Bicycle driving risk early warning method and system based on visual identification

Abstract

The invention discloses a bicycle driving risk early warning method and system based on visual identification, and relates to the technical field of visual identification. The method comprises the following steps: acquiring driving sensing data through a sensing module carried on a bicycle body; identifying a plurality of dynamic traffic participants based on the visual perception data, and predicting future behavior intentions and motion trails; carrying out risk assessment by fusing the predicted behavior intention, the motion trail and the running state data of the bicycle, and constructing a risk perception grid which comprises a traffic event, a participant, a risk level and a movement tracking state; and based on the risk perception grid, identifying the traffic event with the risk level or the mobile tracking state reaching an early warning threshold value and the positioning space, and generating early warning feedback information. The technical problems of insufficient bicycle driving risk identification and untimely early warning in the prior art are solved, and the technical effects of driving risk active perception and timely early warning based on visual prediction are achieved.

Description

Technical Field

[0001] This invention relates to the field of visual recognition technology, specifically to a method and system for early warning of bicycle riding risks based on visual recognition. Background Technology

[0002] With the continuous growth of urban traffic flow, the safety of cyclists in mixed traffic environments is becoming increasingly prominent. Traditional bicycles rely heavily on the rider's subjective observation for risk assessment. Limited by field of vision, reaction speed, and environmental complexity, cyclists often struggle to identify potential risks in time when pedestrians, motor vehicles, electric vehicles, or other dynamic traffic participants are rapidly approaching or suddenly changing direction, leading to a significant increase in collision risk. Existing auxiliary warning technologies are mainly based on simple distance detection or fixed threshold judgments, which cannot predict the behavioral intentions and movement trends of traffic participants. They are unable to effectively respond to sudden or continuously evolving traffic risks, resulting in delayed warnings and insufficient reliability, failing to meet the needs of safe cyclist operation in complex traffic scenarios. Summary of the Invention

[0003] This application provides a method and system for early warning of bicycle riding risks based on visual recognition, which solves the technical problems of insufficient identification of bicycle riding risks and untimely early warning in the prior art.

[0004] The first aspect of this application provides a method for warning of bicycle riding risks based on visual recognition, the method comprising: The bicycle acquires driving perception data through a sensing module mounted on its body, the sensing module including at least a visual sensing device; based on the visual perception data, it identifies multiple dynamic traffic participants and predicts their future behavioral intentions and movement trajectories; based on the dynamic intersection risks of the bicycle in a mixed traffic environment, it integrates the predicted behavioral intentions, movement trajectories, and the bicycle's own operating status data to conduct a risk assessment and construct a risk perception grid, including traffic events, participants, risk levels, and movement tracking status; based on the risk perception grid, it identifies traffic events and their location spaces where the risk level or movement tracking status reaches the warning threshold, and generates warning feedback information.

[0005] A second aspect of this application provides a vision-based bicycle riding risk warning system, the system comprising: Data Acquisition Component: Acquires driving perception data through a sensing module mounted on the bicycle itself, the sensing module including at least a visual perception device; Behavior Prediction Component: Identifies multiple dynamic traffic participants based on visual perception data and predicts future behavioral intentions and movement trajectories; Risk Assessment Component: Based on the dynamic intersection risks of bicycles in mixed traffic environments, integrates predicted behavioral intentions, movement trajectories, and the bicycle's own operating status data to conduct risk assessment and construct a risk perception grid, including traffic events, participants, risk levels, and movement tracking status; Early Warning Component: Identifies traffic events and their location spaces that have reached the early warning threshold based on the risk perception grid, and generates early warning feedback information.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages: First, a perception module mounted on the bicycle acquires driving perception data, including at least a visual perception device. Next, based on the visual perception data, multiple dynamic traffic participants are identified, and their future behavioral intentions and trajectories are predicted. Then, based on the dynamic intersection risks of bicycles in mixed traffic environments, the predicted behavioral intentions, trajectories, and the bicycle's own operational status data are integrated to conduct a risk assessment, constructing a risk perception grid that includes traffic events, participants, risk levels, and movement tracking status. Finally, based on the risk perception grid, traffic events whose risk levels or movement tracking status reach warning thresholds, along with their location spaces, are identified, generating warning feedback information. This solves the technical problems of insufficient risk identification and untimely warnings in existing technologies, achieving the technical effect of proactively perceiving and promptly warning about driving risks based on visual prediction. Attached Figure Description

[0007] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0008] Figure 1 A schematic diagram of the process for a vision-based bicycle riding risk warning method provided in this application embodiment; Figure 2 This is a schematic diagram of the structure of a vision-based bicycle riding risk warning system provided in an embodiment of this application.

[0009] Figure labeling: Data acquisition component 11, behavior prediction component 12, risk assessment component 13, early warning component 14. Detailed Implementation

[0010] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0011] Example 1, as Figure 1 As shown, this application provides a method for early warning of bicycle riding risks based on visual recognition, wherein the method includes: The bicycle acquires driving perception data through a sensing module mounted on the bicycle itself, the sensing module including at least a visual sensing device.

