0041] The risk feature extraction module is configured to perform collaborative feature extraction on the spatio-temporal unified sensing features, and generate a multi-modal dynamic risk vector.

[0042] The environment-adaptive decision module is configured to perform environment-adaptive decision based on the multi-modal dynamic risk vector, and output a risk decision instruction, which is used to perform a hierarchical safety response operation.

[0043] In a third aspect, the present application further provides a computer device including a memory and a processor, wherein the memory stores a computer program, and the processor implements the multi-sensor fusion cycling collision detection method described above when executing the computer program.

[0044] In a fourth aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, and the computer program is executed by a processor to implement the multi-sensor fusion cycling collision detection method described above.

[0045] The multi-sensor fusion cycling collision detection method, device, equipment and medium described above can accurately distinguish real collisions from false positive events such as road bumps and environmental interference in a dynamic traffic scene, and reduce the false alarm rate and the missed detection rate through multi-dimensional data collaborative processing, thereby providing all-weather continuous protection for cyclists. On the premise of ensuring protection accuracy, the data-level fusion and decision optimization can reduce redundant calculation, and realize a low-power safety protection closed loop. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related art, the drawings needed in the embodiment or related art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0047] Figure 1 A schematic flow chart of a multi-sensor fusion riding collision detection method provided by an embodiment of the present invention;

[0048] Figure 2 A schematic diagram of a flow chart of an environment adaptive decision-making method provided by an embodiment of the present invention;

[0049] Figure 3 A schematic structural diagram of a multi-sensor fusion riding collision detection device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0051] First, a brief introduction is given to the terms involved in the embodiments of this application.

[0052] Multi-source raw data refers to the collection of raw physical quantities collected by heterogeneous sensors during riding. It encompasses three categories: kinematic perception data, environmental distance perception data, and equipment status data. Kinematic perception data characterizes the vehicle's dynamic state, including vector sequences such as acceleration and angular velocity. Environmental distance perception data reflects the spatial relationships between obstacles in the riding environment, including time-series information such as target distance and relative speed. Equipment status data describes the operating status of safety equipment, such as helmet fit tightness and sensor activation indicators. This multi-source nature is reflected in the differences in the data's physical dimensions, acquisition frequency, and semantic hierarchy, forming the fundamental input for subsequent fusion processing.

[0053] Spatiotemporal alignment refers to the use of interpolation compensation, coordinate transformation and other technologies to achieve the spatiotemporal benchmark unification of data streams in response to timestamp offsets and spatial coordinate system differences in multi-source raw data; semantic fusion maps data with different physical meanings into a unified risk semantic space through feature layer association. For example, acceleration fluctuations and changes in obstacle distance are associated as collision risk increments. This processing eliminates the heterogeneous contradictions of multi-source data and generates intermediate feature expressions with spatiotemporal consistency and semantic compatibility.

[0054] Based on the above explanation of terms, the implementation environment of the multi-sensor fusion riding collision detection method provided in the embodiment of the present application is explained. Schematically, the implementation environment includes: a terminal, a sensor array, a processor, and a storage device. The sensor array includes but is not limited to a six-axis inertial measurement unit, a millimeter-wave radar module, a distributed pressure sensor array, a photoelectric sensor module, a temperature and humidity composite sensor, and a positioning and navigation module; the processor can be a central processing unit, a graphics processor, a multi-core processor, or an artificial intelligence chip; the storage device can be a distributed storage device or a centralized storage device; the terminal uses a smart helmet as an example, which is not limited here.

[0055] In combination with the above explanations of terms and implementation environments, the application scenarios of the embodiments of this application are explained. The multi-sensor fusion riding collision detection method provided in the embodiments of this application can be applied to, but not limited to, the following scenarios:

[0056] During rush hour on urban roads, riders frequently shuttle between dense traffic and pedestrian areas, facing collision risks such as sudden braking, lane change, and scratches. Traditional methods, which rely solely on acceleration thresholds, can easily misjudge the frequent bumps of manhole covers as collisions, and are also prone to missing reports when they slip and fall at low speeds. This method can analyze the relative motion trajectories of surrounding vehicles / pedestrians in real time, verify the effectiveness of helmet wearing in combination with the equipment status monitoring module, generate dynamic risk vectors in real time, and make adaptive decisions based on the current environment. This not only avoids invalid alarms from interfering with the delivery rhythm, but also immediately initiates a graded safety response when a real collision occurs, significantly reducing the risk of accident handling delays.

[0057] In mountain biking scenarios, faced with the continuous bumps and sharp turns on unpaved roads, traditional single-sensor solutions struggle to distinguish between real rollover risks and ordinary road undulations. This method, through collaborative analysis of multimodal dynamic risk vectors, correlates and models posture instability characteristics with environmental slope data, effectively identifying dangerous rollovers caused by slipping on gravel or losing control on sharp turns. A graded safety response automatically adapts the response strategy based on the risk level: mild risks trigger a vibration warning prompting deceleration, while severe risks directly initiate an emergency location call for help, significantly improving the timeliness of response to accidents in the wild.

[0058] At night and in rainy and foggy environments, where perception is limited by low illumination and high humidity, the risk of failure of conventional vision solutions increases significantly. This method enhances the complementarity of millimeter-wave radar and inertial data through semantic fusion processing, cross-modally correlating the relative motion trajectory of obstacles with the vehicle's dynamic state. The environmental adaptive decision module dynamically compensates for sensor drift based on real-time temperature and humidity parameters, maintaining stable ranging capabilities even in heavy rain and dense fog. A robust data pipeline constructed from unified spatiotemporal sensing features ensures reliable collision detection even in extreme environments with visibility below 50 meters.

