An intelligent early warning method applied to real-time evaluation of non-motor vehicle collision risk

Abstract

This invention discloses an intelligent early warning method for real-time assessment of non-motorized vehicle collision risks. This method integrates YOLOv8 target detection, binocular stereo vision, and deep learning direction recognition technologies to construct a complete process from environmental perception to risk decision-making. The method first identifies non-motorized vehicle targets using YOLOv8, acquires three-dimensional distance using binocular vision, calculates velocity using continuous frame displacement, and determines the driving direction using a CNN model, thereby calculating key risk parameters such as relative speed, distance, and angle. Subsequently, multi-source parameters are input into a risk quantification model based on safety potential energy field theory, achieving nonlinear fusion and outputting a normalized safety coefficient to characterize the real-time collision risk level. Based on this safety coefficient, this invention drives an interactive module integrated with a smart helmet to issue warnings to cyclists through spatialized tactile vibration and contextualized voice, allowing them to perceive risks without shifting their gaze, effectively improving cycling safety and reducing the risk of collision accidents.

Description

Technical Field

[0001] This invention relates to an intelligent early warning method for real-time assessment of non-motorized vehicle collision risks, belonging to the interdisciplinary fields of intelligent transportation, active safety early warning, computer vision, and embedded artificial intelligence. Background Technology

[0002] With the diversification of urban travel demands, the proportion of non-motorized vehicles, represented by electric bicycles and shared bicycles, in the urban transportation system has increased dramatically. At the same time, the number of road traffic accidents involving non-motorized vehicles and the number of injuries and fatalities have also shown a significant upward trend, becoming a major challenge to urban traffic safety. However, current technologies have the following main shortcomings: 1) Limitations in perception capabilities: Most systems rely on a single sensor. For example, using only a monocular camera cannot accurately measure distances, or using only radar makes it difficult to identify target types and attitudes, and cannot fully perceive the status of non-motorized vehicles.

[0003] 2) Innovation in the risk model: In normal driving, rear-end collisions or close-range weaving scenarios with a relative angle of less than 10° are more likely to occur. Conversely, at larger angles, normal drivers will issue warnings or maintain a greater distance, making it relatively safer.

[0004] 3) The warning methods are highly disruptive: Existing warnings mostly rely on buzzers or screen flashing, which can easily distract riders and even cause operational errors. Furthermore, frequent false alarms can lead to "alarm fatigue" among users. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings and deficiencies of existing technologies by proposing an intelligent early warning method for real-time assessment of non-motorized vehicle collision risks. This method identifies non-motorized vehicle targets using YOLOv8, acquires three-dimensional distance using binocular vision, calculates velocity by combining continuous frame displacement, and uses a CNN model to determine the driving direction, forming key risk parameters. The method then inputs a risk quantification model, achieves nonlinear fusion, and outputs a normalized safety coefficient, thereby triggering personalized, low-interference early warnings. This effectively improves riding safety and reduces the risk of collisions without requiring riders to shift their gaze.

[0006] The technical solution adopted by this invention to solve its technical problem is: an intelligent early warning method for real-time assessment of non-motorized vehicle collision risk, the method comprising the following steps: Step S1: Synchronous Acquisition and Preprocessing of Multi-Source Heterogeneous Data: The goal of this stage is to provide high-quality, spatiotemporally aligned input data for subsequent deep learning algorithms and geometric calculations. Specifically, this includes: 1a) Hardware synchronization: Hardware trigger signals or high-precision software synchronization mechanisms are used to ensure that the left and right imaging units of the binocular camera complete the exposure within the same microsecond time window, so as to fundamentally eliminate motion blur and stereo matching errors caused by the difference in sampling time.

[0007] 1b) Stereo calibration and epipolar constraint: using the camera intrinsic parameter matrix obtained from offline calibration ( The original left and right images are stereo-corrected using a rotation matrix R and a translation vector T that describe the relative positions of the left and right cameras. This ensures that the corrected image pairs satisfy the epipolar constraint—that is, the projection of any point in space onto the left and right images lies on the same horizontal scan line. This transformation simplifies the subsequent two-dimensional search problem for stereo matching into a one-dimensional search, greatly improving matching efficiency and accuracy, and laying the foundation for accurate three-dimensional reconstruction.