[0012] A visual sensing device is fixedly installed on the handlebars, front, or frame of a bicycle. This device is preferably a monocular or binocular camera with wide-angle imaging capabilities, used to continuously acquire image sequences of the road scene ahead while the bicycle is in motion. The sensing module automatically enters data acquisition mode after the bicycle is powered on, periodically triggering image acquisition commands and capturing real-time images of the external environment according to a set sampling frame rate. Simultaneously, it can combine auxiliary sensors such as gyroscopes and accelerometers to acquire attitude information, vibration information, and vehicle acceleration data synchronized with the image timestamps. The raw images acquired by the camera are buffered and preliminarily processed, including time synchronization, distortion correction, exposure compensation, and noise suppression, to ensure that subsequent recognition algorithms can obtain stable visual input. Finally, riding perception data containing image sequences, timestamps, and synchronized attitude information is generated.

[0013] Based on visual perception data, multiple dynamic traffic participants can be identified, and their future behavioral intentions and movement trajectories can be predicted.

[0014] Furthermore, based on visual perception data, multiple dynamic traffic participants can be identified, and their future behavioral intentions and movement trajectories can be predicted, including: A pre-trained deep learning model is used to identify visual perception data and obtain multiple dynamic traffic participants, including pedestrians, vehicles, pets and other dynamic obstacles. The identified dynamic traffic participants are continuously tracked to generate a motion trajectory sequence. Based on the motion trajectory sequence, the motion trajectory and behavioral intention classification of the target within a future preset time window are predicted. The behavioral intention classification includes crossing the road, accelerating to approach, decelerating to stop, and turning.

[0015] The system inputs visual perception data collected by the perception module into a pre-trained deep learning recognition model to perform forward inference on the image, automatically extracting key target features in the scene, thereby identifying multiple traffic participants, including pedestrians, motor vehicles, non-motor vehicles, pets, and other dynamic obstacles that may affect bicycle movement, and outputting their category, location, and bounding box information. Subsequently, the recognition results are input into a multi-target tracking algorithm based on an attention mechanism. By calculating the correlation between the appearance feature vectors, temporal position changes, and movement speed of each target, cross-frame matching and trajectory continuity maintenance are achieved, thereby generating a historical motion trajectory sequence for each target. Based on this, the motion trajectory sequence is fed as input into a behavior prediction model built on a spatiotemporal graph convolutional network. The behavior prediction model infers the movement trend of the target within a preset time window based on the dynamic displacement in the time dimension and the interaction in the spatial dimension, and outputs its predicted motion trajectory and corresponding behavioral intention classification. The behavioral intention includes action types with strong risk correlation, such as crossing the road, accelerating to approach, decelerating to stop, and turning, so that the system can grasp the future movement patterns of traffic participants in advance, providing continuous and predictable dynamic input for subsequent risk assessment.

[0016] Preferably, the classification of the target's motion trajectory and behavioral intention within a future preset time window based on the motion trajectory sequence includes: treating each traffic participant as a graph node, using the relative distance, relative speed, and interaction relationship between traffic participants as edge connection features to construct a spatiotemporal graph structure of the traffic scene; performing temporal modeling of the state of the same node at consecutive time steps in the time dimension, and characterizing the interaction influence between different traffic participants in the spatial dimension through graph convolution operations; encoding the historical motion trajectory sequence based on the spatiotemporal graph convolutional network, and outputting the target's displacement sequence and corresponding behavioral intention probability distribution within a future preset time window through a prediction decoding layer; and determining the target's behavioral intention classification result based on the maximum probability principle or a preset confidence threshold.

[0017] Furthermore, by using a pre-trained deep learning model to identify visual perception data, multiple dynamic traffic participants are obtained, including: The visual perception data is standardized, including color space conversion, size normalization, and illumination compensation. Forward inference is performed on the input image to identify multiple candidate boxes. The object attribute category within each candidate box is determined, the confidence score is calculated, and the coordinates of the candidate boxes are corrected to closely surround the target. Duplicate detection boxes for the same target are removed, and the candidate box with the highest confidence score is retained. A list of identified monitoring targets is output, with each target including category, confidence score, and bounding box coordinates.

[0018] The system standardizes visual perception data, including performing color space transformation to improve color separation of target regions, normalizing the size of input images to fit the input size requirements of the detection model, and performing illumination compensation based on local brightness distribution to reduce the impact of ambient light changes on recognition accuracy. The preprocessed image is input into a pre-trained YOLO object detection model for forward inference. The YOLO object detection model generates multiple candidate boxes on the multi-scale feature maps of the feature pyramid network and outputs the corresponding class probability distribution. The system performs class determination for each candidate box, calculates and determines its confidence score, and corrects the coordinates of the candidate boxes based on the regression output to ensure that the bounding boxes tightly surround the detected targets, thereby improving localization accuracy. To avoid the same target being detected repeatedly, non-maximum suppression is performed on overlapping candidate boxes in the same spatial region to remove duplicate detection results with low confidence and retain only the candidate boxes with the highest confidence. Finally, a list of monitored targets is output, with each target including its class type, confidence score, and bounding box coordinates for subsequent multi-target tracking and behavior prediction processing.