[0059] Illustratively, the multi-sensor fusion riding collision detection method provided in the embodiments of the present application can also be applied to other application scenarios. This is only used as an example and is not limited to specific application scenarios.

[0060] In an exemplary embodiment, Figure 1 As shown, a multi-sensor fusion riding collision detection method is provided. This embodiment uses the method as an example of a terminal in the aforementioned implementation environment. It is understandable that the method can also be applied to a server, and can also be applied to a system including a terminal and a server, and implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps 101 to 103:

[0061] Step 101: Acquire multi-source raw data during the riding process, perform spatiotemporal alignment and semantic fusion processing on the multi-source raw data, and generate spatiotemporal unified sensing features, wherein the multi-source raw data includes kinematic perception data, environmental distance perception data, and equipment status data.

[0062] Specifically, kinematic perception data is collected by the vehicle-mounted six-axis inertial measurement unit, including three-dimensional acceleration sequence and angular velocity sequence; environmental distance perception data relies on the millimeter-wave radar module to obtain obstacle distance and relative velocity vector in real time; equipment status data is generated by the distributed pressure sensor array and photoelectric sensor module embedded in the helmet to quantify the helmet wearing tightness and coverage area. The method performs spatiotemporal alignment and semantic fusion processing on the above heterogeneous data. Exemplarily, it can include the following steps: eliminating sampling delays between devices through timestamp synchronization, using coordinate transformation to unify the spatial reference system, and correlating acceleration fluctuations with obstacle motion trajectories based on the risk semantic model to generate spatiotemporal unified sensing features, forming a multi-dimensional tensor structure that integrates vehicle dynamics, environmental interaction and equipment status. This feature enables the acceleration signal generated by bumpy vibration to be cross-verified with the relative motion trajectory of environmental obstacles, avoiding safety false triggering caused by misjudgment of a single sensor.

[0063] Step 102: Perform collaborative feature extraction on the spatiotemporal unified sensing features to generate a multimodal dynamic risk vector.

[0064] Specifically, the acceleration energy integral within a 200-millisecond window can be extracted from the kinematic data to quantify the intensity of the sudden impact. The probability distribution entropy of the angular velocity vector is simultaneously analyzed to characterize the risk of attitude instability. Furthermore, a multi-target trajectory prediction algorithm is used to process the obstacle distance sequence to generate a trajectory conflict threat coefficient. This method does not calculate each feature in isolation, but rather establishes a collaborative mechanism between features. For example, the vehicle's acceleration change rate is used to modify the obstacle trajectory prediction model, and ultimately a weighted fusion is used to generate a multimodal dynamic risk vector. This vector integrates three-dimensional risk indicators: translation, rotation, and trajectory conflict. This vector integrates three weighted parameters: kinematic impact intensity, trajectory conflict threat level, and equipment safety status, to form a dynamically changing risk quantification indicator. Unlike traditional threshold judgments, this method achieves robust expression of risk characteristics through collaborative verification of multi-source data.

[0065] Step 103: Perform environmental adaptive decision-making based on the multimodal dynamic risk vector and output a risk decision instruction, which is used to execute a hierarchical security response operation.

[0066] Specifically, when executing environmentally adaptive decisions based on multimodal dynamic risk vectors, online Bayesian reasoning or reinforcement learning strategies can be employed. These strategies dynamically adjust trigger thresholds based on current vehicle speed, road surface type, and weather conditions. When the risk vector exceeds the adaptive threshold range, a risk decision instruction is immediately generated. The instruction content is divided into three levels: prompt, warning, and emergency call, depending on the severity of the impact. This enables a graded safety response, minimizing user disruption from false alarms while ensuring timely handling of real accidents. Furthermore, the prompt level broadcasts deceleration recommendations through the helmet's speakers, the warning level flashes the lights and limits motor assistance, and the emergency call level transmits location and injury summary information via the cellular network, providing comprehensive protection.

[0067] The above-mentioned multi-sensor fusion riding collision detection method, device, equipment and medium establish a multi-dimensional perception foundation by synchronously collecting multi-source raw data (including kinematic perception data, environmental distance perception data and equipment status data) during the riding process; secondly, a unified spatiotemporal unified sensing feature is generated through spatiotemporal alignment and semantic fusion processing to eliminate the spatiotemporal misalignment and semantic barriers between sensors; then, collaborative feature extraction is performed on the fused features to generate a multimodal dynamic risk vector, realizing cross-modal joint modeling of risk features; relying on the environmental adaptive decision-making mechanism to output risk decision instructions to drive hierarchical safety response operations. The above-mentioned technical solution can accurately distinguish between real collisions and false alarm events such as road bumps and environmental interference in dynamic traffic scenarios, and at the same time reduce the false alarm rate and missed detection rate through multi-dimensional data collaborative processing, providing cyclists with all-weather continuous protection. On the premise of ensuring protection accuracy, redundant calculations are reduced through data-level fusion and decision optimization to achieve a low-power safety protection closed loop.