[0008] Step S2: Multi-target detection and continuous tracking stage based on deep learning. This stage aims to extract the identity, location, and preliminary motion information of each non-motorized vehicle target from the preprocessed image sequence in real time and stably. Specifically, it includes: 2a) Object Detection and Classification: A lightweight variant of the YOLOv8 series (preferably YOLOv8n) is used as the core detection network. This model is pre-trained on large datasets such as COCO and fine-tuned on a non-motorized vehicle-specific dataset containing categories such as "bicycle," "electric bicycle," and "tricycle" to adapt to the target scene. For each input frame, the model outputs a series of bounding boxes, each containing normalized center coordinates (x, y), width w, height h, target presence confidence conf, and specific class label cls.

[0009] 2b) Multi-target tracking and trajectory management: To address the independence issue in frame-by-frame detection and acquire continuous motion information, the ByteTrack algorithm is used for multi-target tracking. Its core process is as follows: 1) Prediction: For each existing trajectory, a Kalman filter based on a constant velocity (CV) or constant velocity turning (CTRV) model is used to predict its bounding box position in the current frame. 2) Association: A similarity matrix is ​​constructed, comprehensively calculating the motion similarity (IoU) and appearance similarity (cosine distance of the 128-dimensional feature vectors extracted using a lightweight ReID network) between the detected and predicted boxes. The Hungarian algorithm is used for optimal matching, and low-confidence detected boxes are innovatively included in the association process, effectively reducing trajectory loss due to occlusion. 3) Management: New trajectories are created for unmatched high-quality detected boxes; for continuously lost-matching trajectories, a lifetime counter is set, and tracking is terminated when the timeout period expires and the target is deemed to have left the field of view. Each successfully tracked target is assigned a unique trajectory ID, and its two-dimensional image coordinates over time are recorded.

[0010] Step S3: Precise Calculation of Multi-Dimensional Interaction State Parameters. Based on the output of S2, this stage calculates in parallel the three core physical quantities that determine the collision risk between the main vehicle and each neighboring vehicle: relative speed, relative distance, and relative angle. Specifically, this includes: 3a) Relative velocity calculation: For a stable tracking trajectory, take the three-dimensional position of the current frame t and the previous frame t-1 in the world coordinate system. (Obtained through subsequent binocular ranging). Calculate the displacement vector. And divide by the precise frame interval time. The instantaneous velocity vector of the target in the two-dimensional plane is obtained (derived from the high-precision timestamp difference). The instantaneous velocity of the target is The relative velocity scalar between vehicle A and adjacent vehicle B is: The magnitude of its velocity vector difference: The component of the relative velocity along the line connecting the two points is called the relative approach velocity. It is a key indicator for assessing the urgency of a collision and can be obtained through vector dot product calculation.

[0011] 3b) Relative Distance Measurement (Based on Binocular Stereo Vision): This is crucial for achieving high-precision spatial perception. Based on epipolar correction, target-level sparse stereo matching is performed: For each tracked and detected non-motorized vehicle target, robust ORB feature points are extracted within its bounding box. ORB features are scale and rotation invariant, resulting in high computational efficiency. A one-dimensional search is performed on the corresponding epipolar (horizontal) line in the right-hand image to find the optimal matching point for each ORB feature point, using Hamming distance as the similarity metric. The disparity of all successfully matched point pairs within the target region is collected. ,in These refer to the horizontal coordinates of the same pixel on the left and right cameras, respectively. Median filtering is used to remove mismatched outliers, thus obtaining the representative disparity of the target. According to the binocular vision triangulation formula: ; Where B is the baseline distance and f is the corrected equivalent focal length (in pixels).