[0019] Preferably, the pre-trained YOLO object detection model includes: constructing a training sample set based on public traffic scene datasets and real-world bicycle driving data, wherein the training sample set contains labeled images of pedestrians, motor vehicles, non-motor vehicles, pets, and other dynamic obstacles under various road environments, lighting conditions, and viewpoints; performing data augmentation processing on the training sample set, including random cropping, scaling, brightness perturbation, and angle rotation, to improve the model's robustness to complex environments; employing a YOLO network structure including a backbone feature extraction network, a feature pyramid fusion network, and a detection head to perform multi-scale feature extraction and fusion on the input images; using the target category probability, bounding box regression error, and confidence loss as a joint loss function, iteratively updating the model parameters through a backpropagation algorithm until the loss function converges or reaches a preset number of training rounds; after training, the model parameters are solidified and deployed in the bicycle itself or an edge computing unit for online object recognition of real-time acquired visual perception data.

[0020] Furthermore, the identified dynamic traffic participants are continuously tracked to generate motion trajectory sequences, including: Extract the appearance and motion features of the tracked target from the previous frame and the newly detected target in the current frame; based on the features of all currently detected targets and existing tracking trajectories, identify the correlation based on the appearance and motion features, analyze the spatial relationship between all targets in the current frame and identify the target most similar to the current detected target in the historical trajectory to determine the matching relationship; based on the matching relationship, update the position and state of the trajectory using the new detection box to form a continuous motion trajectory sequence, wherein each trajectory ID corresponds to a data sequence.

[0021] After completing target detection in the current frame, the system first reads the corresponding trajectory identifiers from the active tracking trajectories in the previous frame and extracts the historical state data of the trajectory within the most recent preset frame window. This historical state data includes at least the target center point coordinate sequence, velocity change sequence, and corresponding appearance feature vector. Simultaneously, for each newly detected target in the current frame, appearance features are extracted based on its bounding box region, and instantaneous velocity vectors are calculated based on displacement changes between adjacent frames, forming a target feature set for the current frame. Based on this, the historical trajectory feature set and the current target feature set are mapped to the same feature space and input to the attention mechanism calculation module. By weighting the similarity between historical trajectory features and current target features, an attention weight matrix representing the degree of correlation between the detected target and each historical trajectory is obtained. This similarity considers at least appearance feature similarity, motion direction consistency, and spatial distance constraints. The system then uses this attention weight matrix to... The system determines a one-to-one correspondence between the detected target in the current frame and the historical trajectory according to the principle of maximum weight or global optimality. After determining the matching relationship, the system updates the center point coordinates, velocity status, and appearance features of the matched trajectory using the detection box position corresponding to the current frame, and appends the updated status data to the time-series data of the corresponding trajectory identifier, thereby realizing continuous trajectory updates. For historical trajectories that do not match any detected target in the current frame, the system predicts the position of the next moment based on its historical motion state, and maintains the active state of the trajectory as long as the number of consecutive unmatched frames does not exceed a preset threshold. When the number of consecutive unmatched frames exceeds the preset threshold, the corresponding trajectory is terminated. For newly detected targets in the current frame that do not match any historical trajectory, the system assigns a new trajectory identifier and initializes the trajectory state, thereby forming a continuous motion trajectory sequence indexed by the trajectory identifier. Each trajectory identifier corresponds to a state data sequence updated over time, which is used for subsequent behavior prediction and risk assessment processing.

[0022] Based on the dynamic intersection risks of bicycles in mixed traffic environments, risk assessment is carried out by integrating predicted behavioral intentions, movement trajectories, and bicycle operating status data, and a risk perception grid is constructed, which includes traffic events, participants, risk levels, and movement tracking status.

[0023] Based on the predicted behavioral intentions and future trajectories of each dynamic traffic participant, the system simultaneously collects the bicycle's own operational status data, including current speed, acceleration, heading angle change, handlebar deflection angle, and predicted trajectory. Subsequently, based on the relative position changes between each traffic participant and the bicycle, the probability of intersection of movement trends, and the degree of danger of behavioral intentions, the system calculates risk assessment results for different traffic interaction scenarios, generating a risk description for each participant that includes risk level, approach speed, potential intersection location, and time. The system uses traffic participants as nodes and traffic events or interaction relationships as edges, dynamically updating node attributes (target type, behavioral intention, trajectory trend) and edge attributes (proximity, potential conflict probability) to construct a risk perception grid based on the current road scenario. This risk perception grid includes traffic events, the status information of relevant participants, the risk level of each participant, and their movement tracking status in continuous frames, thus forming a continuously updated spatiotemporal risk expression structure, providing a unified data foundation for subsequent risk threshold judgment and early warning information generation.

[0024] Furthermore, based on the dynamic intersection risks of bicycles in mixed traffic environments, risk assessment is conducted by integrating predicted behavioral intentions, movement trajectories, and the bicycle's own operational status data, constructing a risk perception grid, including: Each traffic participant is treated as a graph node, with node attributes including target type, historical location, speed, and acceleration. Edges are used to describe the distance and interaction relationships between traffic participants, thus constructing a spatiotemporal graph of the traffic scene. By fusing the behavioral intentions, movement trajectories, and the bicycle's own operating status data, the risk status of each node is identified, and the corresponding events, participants, and trajectory status are associated. The spatiotemporal graph of the traffic scene is dynamically updated to obtain the risk perception grid.