[0068] In one embodiment, multi-source raw data is obtained during a ride, and spatiotemporal alignment and semantic fusion processing are performed on the multi-source raw data to generate spatiotemporal unified sensing features, including:

[0069] Obtain acceleration sequence and angular velocity sequence in kinematic perception data;

[0070] Obtain target distance sequence and relative speed sequence in environmental distance perception data;

[0071] Performing time interpolation processing on the target distance sequence and relative velocity sequence to generate time-synchronized distance dynamic vectors;

[0072] Decompose the acceleration sequence into gravity components to separate the vertical motion acceleration component and the horizontal interference component;

[0073] The time-synchronized distance dynamic vector, vertical motion acceleration component and equipment status identification are integrated to generate a unified spatiotemporal sensing feature.

[0074] Specifically, the acceleration sequence and angular velocity sequence in the kinematic perception data are acquired, which can be collected by a six-axis inertial measurement unit (IMU) built in the smart helmet in real time, wherein the acceleration sequence contains the linear motion acceleration components of the helmet in three-dimensional space, and the angular velocity sequence records the rotational angular velocity of the helmet around three coordinate axes. The target distance sequence and relative velocity sequence in the environmental distance perception data are synchronously acquired, which can be generated by the millimeter wave radar module integrated in front of the helmet, and the absolute distance change trend and the motion speed relative to the rider of the obstacle within the range of 5-80 meters in front are captured in real time. Further, the embodiment performs time interpolation processing on the target distance sequence and the relative velocity sequence, and specifically adopts a linear interpolation algorithm to align the radar data sampling rate (typical value 10 Hz) with the IMU sampling rate (typical value 100 Hz). The processing is based on the IMU timestamp, and the radar data is resampled and filled to generate a time-synchronized distance dynamic vector. The vector integrates the distance decay rate and the relative speed direction of the obstacle in units of 20 milliseconds. For example, when the radar detects that a vehicle in front suddenly slows down, the vector will synchronously reflect the sharp shortening of the distance and the negative speed increment, providing time-consistent input for subsequent trajectory analysis. Further, the acceleration sequence is subjected to gravity component decomposition processing, and the original acceleration in the device coordinate system is converted to the geodetic coordinate system through a direction cosine matrix. The specific decomposition process is as follows: based on the angular velocity sequence, the real-time orientation of the helmet is estimated, and the motion acceleration component in the vertical ground direction is separated from the original acceleration, while the horizontal high-frequency noise caused by road bumps is filtered out. The vertical motion acceleration component after this processing can accurately reflect key motion characteristics such as vehicle lifting, collision impact, etc. For example, when riding uphill, a positive and continuous acceleration is presented, and a pulse-like spike appears at the moment of collision. Finally, the time-synchronized distance dynamic vector, the vertical motion acceleration component, and the equipment state identifier are fused to generate a spatiotemporal unified sensing feature, wherein the equipment state identifier is generated by the internal pressure sensor array and the photoelectric sensor of the helmet: the pressure sensor detects the headband tension value, and the photoelectric sensor quantifies the contact area proportion of the helmet and the head. The two are combined into a binary state identifier (e.g., 1 represents correct wearing, and 0 represents abnormal wearing). Exemplarily, the fusion process maps the distance dynamic vector to a polar coordinate obstacle distribution matrix, converts the vertical motion acceleration component to a frequency energy spectrum, and takes the equipment state identifier as a feature marker. The three are spliced into an 8-dimensional feature vector through tensor. The feature vector simultaneously carries three pieces of information: environmental risk, motion impact, and equipment state. For example, when an obstacle is detected to approach quickly (distance dynamic vector anomaly), the vertical acceleration mutates (energy spectrum spike), and the helmet is effectively worn, the feature vector will present a high-dimensional risk pattern, providing a fused perception benchmark for subsequent collision decision-making.

[0075] In one of the embodiments, the spatiotemporal unified sensing feature is subjected to collaborative feature extraction to generate a multi-modal dynamic risk vector, including:

[0076] Performing time window fluctuation analysis on the vertical motion acceleration component to generate a translation feature quantity representing the sudden impact intensity, wherein the time window is a preset time window;

[0077] The angular velocity sequence is subjected to band-pass filtering to extract the rotational instability characteristic, where the frequency band of the band-pass filtering is the sideslip sensitive frequency band;

[0078] Perform multi-target trajectory analysis processing on the time-synchronized range dynamic vectors to generate trajectory conflict threat coefficients;

[0079] The translation feature, rotational instability feature and trajectory conflict threat coefficient are weightedly fused to generate a multimodal dynamic risk vector.