[0012] Calculate the target's 3D coordinates (X, Y, Z) relative to the left camera coordinate system. Obtain the 3D coordinates of the main vehicle A and the adjacent vehicle B (usually taken as the center point of their detection box bottom or centroid), and the Euclidean distance between them. This refers to the precise real-time relative distance.

[0013] 3c) Driving Direction Classification and Relative Angle Calculation (Key Innovation): To quantify the impact of driving direction on risk, this invention trains a dedicated lightweight CNN model (e.g., based on the MobileNetV2 architecture). The model's input is a cropped image of a non-motorized vehicle target (resized to 64x64), and the output layer is designed with 7 neurons, representing seven discrete driving direction angles (relative to the horizontal axis of the image coordinate system): 0°, 5°, 10°, 15°, 20°, 25°, and 30°. The model is trained on a large dataset with manually labeled directions using a cross-entropy loss function with class weights to balance the difference in the number of samples at different angles.

[0014] This invention inputs cropped images of the main vehicle A and the adjacent vehicle B into the model to obtain their most likely driving direction angles θA and θB. The absolute value of the difference between the two is... That is, relative angle.

[0015] Based on statistical analysis of a large amount of non-motorized vehicle accident data, this invention provides a revolutionary redefinition of the risk level of relative angles: High-risk area (small angle): This angle corresponds to when two vehicles are traveling in the same or opposite directions. At this angle, there is a high probability of a head-on collision or a minor scrape with the driver. Therefore, the chance of danger is actually higher within this angle range.

[0016] Medium danger zone (medium angle): This section corresponds to a certain angular deviation between the two vehicles. Risk mechanism: There is a possibility of intersection, such as a collision between a turning vehicle and a straight-going vehicle at an intersection. However, since the angle is already relatively obvious, both drivers have relatively more time to observe and anticipate the intentions, and the avoidance space and reaction time are better than in small-angle scenarios.

[0017] Low-risk zone (wide angle): This section corresponds to a significant difference in the direction of the two vehicles. Risk mechanism: This usually occurs when there is a clear intention to turn or change lanes. Due to the large angle, the projected components of the motion vectors of both vehicles in the potential collision direction are relatively small, and these actions are often accompanied by significant changes in spatial position, giving the other party more time to warn and avoid collisions.

[0018] Step S4: Comprehensive Risk Quantification and Decision-Making Stage Based on the Safety Potential Field. This stage is the core of the method, aiming to construct a risk quantification model with clear physical meaning and adjustable parameters, and to integrate the heterogeneous parameters obtained in stage S3 ( These are combined into a normalized comprehensive safety factor (SF). This model, known as the "Safety Potential Field Model," includes the following sub-modules: Virtual quality calculation module (S41): Definition ; in, To calculate the average weight (e.g., for electric bicycles including the rider, take 80kg); Here, represents the type coefficient, reflecting the inherent risk of different vehicle types (electric bicycles: 1.0-1.2, bicycles: 0.8-1.0); f(v) is the velocity effect function. Preferred parameters This function amplifies the contribution of speed to the "dangerous equivalent" in a high-power form.

[0019] Basic safety factor modules (S42 & S43): Simulate the attenuation of the risk field with distance. , . These are preset parameters.

[0020] Directional safety factor module (S44, key innovation): To reflect the new understanding that "the smaller the angle, the higher the risk", a directional risk function is designed. ; Where θ is in radians. A, B, and C are preset constants. The directional safety factor is... Calculations show that as θ increases from 0° to 30°, It will get smaller.

[0021] Speed ​​safety factor module (S45): Uses a piecewise linear function to measure the relative approach speed. Perform highly nonlinear mapping: ; ; ; ; Quality and safety factor module (S46): ); Preferred To prevent extreme values ​​from dominating.

[0022] f) Integrated safety factor fusion module (S47): performs a weighted summation of the above four sub-factors: The weighting allocation reflects the priority of risk factors: ; All weights sum to 1.

[0023] g) Risk Level Determination Module (S48): Maps continuous SF to three levels of discrete risk: Safe (Green): SF ≥ 0.5; Alert (Yellow): 0.28 ≤ SF < 0.5; Danger (Red): SF < 0.28.