[0025] Preferably, an adjacency matrix or adjacency list data structure is used to store the spatiotemporal graph, where each graph node corresponds to a unique identifier for a traffic participant. Node attributes are stored in vector form as target type code, historical location coordinates of the most recent N frames, velocity vector, and acceleration vector. Edge attributes are quantified by real-time calculation of Euclidean distance, relative velocity projection value, and trajectory angle between nodes to characterize the interaction intensity between traffic participants. During the temporal update process, the system performs incremental update operations on the spatiotemporal graph at a preset refresh cycle, dynamically creating nodes for newly appearing traffic participants, freezing or deleting nodes for disappearing targets, and performing other operations as needed. Nodes that persist continuously update their attribute vectors. During the risk status identification phase, the corresponding risk score is obtained by calculating the node risk feature vector with preset risk judgment rules or risk scoring functions. This score is then mapped to discrete risk levels based on the scoring interval. Simultaneously, the risk level is bound to the corresponding traffic event type, participant identifier, and trajectory status. Finally, the spatiotemporal map is mapped to a regularized risk perception grid using a spatial region index. Each grid cell contains a set of traffic events, participant status information, risk level identifier, and movement tracking status within the current region, enabling structured storage and real-time retrieval of risk information.

[0026] Furthermore, based on the dynamic intersection risks of bicycles in mixed traffic environments, risk assessment is conducted by integrating predicted behavioral intentions, movement trajectories, and the bicycle's own operational status data, including: Based on the bicycle's own operating status data, a predicted trajectory of its own movement is fitted; using the predicted trajectory of its own movement as the aiming target, aiming analysis is performed using the behavioral intention and movement trajectory to determine the aiming intersection probability, the aiming intersection point, and the intersection state description; based on the aiming intersection probability, the aiming intersection point, and the intersection state description, a risk assessment is performed to generate a risk state assessment level.

[0027] The system uses a kinematic model to fit the bicycle's predicted trajectory within a preset time window based on its current speed, acceleration, handlebar deflection angle, yaw angle changes, and historical trajectory. This trajectory represents the bicycle's possible short-term path. Subsequently, this predicted trajectory is used as the target trajectory and correlated with the future paths derived from the behavioral intentions and predicted trajectories of various traffic participants. Specifically, the predicted trajectories of the bicycle and those of traffic participants are mapped to the same ground coordinate system, and the minimum Euclidean distance sequence between each pair of trajectories is calculated within the preset time window. Based on this minimum distance sequence and considering the uncertainty parameters of the traffic participants' trajectory predictions, a Gaussian probability model or Monte Carlo sampling method is used to calculate the trajectory occurrence space. The overlapping probability values ​​are used to obtain the target intersection probability of possible trajectory intersection, and the spatial location of the corresponding probability peak time is determined as the target intersection point. At the same time, based on the approach direction, speed difference, and behavioral intention category of the two trajectories, a corresponding intersection state description is generated. The intersection state description includes at least the expected contact time, relative motion direction, and potential conflict type. On this basis, the target intersection probability, target intersection point location, and intersection state description are weighted according to a preset risk assessment rule. The weighting rule is set based on the behavioral intention risk coefficient, intersection distance threshold, and approach time threshold. When the comprehensive risk value falls into different threshold ranges, the corresponding risk state assessment level is output to indicate the potential danger level of each traffic participant relative to the bicycle, providing a basis for subsequent early warning triggering.

[0028] Furthermore, using the predicted trajectory of its own motion as the aiming target, and performing aiming analysis using the behavioral intent and motion trajectory, includes: The predicted trajectory of its own motion is regarded as the target trajectory, and the predicted trajectory of the traffic participant is regarded as the trajectory to be compared. The minimum spatial distance between the target trajectory and the trajectory to be compared at any time point within a preset future time window is calculated. Based on the sequence of minimum spatial distances and combined with the motion uncertainty of both parties, the probability of spatial overlap between one or more trajectories is calculated to obtain the aiming intersection probability, aiming intersection point, and intersection state description.

[0029] The system considers its own predicted trajectory, generated based on bicycle running status data, as the target trajectory, and the predicted trajectories of each traffic participant, calculated based on their behavioral intentions and future movement trends, as the trajectories to be compared. Subsequently, both the self-predicted trajectory and the predicted trajectories of each traffic participant are uniformly represented as a discrete trajectory sequence generated at a fixed time step. Within a preset future time window, the spatial positions of the two trajectories are compared moment-by-moment, calculating the relative spatial distance between them at corresponding time points. The minimum distance is extracted from the relative spatial distances of all time points within the time window to characterize the closest approach between the two trajectories in future movement. Simultaneously, based on the speed change amplitude, direction fluctuations, and historical prediction errors obtained during trajectory prediction, motion uncertainty descriptions (including speed fluctuations, etc.) are constructed for both the self-predicted trajectory and the traffic participant trajectories. The uncertainty parameters (such as changes in direction and acceleration disturbances) are integrated to form a comprehensive uncertainty parameter characterizing the reliability of trajectory prediction. Based on this, the minimum spatial distance and the comprehensive uncertainty parameter are input into a preset conflict probability calculation model to quantify the possibility of spatial overlap between the target trajectory and the trajectory to be compared within a future time window, thus obtaining the corresponding aiming intersection probability. When the aiming intersection probability reaches or exceeds a preset threshold, a potential trajectory intersection risk is determined. The time point at which the minimum spatial distance is generated is taken as the aiming intersection time, and the corresponding spatial location is taken as the aiming intersection point. At the same time, the relative motion direction, approach speed, and behavioral intention type of both parties at that time point are combined to generate an intersection state description characterizing the interaction pattern, thereby providing an executable judgment basis for subsequent risk level assessment and early warning triggering.