[0080] Specifically, in this embodiment, the vertical motion acceleration component obtained by the gravity decomposition process is subjected to time window fluctuation analysis processing, which characterizes the motion impact characteristics of the vehicle in the up-down direction. Exemplarily, a fixed time length sliding window (typically a window covering a 200-500 millisecond period) can be used to calculate the root mean square fluctuation value of the acceleration within the window, generating a translation feature quantity that quantifies the intensity of the sudden impact, for example, when the vehicle is in a rear-end collision, a continuous positive high amplitude fluctuation will occur within the window, while road bumps only produce transient spikes, thereby effectively distinguishing between real collisions and interference vibrations. Further, the angular velocity sequence, which can be derived from the real-time acquisition data of the helmet six-axis IMU, is subjected to band-pass filtering processing, wherein the filter band can be set to a 0.5-2Hz side slip sensitive frequency band, which covers the typical roll frequency when the vehicle loses balance. By suppressing high-frequency noise (such as helmet fine adjustment actions) and low-frequency shaking (such as uniform turning), the rotation instability feature quantity is extracted, which presents an amplitude steep increase characteristic when the vehicle is side slipping, for example, when the wheels suddenly hit the curb, the vertical direction pulse signal captured by the helmet sensor will be identified as a high amplitude feature quantity, which is different from the low frequency oscillation generated by the regular jolt of the vehicle. This processing effectively suppresses the interference signals caused by road unevenness, providing a core mechanical basis for subsequent collision judgment. Further, the time-synchronized distance dynamic vector is subjected to multi-target trajectory analysis processing, generating a trajectory conflict threat coefficient. This method establishes a dynamic interaction model based on the obstacle distance and relative speed data provided by the millimeter wave radar in real time. Specifically, it can include: distinguishing different target individuals through clustering algorithms, calculating the intersection time window of each target and the rider's travel path, and generating an exponentially increasing threat coefficient when a target object is detected to enter the 2.5 second collision countdown threshold range. For example, when a left-turning car is detected to cut into the path during a turn, the threat coefficient will automatically increase and the helmet sound and light alarm will be activated. This processing upgrades the environmental perception from static distance monitoring to dynamic behavior prediction. Further, the translation feature quantity, the rotation instability feature quantity and the trajectory conflict threat coefficient are fused to generate a multi-modal dynamic risk vector, and the weight distribution strategy can be dynamically configured according to the riding scene, for example: the trajectory conflict weight accounts for the highest proportion (about 50%) in urban roads, and the rotation instability weight is increased to 60% in mountain scenes. The fusion process adopts feature normalization and weighted summation, and finally outputs a three-dimensional risk vector. In the scenario of a delivery rider crossing an intersection diagonally, this vector can simultaneously reflect the combined risk state of a vehicle in front suddenly braking (high translation feature), a wet road side slipping (high rotation feature), and an oncoming vehicle approaching (high trajectory threat).This embodiment uses a three-dimensional feature extraction architecture with complementary physical properties: translation features capture longitudinal collisions, rotation features identify lateral instability, and trajectory features warn of environmental conflicts. It breaks through the limitations of traditional single-dimensional detection and can identify the concurrent dangers of vehicles that illegally change lanes but are not captured by the camera due to backlight (a surge in trajectory threats) and the risk of tailspin when the rider dodges (abnormal rotation features), thereby reducing the missed detection rate of accidental risks such as collisions.

[0081] like Figure 2 As shown, in one embodiment, an environment adaptive decision is made based on a multimodal dynamic risk vector, and a risk decision instruction is output, including:

[0082] Step 201, obtain current GPS location information and query the regional risk level through the geo-fence database;

[0083] Step 202: querying a preset sensitivity mapping table based on the regional risk level to generate a dynamic sensitivity coefficient;

[0084] Step 203: Input the multimodal dynamic risk vector into the random forest decision model for processing, and output a basic collision risk value;

[0085] Step 204 , performing weighted correction processing on the basic collision risk value based on the dynamic sensitivity coefficient to generate a geographically weighted risk value;

[0086] Step 205: When the geographically weighted risk value exceeds the dynamic decision threshold, a risk decision instruction is output.

[0087] Specifically, the current GPS position information is acquired, which is generated in real time by the GNSS module built-in or bound to the helmet by the rider, and the regional risk level is queried by a geo-fencing database, which preloads the accident heat map layer and road topological features of different geographic locations, such as marking a 50-meter range of city intersections as a high-risk area and marking a non-motor vehicle lane as a medium-risk area, and the query outputs a regional risk level quantization value (1-5 levels) directly reflecting the probability of accident occurrence at the current location. Based on the regional risk level query, a preset sensitivity mapping table is generated to generate a dynamic sensitivity coefficient, and the mapping table establishes a nonlinear correspondence between the risk level and the decision sensitivity, for example: the high-risk area (such as the intersection) triggers a sensitivity enhancement coefficient (typical value 1.5-2.0 times), and the low-risk area (such as a closed greenway) adopts a sensitivity reduction coefficient (0.7-0.9 times), which is used to dynamically adjust the decision threshold, for example, automatically reducing the collision judgment threshold by 20% in the bus stop area to cope with the sudden risk of pedestrians crossing. Further, a multi-modal dynamic risk vector is input into a random forest decision model for processing, and a basic collision risk value is output. The model is trained by ten thousand accident samples, and the input vector includes three dimensions of translation feature quantity, rotation instability feature quantity and trajectory conflict threat coefficient. The model calculates in parallel through multiple decision trees, for example, when the translation feature quantity exceeds the tree node split threshold and the trajectory conflict feature continuously increases, a high-risk probability value is output. The output result is a basic collision risk value (0-1 interval), representing the comprehensive danger degree of the current state. Further, the method performs weighted correction on the basic collision risk value and the dynamic sensitivity coefficient, for example, the influence of the geographical environment on risk perception can be superimposed on the instantaneous physical risk through a multiplication operation to generate a geographical weighted risk value; in a low-risk street, even if the vehicle has a large impact, the geographical weighted risk value may still be suppressed below the alarm threshold, while in a high-risk street, a slight abnormality may be amplified to the trigger level, ensuring that the alarm strategy is synchronized with the real road danger. Further, when the geographical weighted risk value exceeds the dynamic decision threshold, a risk decision instruction is output, wherein the dynamic decision threshold can be set to adaptively adjust according to the real-time vehicle speed: a low-speed ride (<15km / h) adopts a 0.75 threshold, and a high-speed ride (>30km / h) is reduced to a 0.6 threshold. When the geographical weighted risk value exceeds the threshold for 0.5 seconds, a high-risk collision instruction is output; if it fluctuates around the threshold, a risk warning instruction is output, which can be transmitted to the rider's mobile phone APP through the Bluetooth module of the helmet, triggering corresponding level sound and light alarm or automatic help.This embodiment breaks through the scenario adaptability limitations of traditional collision detection systems through collaborative decision-making based on geo-fence dynamic parameters and real-time sensor data. The environmental adaptive decision-making mechanism enables the method to automatically increase the protection level in accident blackspot areas, while reducing the response frequency in safe sections to avoid disturbing the public. The weighted correction processing overcomes the mechanical judgment defects of fixed thresholds, and the multi-feature fusion of the random forest model significantly improves the risk assessment accuracy in complex scenarios. The resulting hierarchical response strategy ensures timely intervention while minimizing riding interference caused by false triggering, thereby achieving a substantial improvement in the active safety protection capabilities of smart helmets.