[0024] Step S5: Personalized, Low-Interference Proactive Safety Warning Output Stage: When the system determines that the risk level has reached the warning threshold, it enters the warning execution process: Warning information coding: Generate a structured warning instruction package, including: risk level (yellow / red), risk direction (azimuth angle with the main vehicle as the origin, such as 30° to the left front), risk distance level (near / medium / far), and risk dominant factors (such as "high-speed approach", "small-angle rear-end collision").

[0025] This invention employs a "ghost peek" risk warning method based on infrared thermal sensing and area determination, comprising: 1. Thermal imaging image acquisition and three-zone division: a) Hardware deployment and data source Sensing unit: An uncooled infrared focal plane array (IRFPA) thermal imaging camera deployed at the front of the vehicle (such as behind the windshield or at the grille).

[0026] The simulation experimental parameters of this invention include the following: Operating wavelength: 8-14 μm (long-wave infrared, suitable for human body temperature measurement).

[0027] Resolution: ≥ 320 × 240 pixels (supports basic analysis).

[0028] Frame rate: ≥ 25 Hz (meets real-time requirements).

[0029] Field of view (FOV): Horizontal ≥ 60°, vertical ≥ 40° (covering the risk areas in front and to the sides).

[0030] Output: Real-time single-mode infrared thermal imaging video stream, each frame is a grayscale or pseudo-color image, and the pixel value corresponds to the temperature or radiation intensity.

[0031] b) Key Region Delineation Algorithm For each input frame Perform static space partitioning: Get the image dimensions: Let the image width be W and the height be H.

[0032] Divide the sample into three equal parts: Left region ; Central region ; Right region ; The three areas are equal in size ; Design rationale: To compress the high-risk concern area (directly in front and immediately to the side) into the medium zone. This area corresponds to the "core conflict zone" into which the vehicle is about to enter. The left and right areas are used for situational awareness, but not as triggers for Level 1 warnings, thus filtering out interference from irrelevant targets at a distance from the sides.

[0033] High-temperature target detection in the middle area: a) Detection logic: Only detects the central region pixel subset Processing: Background temperature estimation: Calculation Median temperature of all pixels This serves as a robust estimate of the current frame's ambient background temperature. Then, a dynamic threshold is set. : ; in, A preset temperature rise threshold, typically 2℃-3℃, is used to distinguish between living organisms and the environment.

[0034] Generate binary image (High temperature), otherwise 0. Apply morphological opening operation (erosion followed by dilation) to B(x,y) to eliminate small thermal noise points. Mark all connected components (high temperature patches). .

[0035] b) Target filtering and labeling: For each connected component Calculation feature: pixel area (Patch size). Center position of the circumscribed rectangle. Average temperature For all detected high-temperature connected components, the following filtering rules are applied sequentially to remove interfering targets that clearly do not conform to the characteristics of a human body or bicycle: Area filter: Calculates the pixel area of ​​each connected component. .like (200px) or If the value is less than 400px, then the connected component is removed. This rule is used to filter out excessively small noise points and excessively large heat sources (such as car engines and building heat sources).

[0036] Shape Filter: For connected components filtered by area, calculate the aspect ratio (width to height ratio) of their smallest bounding rectangle. Only connected components with aspect ratios falling within the typical human or bicycle target range (between 0.2 and 0.8) are retained. This rule is used to filter based on the approximate shape of the target.

[0037] Output: Connected components that pass all the above filtering rules are marked as "high-temperature targets to be evaluated" by the system. The system will record the minimum bounding rectangle parameters (such as center point coordinates, width, height, and angle) for use by subsequent tracking or classification modules.

[0038] 3. Distance determination and early warning triggering a) Distance estimation model: for any marked target within the central region Obtain key image parameters: pixel height of the target in the image. (Height of the circumscribed rectangle). The ordinate of the bottom edge of the target. (With the top of the image as 0). Applying a simplified monocular ranging model: ; in The estimated physical distance (in meters). The target's true typical height is defined as a constant (1.7m for pedestrians, 2.0m for cyclists), and k2 is a calibration coefficient, combined with the camera's focal length. (mm), pixel size (μm / px) and installation height (m) Pre-calibrated results: .