[0030] The optimized uncertainty parameters include: while generating the predicted trajectory of traffic participants, recording the changes in velocity, heading angle, and acceleration of the predicted trajectory within consecutive prediction time steps, and combining the deviation between the actual motion trajectory and the predicted trajectory of the corresponding target in several historical frames, statistically analyzing the deviation to obtain error characteristic parameters reflecting prediction stability; based on the error characteristic parameters, calculating the fluctuation degree of the target in the position, velocity, and direction dimensions respectively, and fusing the fluctuation degree of each dimension according to preset weights to form motion uncertainty parameters used to characterize the reliability of the target motion prediction; wherein, the motion uncertainty parameters are dynamically adjusted as the trajectory prediction results are updated, and when the target motion state changes abruptly or the prediction deviation continues to increase, the uncertainty weight is increased accordingly, so that the subsequent aiming analysis and risk assessment processes can adaptively reflect the changes in trajectory prediction reliability.

[0031] Furthermore, it also includes: When the stated behavioral intent is a turning intent, based on the inherent physical dimensions of the vehicle type, the high-risk sweep area generated by the inner wheel difference effect during the turning process is dynamically calculated; it is determined whether the predicted trajectory of the vehicle's own motion will enter the high-risk sweep area within a future time window; if it is determined to have entered, the highest risk level is assigned; if it has not entered but is in the edge buffer zone of the high-risk sweep area, a medium risk level is assigned based on distance and approach time; wherein, the dynamic calculation of the high-risk sweep area generated by the inner wheel difference effect during the turning process includes: performing fine-grained analysis of the vehicle, based on the fine-grained classification results, The inherent physical attributes of the corresponding vehicle type are obtained from a predefined vehicle physical parameter library, including at least typical length, width, and wheelbase. A multi-target tracking algorithm is used to continuously track the classified vehicle targets, outputting dynamic state parameters in real time. These dynamic state parameters include: center point coordinates, instantaneous speed, real-time heading angle, and bounding box size based on image pixels. Based on the vehicle's wheelbase and real-time heading angle, the theoretical trajectory of the rear wheels under turning conditions is dynamically calculated. Using the theoretical trajectory of the rear wheels as a reference, and combined with the vehicle's width, a safety distance threshold margin is extended outwards to calculate the high-risk sweep area during turning.

[0032] When a traffic participant's intention is determined to be a turning action, the system dynamically calculates the high-risk sweep area generated by the inner wheel difference effect during the turning process based on the inherent physical dimensions of the vehicle type. It then uses the relative relationship between this area and the bicycle's predicted trajectory to determine the risk level. Specifically, the system first performs a fine-grained analysis of the vehicle target, extracting the inherent physical attributes of the corresponding vehicle type from a predefined vehicle physical parameter library based on the identification results. These attributes include typical length, width, and wheelbase. Subsequently, a multi-target tracking algorithm continuously tracks the vehicle target, updating its center point coordinates, instantaneous speed, real-time heading angle, and dynamic state parameters such as the bounding box size based on image pixels. Based on this, the system calculates the theoretical trajectory of the rear wheels during the turning process using the vehicle's wheelbase and real-time heading angle. Using this theoretical rear wheel trajectory as a benchmark, combined with a safety distance threshold margin set by extending the vehicle width outwards, the system identifies the high-risk sweep area that the vehicle may sweep during the turning process. After obtaining the high-risk sweeping area, the system determines whether the bicycle's predicted trajectory will enter the area within a preset time window in the future: if the predicted trajectory will enter the high-risk sweeping area, it will be directly assigned the highest risk level; if it does not enter but is located in the buffer zone at the outer edge of the sweeping area, it will be assigned a medium risk level based on the minimum distance between the bicycle and the boundary of the area and the expected approach time.

[0033] Furthermore, constructing a risk awareness grid also includes: The communication module connects to the collaborative vehicle connectivity interface, including the vehicle network, traffic data network, and vehicle broadcasting; the collaborative vehicle connectivity interface is used to obtain the associated traffic status and extract the associated traffic participants to construct a vehicle connectivity perception network; the vehicle connectivity perception network is spatiotemporally spliced ​​with its own risk perception grid according to the spatiotemporal relationship to construct the risk perception grid of vehicle connectivity collaboration.

[0034] The system establishes a connection with an external collaborative vehicle-to-everything (V2X) interface through a communication module configured on the bicycle itself. This interface includes a vehicle-to-everything (V2X) platform, a regional traffic data network, and road vehicle broadcast channels, used to acquire real-time traffic status information released by road infrastructure or surrounding vehicles. After accessing the V2X interface, the system extracts traffic participant data and traffic event data related to the current driving environment, forming a V2X perception network that includes information such as vehicle speed, direction, location, and intent. Subsequently, the system performs spatiotemporal alignment between the V2X perception network and its own vision-based risk perception grid based on timestamps and geospatial coordinates. It matches, merges, and completes participant nodes from the V2X side with nodes from the visual recognition side, ultimately generating a collaborative V2X risk perception grid. This allows the system to simultaneously utilize its own perception and external traffic infrastructure information to achieve a more comprehensive and reliable risk characterization and dynamic update of the road environment.