[0088] In one embodiment, querying a preset sensitivity mapping table based on the regional risk level to generate a dynamic sensitivity coefficient includes:

[0089] Parsing road topology features stored in the geofence database;

[0090] Conduct spatial analysis based on road topology characteristics to identify high-risk area types;

[0091] Based on the high-risk area type matching preset sensitivity improvement rules, a dynamic sensitivity coefficient is obtained.

[0092] For example, when generating the dynamic sensitivity coefficient, the road topology features stored in the geofence database are first analyzed. These features include vector layer data such as intersection density, lane change points, and non-motorized vehicle lane interruption locations. These features can be accessed in real time by invoking the geofence database interface through the helmet's built-in processor. For example, in an urban delivery scenario, the number of road intersections and the density of crosswalks within a 100-meter radius of the current location are extracted to form a topological feature vector. This vector quantifies the complexity of the regional road network and provides a spatial structural basis for risk classification. Furthermore, spatial analysis based on road topology features can be used to identify high-risk areas using a rule-based spatial relationship determination engine. When intersection density exceeds a threshold and a crosswalk is detected, it is marked as a "high-risk pedestrian crossing area." When a sudden lane change point overlaps with a non-motorized vehicle lane interruption, it is marked as a "high-risk lane conflict area." This analysis avoids relying solely on historical accident data and identifies potential danger zones through topological logical relationships. For example, a typical high-risk section where the non-motorized vehicle lane suddenly narrows before a bus stop can be identified. Furthermore, based on the matching of high-risk area types, the preset sensitivity enhancement rules are used to obtain a dynamic sensitivity coefficient, wherein the preset rule base on which the matching is based contains three core strategies: a linear enhancement rule is used for high-risk areas for pedestrians (the coefficient is enhanced by 0.1 times for each additional intersection, with an upper limit of 2.0 times); a step-by-step enhancement rule is enabled for high-risk areas for lane conflicts (a fixed enhancement of 1.8 times when a lane interruption is detected); and a baseline coefficient of 1.0 is maintained for other areas. The matching process outputs a dynamic sensitivity coefficient. For example, when a bus stop area is identified, a 1.8-fold coefficient is automatically triggered, so that the subsequent decision-making module remains highly sensitive to sudden risks in the area. This embodiment replaces the traditional reliance on accident statistics by identifying road structure features, and can still effectively warn in newly built sections without historical accident data to ensure the personal safety of cyclists.

[0093] In one embodiment, a method for constructing a geo-fence database includes:

[0094] Obtain historical accident thermal map data and extract the coordinates of areas where accident density exceeds the limit;

[0095] Obtain real-time traffic flow data and weather characteristics data;

[0096] Perform linear weighting processing on the coordinates of the accident density exceeding limit area, traffic flow data and weather characteristic data to generate the regional hazard index;

[0097] The regional hazard index is spatially classified through a spatial clustering algorithm to generate a hierarchical risk geofence map and load it into the geofence database. The hierarchical risk geofence map is used to characterize the susceptibility of regional intersection accidents.