[0035] Based on the risk perception grid, traffic events whose risk level or movement tracking status reaches the warning threshold and their location space are identified, and warning feedback information is generated.

[0036] After completing the dynamic construction of the risk perception grid, the system continuously monitors the risk level, trajectory trend, and movement tracking status of each node in the grid. When the risk level of any traffic participant exceeds the preset risk threshold, or when its movement tracking status shows drastic changes (including sudden acceleration, sharp turns, rapid approach, etc.) and meets the warning triggering conditions, the system determines the spatial area and specific direction of the corresponding traffic event through the location information of the grid node and related edge attributes. Subsequently, based on the level of risk, the type of event, and the predicted convergence time, the system generates warning feedback information that includes the risk level, source of danger, directional indication, and suggested avoidance actions. This information is then provided to the cyclist in real time through the bicycle terminal's audio-visual prompts, vibration feedback module, or display interface, enabling the cyclist to take countermeasures such as slowing down, avoiding, or adjusting the route before potential risks occur, thereby improving cycling safety.

[0037] In summary, the embodiments of this application have at least the following technical effects: First, a perception module mounted on the bicycle acquires driving perception data, including at least a visual perception device. Next, based on the visual perception data, multiple dynamic traffic participants are identified, and their future behavioral intentions and trajectories are predicted. Then, based on the dynamic intersection risks of bicycles in mixed traffic environments, the predicted behavioral intentions, trajectories, and the bicycle's own operational status data are integrated to conduct a risk assessment, constructing a risk perception grid that includes traffic events, participants, risk levels, and movement tracking status. Finally, based on the risk perception grid, traffic events whose risk levels or movement tracking status reach warning thresholds, along with their location spaces, are identified, generating warning feedback information. This solves the technical problems of insufficient risk identification and untimely warnings in existing technologies, achieving the technical effect of proactively perceiving and promptly warning about driving risks based on visual prediction.

[0038] Example 2, based on the same inventive concept as the visual recognition-based bicycle riding risk warning method in the foregoing examples, such as... Figure 2 As shown, this application provides a vision-based bicycle riding risk warning system, wherein the system includes: Data acquisition component 11: Acquires driving perception data through a sensing module mounted on the bicycle body, the sensing module including at least a visual perception device; Behavior prediction component 12: Identifies multiple dynamic traffic participants based on visual perception data and predicts future behavioral intentions and movement trajectories; Risk assessment component 13: Based on the dynamic intersection risk of bicycles in mixed traffic environments, integrates predicted behavioral intentions, movement trajectories, and the bicycle's own operating status data to conduct risk assessment and construct a risk perception grid, including traffic events, participants, risk levels, and movement tracking status; Early warning component 14: Identifies traffic events and their location spaces that have reached the early warning threshold based on the risk perception grid, and generates early warning feedback information.

[0039] Furthermore, the behavior prediction component 12 is used to perform the following methods: A pre-trained deep learning model is used to identify visual perception data and obtain multiple dynamic traffic participants, including pedestrians, vehicles, pets and other dynamic obstacles. The identified dynamic traffic participants are continuously tracked to generate a motion trajectory sequence. Based on the motion trajectory sequence, the motion trajectory and behavioral intention classification of the target within a future preset time window are predicted. The behavioral intention classification includes crossing the road, accelerating to approach, decelerating to stop, and turning.

[0040] Furthermore, the behavior prediction component 12 is used to perform the following methods: The visual perception data is standardized, including color space conversion, size normalization, and illumination compensation. Forward inference is performed on the input image to identify multiple candidate boxes. The object attribute category within each candidate box is determined, the confidence score is calculated, and the coordinates of the candidate boxes are corrected to closely surround the target. Duplicate detection boxes for the same target are removed, and the candidate box with the highest confidence score is retained. A list of identified monitoring targets is output, with each target including category, confidence score, and bounding box coordinates.

[0041] Furthermore, the behavior prediction component 12 is used to perform the following methods: Extract the appearance and motion features of the tracked target from the previous frame and the newly detected target in the current frame; based on the features of all currently detected targets and existing tracking trajectories, identify the correlation based on the appearance and motion features, analyze the spatial relationship between all targets in the current frame and identify the target most similar to the current detected target in the historical trajectory to determine the matching relationship; based on the matching relationship, update the position and state of the trajectory using the new detection box to form a continuous motion trajectory sequence, wherein each trajectory ID corresponds to a data sequence.

[0042] Furthermore, the risk assessment component 13 is used to perform the following methods: Each traffic participant is treated as a graph node, with node attributes including target type, historical location, speed, and acceleration. Edges are used to describe the distance and interaction relationships between traffic participants, thus constructing a spatiotemporal graph of the traffic scene. By fusing the behavioral intentions, movement trajectories, and the bicycle's own operating status data, the risk status of each node is identified, and the corresponding events, participants, and trajectory status are associated. The spatiotemporal graph of the traffic scene is dynamically updated to obtain the risk perception grid.