[0098] For example, when constructing a geo-fence database, historical accident heat map data is first obtained. This data can be sourced from the public accident database of the traffic management department and contains fields such as the accident latitude and longitude coordinates, timestamp, and accident type. The accident coordinate points are processed using a spatial density analysis algorithm (such as kernel density estimation) to extract high-risk coordinate clusters where the accident density exceeds two standard deviations of the regional mean. For example, red hot zones where accident points are concentrated within a 50-meter radius of the intersection are identified on urban roads, and a coordinate set of areas with excessive accident density is generated. Real-time traffic flow data and weather characteristic data are obtained. Traffic flow data can be collected in real time through the interface of the urban traffic information platform, including the number of vehicles passing through a specified road section per unit time and the average speed; weather characteristic data is obtained from the meteorological service API, quantifying parameters such as rainfall intensity and visibility as standardized values ​​in the range of 0-1. This data is updated every 5 minutes to ensure the timeliness of the hazard index assessment, such as capturing the concurrent scenario of a surge in traffic on main roads and heavy rain during the evening rush hour. Furthermore, linear weighted processing is performed on the multidimensional data. This method establishes a quantitative assessment system for accident density, traffic flow, and weather characteristics. For example, accident density is set to a maximum weight coefficient of 0.6, traffic flow dynamic weight range is 0.2-0.4 (the more congested the traffic, the higher the weight), and weather characteristics use a fixed influence factor of 0.1. Specifically, the parameter values ​​are first normalized to the range of 0-1, and then a comprehensive score is calculated according to the formula (regional hazard index = 0.6 × accident density value + real-time traffic weight × congestion coefficient + 0.1 × weather index). This process fuses discrete data into a single risk indicator. For example, at an intersection, the historical frequency of accidents (high density weight), real-time congestion (high flow value), and slippery road surface (high weather value) are superimposed to generate a hazard index of 0.85. This algorithm breaks through the limitations of single accident statistics, allowing the calculation results to reflect the instantaneous risks of the real environment. Furthermore, the regional hazard index can be spatially classified by a spatial clustering algorithm (such as DBSCAN), and continuous areas with spatial proximity and a hazard index difference of less than 0.1 can be merged with the road network as a constraint. For example, the generated hierarchical risk geo-fence map can include three levels of fences, such as: core area (hazard index>0.8), warning area (0.6-0.8), and safety area (<0.6). The boundaries of each fence are stored as vector polygons. The map is eventually loaded into the geo-fence database and the query is accelerated through spatial indexing. For example, the 500-meter range around the school is marked as the core area and associated with high-risk decision parameters. In this embodiment, the database construction method can break through the limitations of traditional static geo-fences, lock the spatial risk source points through accident density analysis, integrate real-time traffic and weather elements to establish a dynamic assessment model, and generate a hierarchical risk map based on spatial clustering.For example, in the application of smart helmets, when a rider enters an area with no historical data, such as a newly built road, the method can automatically generate a temporary danger index based on the real-time traffic speed drop (traffic data) and fog warning (weather data), dynamically expand the coverage of the geographic fence, and enable the risk perception capability to cover urban blind spots, thereby improving the level of risk warning.

[0099] In one embodiment, a linear weighted process is performed on the coordinates of the accident density exceeding limit area, traffic flow data, and weather characteristic data to generate a regional danger index, including:

[0100] The regional hazard index is calculated using the following formula:

[0101]

[0102] Among them, R risk is the regional hazard index, α is the accident density weight coefficient, is the normalized value of accident point density, β is the flow sensitivity coefficient, F traffic is the number of vehicles per unit time, γ is the weather correction coefficient, W weather Quantify values ​​for weather characteristics.

[0103] Specifically, the normalized value of accident point density The spatial statistical analysis of the historical accident heat map is carried out by performing grid density calculation on the accident heat map, and the original number of accident points is converted into a normalized density score in the range of 0-1. For example, the score of the accident-prone area at a certain intersection is 0.92; the number of vehicles per unit time F traffic The number of vehicles passing through a specific road section per minute is quantified through real-time acquisition of roadside intelligent sensing equipment or vehicle-mounted communication modules; the weather characteristic quantification value W weather The parameters such as precipitation intensity and visibility released by the meteorological department are mapped into numerical scales. For example, the corresponding quantitative value of rainstorm is 3.0, and that of sunny day is 0.5. The accident density weight coefficient α is usually set in the range of 0.6-0.8, giving the historical accident data the highest decision weight. The flow sensitivity coefficient β is set in the range of 0.2-0.3, and the logarithmic function log(F traffic +1) compresses the numerical expansion effect of high-traffic areas; the weather correction coefficient γ adopts a dynamic adjustment strategy and automatically increases to 0.5 in foggy weather to strengthen the impact of adverse environments. This formula achieves risk fusion of three types of heterogeneous data through weighted combination: historical accident density constitutes the basic risk base, real-time traffic flow reflects the instantaneous risk increase, and weather factors provide environmental correction factors. Further, the regional risk index R is generated. risk Then, the index values ​​of adjacent grids are aggregated into continuous geographic fences through spatial clustering algorithm. For example, when R is calculated for a certain intersection area, risk= 7.2 (e.g., full score of 10), automatically marking it as a red high-risk fence, and the generated danger index is used as an input for a spatial clustering algorithm, so that new road areas (without historical accident data) can still generate risk values dynamically according to real-time traffic and weather, improving the coverage rate of blind area road section early warning.

[0104] In summary, the multi-sensor fusion riding collision detection method provided in the present application collects kinematic perception data, environmental distance perception data and equipment state data simultaneously by the sensor array in the intelligent helmet, forms a spatio-temporal unified sensing feature through spatio-temporal alignment and semantic fusion, and then generates a multi-modal dynamic risk vector that can reflect the dangerous situation in real time through collaborative feature extraction; on this basis, the current GPS position is queried in the geographic fence database, the road topological features are analyzed and the regional risk level is matched, and then the dynamic sensitivity coefficient is generated by calling the preset sensitivity mapping table, and the basic collision risk value output by the random forest is weighted and corrected by using the coefficient, and finally the geographic weighted risk value is obtained to trigger the graded safety response; wherein the geographic fence database itself is constructed offline through linear weighting and spatial clustering of historical accident heat map, real-time traffic flow and weather feature data, and the hierarchical risk map is resident in the helmet with the firmware, realizing millisecond-level query without network dependence. The above technical solution can accurately distinguish between real collision events and road bumps, effectively identify dangerous scenarios such as low-speed side slipping that are easily missed by traditional solutions, and at the same time optimize system energy consumption through a graded response mechanism, so as to accurately distinguish between real collisions and interference events in complex environments, achieving a triple breakthrough in protection accuracy, environmental adaptability and energy economy, and building an intelligent active safety barrier for cyclists.

[0105] It should be understood that, although each step in the flowchart involved in each embodiment as described above is shown in sequence according to the direction of the arrow, these steps are not necessarily executed in sequence according to the direction of the arrow. Unless otherwise specified herein, there is no strict order limitation for the execution of these steps, and these steps can be executed in other orders. Moreover, at least part of the steps in the flowchart involved in each embodiment as described above can include multiple steps or stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but can be executed in rotation or alternation with at least part of other steps or steps or stages in other steps.

[0106] Based on the same inventive concept, embodiments of the present application also provide a multi-sensor fusion riding collision detection device 10 for implementing the multi-sensor fusion riding collision detection method described above. The solution provided by this device is similar to the solution described in the aforementioned method. Therefore, the specific limitations of one or more embodiments of the multi-sensor fusion riding collision detection device 10 provided below can be found in the above-mentioned limitations of the multi-sensor fusion riding collision detection method and will not be further elaborated here.

[0107] In an exemplary embodiment, Figure 3 As shown, a multi-sensor fusion riding collision detection device 10 is provided, comprising:

[0108] A multi-source data fusion module 11 is used to obtain multi-source raw data during the riding process, perform spatiotemporal alignment and semantic fusion processing on the multi-source raw data, and generate spatiotemporal unified sensing features, wherein the multi-source raw data includes kinematic perception data, environmental distance perception data, and equipment status data;

[0109] The risk feature extraction module 12 is used to perform collaborative feature extraction on the spatiotemporal unified sensing features to generate a multimodal dynamic risk vector;

[0110] The environment adaptive decision module 13 is used to make environment adaptive decisions based on the multimodal dynamic risk vector and output risk decision instructions, which are used to execute hierarchical security response operations.

[0111] In one embodiment, the multi-source data fusion module 11 includes:

[0112] A motion data acquisition unit, used to acquire acceleration sequences and angular velocity sequences in kinematic perception data;

[0113] An environmental perception unit, used to obtain a target distance sequence and a relative speed sequence in environmental distance perception data;

[0114] A time alignment unit is used to perform time interpolation processing on the target distance sequence and the relative velocity sequence to generate a time-synchronized distance dynamic vector;

[0115] The gravity decomposition unit is used to decompose the acceleration sequence into gravity components and separate the vertical motion acceleration component and the horizontal interference component;

[0116] The feature fusion unit is used to fuse the time-synchronized distance dynamic vector, vertical motion acceleration component and equipment status identification to generate a unified spatiotemporal sensing feature.

[0117] In one embodiment, the risk feature extraction module 12 includes:

[0118] An impact analysis unit is used to perform time window fluctuation analysis on the vertical motion acceleration component to generate a translation feature quantity representing the sudden impact intensity, wherein the time window is a preset time window;

[0119] An instability detection unit is used to perform band-pass filtering on the angular velocity sequence to extract the rotational instability characteristic, wherein the frequency band of the band-pass filtering is the sideslip sensitive frequency band;

[0120] a trajectory prediction unit for performing multi-target trajectory analysis processing on the time-synchronized range dynamic vectors to generate a trajectory conflict threat coefficient;

[0121] The vector synthesis unit is used to weightedly fuse the translation feature quantity, rotational instability feature quantity and trajectory conflict threat coefficient to generate a multimodal dynamic risk vector.

[0122] In one embodiment, the environment adaptation decision module 13 includes:

[0123] Geographic query unit, used to obtain current GPS location information and query regional risk level through the geo-fence database;

[0124] A sensitivity mapping unit, configured to query a preset sensitivity mapping table based on the regional risk level and generate a dynamic sensitivity coefficient;

[0125] The model decision unit is used to input the multimodal dynamic risk vector into the random forest decision model for processing and output the basic collision risk value;

[0126] a risk correction unit, configured to perform weighted correction processing on the basic collision risk value based on the dynamic sensitivity coefficient to generate a geographically weighted risk value;

[0127] The instruction triggering unit is used to output a risk decision instruction when the geographically weighted risk value exceeds the dynamic decision threshold.

[0128] In one embodiment, the sensitivity mapping unit includes:

[0129] a topology parsing subunit, configured to parse road topology features stored in the geo-fence database;

[0130] The spatial identification subunit is used to perform spatial analysis based on road topological characteristics and identify high-risk area types;

[0131] The rule matching subunit is used to match the preset sensitivity improvement rules based on the high-risk area type to obtain the dynamic sensitivity coefficient.

[0132] In one embodiment, in the topology parsing subunit, the method for constructing a geo-fence database includes the following steps:

[0133] Obtain historical accident thermal map data and extract the coordinates of areas where accident density exceeds the limit;

[0134] Obtain real-time traffic flow data and weather characteristics data;

[0135] Perform linear weighting processing on the coordinates of the accident density exceeding limit area, traffic flow data and weather characteristic data to generate the regional hazard index;

[0136] The regional hazard index is spatially classified through a spatial clustering algorithm to generate a hierarchical risk geofence map and load it into the geofence database. The hierarchical risk geofence map is used to characterize the susceptibility of regional intersection accidents.

[0137] In one embodiment, the index calculation unit may be configured to perform the following steps:

[0138] The regional hazard index is calculated using the following formula:

[0139]

[0140] Among them, R risk is the regional hazard index, α is the accident density weight coefficient, is the normalized value of accident point density, β is the flow sensitivity coefficient, F traffic is the number of vehicles per unit time, γ is the weather correction coefficient, W weather Quantify values ​​for weather characteristics.

[0141] In one embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, the steps of the multi-sensor fusion riding collision detection method as described above are implemented.