[0043] Furthermore, the risk assessment component 13 is used to perform the following methods: Based on the bicycle's own operating status data, a predicted trajectory of its own movement is fitted; using the predicted trajectory of its own movement as the aiming target, aiming analysis is performed using the behavioral intention and movement trajectory to determine the aiming intersection probability, the aiming intersection point, and the intersection state description; based on the aiming intersection probability, the aiming intersection point, and the intersection state description, a risk assessment is performed to generate a risk state assessment level.

[0044] Furthermore, the risk assessment component 13 is used to perform the following methods: The predicted trajectory of its own motion is regarded as the target trajectory, and the predicted trajectory of the traffic participant is regarded as the trajectory to be compared. The minimum spatial distance between the target trajectory and the trajectory to be compared at any time point within a preset future time window is calculated. Based on the sequence of minimum spatial distances and combined with the motion uncertainty of both parties, the probability of spatial overlap between one or more trajectories is calculated to obtain the aiming intersection probability, aiming intersection point, and intersection state description.

[0045] Furthermore, the risk assessment component 13 is used to perform the following methods: When the stated behavioral intent is a turning intent, based on the inherent physical dimensions of the vehicle type, the high-risk sweep area generated by the inner wheel difference effect during the turning process is dynamically calculated; it is determined whether the predicted trajectory of the vehicle's own motion will enter the high-risk sweep area within a future time window; if it is determined to have entered, the highest risk level is assigned; if it has not entered but is in the edge buffer zone of the high-risk sweep area, a medium risk level is assigned based on distance and approach time; wherein, the dynamic calculation of the high-risk sweep area generated by the inner wheel difference effect during the turning process includes: performing fine-grained analysis of the vehicle, based on the fine-grained classification results, The inherent physical attributes of the corresponding vehicle type are obtained from a predefined vehicle physical parameter library, including at least typical length, width, and wheelbase. A multi-target tracking algorithm is used to continuously track the classified vehicle targets, outputting dynamic state parameters in real time. These dynamic state parameters include: center point coordinates, instantaneous speed, real-time heading angle, and bounding box size based on image pixels. Based on the vehicle's wheelbase and real-time heading angle, the theoretical trajectory of the rear wheels under turning conditions is dynamically calculated. Using the theoretical trajectory of the rear wheels as a reference, and combined with the vehicle's width, a safety distance threshold margin is extended outwards to calculate the high-risk sweep area during turning.

[0046] Furthermore, the risk assessment component 13 is used to perform the following methods: The communication module connects to the collaborative vehicle connectivity interface, including the vehicle network, traffic data network, and vehicle broadcasting; the collaborative vehicle connectivity interface is used to obtain the associated traffic status and extract the associated traffic participants to construct a vehicle connectivity perception network; the vehicle connectivity perception network is spatiotemporally spliced ​​with its own risk perception grid according to the spatiotemporal relationship to construct the risk perception grid of vehicle connectivity collaboration.

[0047] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for early warning of bicycle riding risks based on visual recognition, characterized in that, The method includes: The bicycle acquires driving perception data through a sensing module mounted on the bicycle body, wherein the sensing module includes at least a visual sensing device. Identify multiple dynamic traffic participants based on visual perception data and predict their future behavioral intentions and movement trajectories; Based on the dynamic intersection risks of bicycles in mixed traffic environments, risk assessment is carried out by integrating predicted behavioral intentions, movement trajectories and bicycle operating status data, and a risk perception grid is constructed, which includes traffic events, participants, risk levels and movement tracking status. Based on the risk perception grid, traffic events whose risk level or movement tracking status reaches the warning threshold and their location space are identified, and warning feedback information is generated.

2. The method for early warning of bicycle riding risks based on visual recognition according to claim 1, characterized in that, Based on visual perception data, multiple dynamic traffic participants are identified, and their future behavioral intentions and movement trajectories are predicted, including: By using a pre-trained deep learning model, visual perception data is identified to obtain multiple dynamic traffic participants, including pedestrians, vehicles, pets and other dynamic obstacles. Continuously track identified dynamic traffic participants and generate motion trajectory sequences; Based on the motion trajectory sequence, the target's motion trajectory and behavioral intention classification within a future preset time window are predicted. The behavioral intention classification includes crossing the road, accelerating to approach, decelerating to stop, and turning.

3. The method for early warning of bicycle riding risks based on visual recognition according to claim 2, characterized in that, By using a pre-trained deep learning model to identify visual perception data, multiple dynamic traffic participants are obtained, including: The visual perception data is standardized, including color space conversion, size normalization, and illumination compensation. Perform forward reasoning on the input image to identify multiple candidate boxes; Determine the object attribute category within each candidate box, determine the confidence score, and adjust the coordinates of the candidate boxes to tightly surround the target. Remove duplicate detection boxes for the same target, retain the candidate box with the highest confidence, and output a list of identified monitoring targets. Each target includes its category, confidence, and bounding box coordinates.