[0142] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0143] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to the partial description of the method embodiments. The device embodiments described above are merely illustrative, wherein the components described as separate parts may or may not be physically separated, and the parts displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the disclosed solution. A person of ordinary skill in the art can understand and implement it without expending creative work.

[0144] The above-described embodiments only express several implementation manners of the application, the description is more specific and detailed, but it cannot be understood as the limitation of the patent scope of the application. It should be pointed out that for ordinary skilled in the art, without departing from the concept of the application, several modifications and improvements can be made, which are within the protection scope of the application.

Claims

1. A multi-sensor fusion riding collision detection method, characterized in that: The method comprises: Acquiring multi-source raw data during riding, performing spatiotemporal alignment and semantic fusion processing on the multi-source raw data to generate spatiotemporal unified sensing features, wherein the multi-source raw data includes kinematic perception data, environmental distance perception data, and equipment status data; Performing collaborative feature extraction on the spatiotemporal unified sensing features to generate a multimodal dynamic risk vector; An environment adaptive decision is made based on the multimodal dynamic risk vector, and a risk decision instruction is output, where the risk decision instruction is used to execute a graded security response operation.

2. The method according to claim 1, characterized in that The step of acquiring multi-source raw data during the riding process, performing spatiotemporal alignment and semantic fusion processing on the multi-source raw data, and generating spatiotemporal unified sensing features includes: Acquiring an acceleration sequence and an angular velocity sequence from the kinematic perception data; Acquiring a target distance sequence and a relative speed sequence from the environmental distance perception data; performing time interpolation processing on the target distance sequence and the relative speed sequence to generate a time-synchronized distance dynamic vector; Performing gravity component decomposition processing on the acceleration sequence to separate the vertical motion acceleration component and the horizontal interference component; The time-synchronized distance dynamic vector, vertical motion acceleration component and equipment status identifier are integrated to generate the time-space unified sensing feature.

3. The method according to claim 2, characterized in that The performing collaborative feature extraction on the spatiotemporal unified sensing features to generate a multimodal dynamic risk vector includes: Performing a time window fluctuation analysis on the vertical motion acceleration component to generate a translation feature quantity characterizing the sudden impact intensity, wherein the time window is a preset time window; performing band-pass filtering on the angular velocity sequence to extract a rotational instability feature, wherein the frequency band of the band-pass filtering is a sideslip sensitive frequency band; performing multi-target trajectory analysis processing on the time-synchronized range dynamic vector to generate a trajectory conflict threat coefficient; The translation feature quantity, the rotational instability feature quantity and the trajectory conflict threat coefficient are weightedly fused to generate the multimodal dynamic risk vector.

4. The method according to claim 1, wherein The step of performing an environment adaptive decision based on the multimodal dynamic risk vector and outputting a risk decision instruction includes: Get the current GPS location information and query the regional risk level through the geo-fence database; Querying a preset sensitivity mapping table based on the regional risk level to generate a dynamic sensitivity coefficient; Inputting the multimodal dynamic risk vector into a random forest decision model for processing, and outputting a basic collision risk value; performing weighted correction processing on the basic collision risk value based on the dynamic sensitivity coefficient to generate a geographically weighted risk value; When the geographically weighted risk value exceeds a dynamic decision threshold, the risk decision instruction is output.

5. The method according to claim 4, characterized in that The querying of a preset sensitivity mapping table based on the regional risk level to generate a dynamic sensitivity coefficient includes: Parsing road topology features stored in the geofence database; Perform spatial analysis based on the road topology characteristics to identify high-risk area types; The dynamic sensitivity coefficient is obtained based on the preset sensitivity improvement rule matched with the high-risk area type.

6. The method according to claim 4, characterized in that The method for constructing the geo-fence database includes: Obtain historical accident thermal map data and extract the coordinates of areas where accident density exceeds the limit; Obtain real-time traffic flow data and weather characteristics data; Performing linear weighted processing on the coordinates of the accident density exceeding limit area, the traffic flow data, and the weather characteristic data to generate a regional hazard index; The regional hazard index is spatially classified by a spatial clustering algorithm to generate a hierarchical risk geofence map and load it into the geofence database. The hierarchical risk geofence map is used to characterize the susceptibility of regional intersection accidents.

7. The method according to claim 6, characterized in that The performing of linear weighted processing on the coordinates of the accident density exceeding limit area, the traffic flow data, and the weather characteristic data to generate a regional danger index includes: The regional hazard index is calculated using the following formula: Among them, R risk is the regional hazard index, α is the accident density weight coefficient, is the normalized value of accident point density, β is the flow sensitivity coefficient, F traffic is the number of vehicles per unit time, γ is the weather correction coefficient, W weather Quantify values ​​for weather characteristics.

8. A multi-sensor fusion riding collision detection device, characterized in that: The device comprises: A multi-source data fusion module is used to obtain multi-source raw data during the riding process, perform spatiotemporal alignment and semantic fusion processing on the multi-source raw data, and generate spatiotemporal unified sensing features, wherein the multi-source raw data includes kinematic perception data, environmental distance perception data, and equipment status data; a risk feature extraction module, configured to perform collaborative feature extraction on the spatiotemporal unified sensing features to generate a multimodal dynamic risk vector; An environment adaptive decision module is used to make an environment adaptive decision based on the multimodal dynamic risk vector and output a risk decision instruction, wherein the risk decision instruction is used to execute a graded security response operation.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the method according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images