4. The method for early warning of bicycle riding risks based on visual recognition according to claim 3, characterized in that, Continuous tracking of identified dynamic traffic participants generates motion trajectory sequences, including: Extract the appearance and motion features of the tracked target from the previous frame and the newly detected target in the current frame; Based on the features of all currently detected targets and existing tracking trajectories, the correlation degree is identified based on the appearance and motion features. The spatial relationship between all targets in the current frame is analyzed and the target most similar to the current detected target in the historical trajectory is identified to determine the matching relationship. Based on the matching relationship, the position and state of the trajectory are updated using the new detection box to form a continuous motion trajectory sequence, wherein each trajectory ID corresponds to a data sequence.

5. The method for early warning of bicycle riding risks based on visual recognition according to claim 1, characterized in that, Risk assessment is conducted by integrating predicted behavioral intentions, movement trajectories, and the bicycle's own operational status data to construct a risk perception grid, including: Each traffic participant is treated as a graph node, with node attributes including target type, historical location, speed, and acceleration. Edges are used to describe the distance and interaction relationships between traffic participants, thus constructing a spatiotemporal graph of the traffic scene. By integrating the behavioral intent, movement trajectory, and bicycle's own operating status data, the risk status of each node is identified, and the corresponding events, participants, and trajectory status are associated to dynamically update the spatiotemporal map of the traffic scene, thereby obtaining the risk perception grid.

6. The method for early warning of bicycle riding risks based on visual recognition according to claim 1, characterized in that, Risk assessment is conducted by integrating predicted behavioral intentions, movement trajectories, and the bicycle's own operational status data, including: Based on the bicycle's own operating status data, fit its own predicted trajectory; Using the predicted trajectory of its own motion as the aiming target, the aiming analysis is performed using the behavioral intention and motion trajectory to determine the aiming intersection probability, the aiming intersection point, and the intersection state description. Based on the target intersection probability, target intersection point, and intersection state description, a risk assessment is performed, and a risk state assessment level is generated.

7. The method for early warning of bicycle riding risks based on visual recognition according to claim 6, characterized in that, Using the predicted trajectory of its own motion as the aiming target, aiming analysis is performed using the behavioral intent and motion trajectory, including: The predicted trajectory of its own motion is regarded as the target trajectory, and the predicted trajectory of the traffic participants is regarded as the trajectory to be compared. Calculate the minimum spatial distance between the target trajectory and the trajectory to be compared at any point in time within a future preset time window; Based on the minimum spatial distance sequence and combined with the motion uncertainty of both parties, the probability of spatial overlap between one or more trajectories is calculated to obtain the aiming intersection probability, aiming intersection point, and intersection state description.

8. The method for early warning of bicycle riding risks based on visual recognition according to claim 6, characterized in that, Also includes: When the behavioral intent is a turning intent, the high-risk sweeping area generated by the inner wheel difference effect of the corresponding vehicle during the turning process is dynamically calculated based on the inherent physical size of the vehicle type. Determine whether the predicted trajectory of its own motion will enter the high-risk sweeping area within a future time window; If entry is deemed appropriate, the highest risk level will be assigned. If the area is not entered but is located in the buffer zone on the edge of a high-risk sweeping area, a medium risk level will be assigned based on the distance and time of approach. The dynamic calculation of the high-risk sweep area caused by the inner wheel difference effect during vehicle turning includes: Fine-grained analysis is performed on the vehicle. Based on the fine-grained classification results, the inherent physical attributes of the corresponding vehicle type are obtained from the predefined vehicle physical parameter library, including at least typical length, width, and wheelbase. Using a multi-target tracking algorithm, classified vehicle targets are continuously tracked, and dynamic state parameters are output in real time. The dynamic state parameters include: center point position coordinates, instantaneous speed, real-time heading angle, and bounding box size based on image pixels. Based on the vehicle's wheelbase and real-time heading angle, the theoretical trajectory of the rear wheels under turning conditions is dynamically calculated. Based on the theoretical trajectory of the rear wheels and combined with the width of the vehicle, the safe distance threshold margin is extended outward to calculate the high-risk sweeping area during turning.

9. The method for early warning of bicycle riding risks based on visual recognition according to claim 1, characterized in that, Building a risk awareness grid also includes: The communication module connects to the collaborative vehicle connectivity interface, including vehicle networking, traffic data network, and vehicle broadcasting. The associated traffic status is obtained through the collaborative vehicle connectivity interface, and the associated traffic participants are extracted to construct a vehicle connectivity perception network. The vehicle-to-everything (V2X) perception network is spatiotemporally spliced ​​with its own risk perception grid according to the spatiotemporal relationship to construct the V2X collaborative risk perception grid.

10. A bicycle riding risk warning system based on visual recognition, characterized in that, The system is used to implement the vision-based bicycle riding risk warning method according to any one of claims 1-9, the system comprising: Data acquisition component: Acquires riding perception data through a sensing module mounted on the bicycle body, wherein the sensing module includes at least a visual sensing device; Behavior prediction component: Identifies multiple dynamic traffic participants based on visual perception data and predicts their future behavioral intentions and movement trajectories; Risk assessment component: Based on the dynamic intersection risk of bicycles in mixed traffic environments, risk assessment is carried out by integrating predicted behavioral intentions, movement trajectories and bicycle operation status data, and a risk perception grid is constructed, which includes traffic events, participants, risk levels and movement tracking status. Early warning component: Based on the risk perception grid, identify traffic events and their location spaces that have reached the early warning threshold in terms of risk level or movement tracking status, and generate early warning feedback information.

Images