[0039] Logical simplification: This method employs a fixed height assumption and an approximation of the image's bottom edge, sacrificing absolute accuracy for extremely fast computation. For the binary judgment of "whether it is less than 5 meters," this simplified model has been verified to be sufficiently reliable.

[0040] b) Level 1 Warning Trigger Logic: Conditional Judgment: If in the current frame... of The region contains at least one target that passes the screening. And its estimated distance If so, it will be triggered immediately. Key threshold. The preset distance is 5 meters. Engineering basis: At a distance of 5 meters, for vehicles traveling at 20-30 km / h on urban roads, this means the driver has only 0.36-0.6 seconds of reaction time, already at a critical point where immediate emergency measures (such as full braking) are necessary. This threshold is designed to only alert for imminent collision risks, minimizing interfering alarms. A Level 1 warning signal is generated as a Boolean flag WARNING_LEVEL_1 = TRUE, and can include the location information of the most dangerous target (based on...). (Determine whether it is left center or right center).

[0041] 4. Upon receiving a Level 1 warning signal, the warning execution module will drive a miniature vibration motor installed inside the non-motorized vehicle helmet to generate strong, continuous vibration. The vibration mode uses high-frequency pulses (approximately 3-5Hz), directly transmitting the tactile warning to the wearer's head through the helmet's internal structure, forming a physical sensory signal that cannot be ignored. This vibration warning does not rely on sound or vision and can effectively alert riders to the risk of close-range collisions in noisy environments, at night, or under conditions of limited visibility, prompting them to take immediate measures to slow down or avoid the collision.

[0042] Beneficial effects: 1. This invention can be applied to intelligent early warning for real-time assessment of non-motorized vehicle collision risks. It integrates YOLOv8 target detection, binocular stereo vision, and deep learning direction recognition technologies to construct a complete process from environmental perception to risk decision-making. The system identifies non-motorized vehicle targets using YOLOv8, obtains three-dimensional distance using binocular vision, calculates speed by combining continuous frame displacement, and determines the driving direction through a CNN model to form key risk parameters such as relative speed, distance, and angle.

[0043] 2. This invention inputs multi-source parameters into a risk quantification model based on the theory of safety potential energy field, achieving nonlinear fusion and outputting a normalized safety coefficient to characterize the real-time collision risk level. Based on this safety coefficient, the system uses the human-computer interaction module of the smart helmet to issue warnings to the rider through spatial tactile vibration and contextualized voice, effectively improving riding safety and reducing the risk of collision accidents without requiring the rider to shift their gaze. Attached Figure Description

[0044] Figure 1 This is a flowchart of the model of the present invention.

[0045] Figure 2 This is a structural framework diagram of the model of the present invention.

[0046] Figure 3 This describes the specific deployment of the invention on a helmet.

[0047] Figure 4 Establish a hotspot map of the spatial distribution of safety factors.

[0048] Figure 5 Analysis of the safety factor as a function of angle.

[0049] Figure 6 Analysis of safety factor variation with distance.

[0050] Figure 7 Multi-target recognition results in real-world scenarios.

[0051] Figure 8 The identification results of other non-motorized vehicles in the surrounding area when driving normally in the actual non-motorized vehicle lane.

[0052] Figure 9 The results of judgments under simulated dangerous driving conditions. Detailed Implementation

[0053] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0054] like Figure 1 and Figure 2As shown, this invention provides an intelligent early warning method for real-time assessment of non-motorized vehicle collision risks. This method constructs a dynamic safety potential energy field model that closely reflects the actual conflict patterns of non-motorized vehicles through the fusion of heterogeneous information from multiple modules. The system first achieves three-dimensional perception through high-precision synchronous binocular vision, then uses deep learning and filtering algorithms to stably track traffic participants and accurately calculate their relative motion states. Finally, multi-dimensional parameters such as relative speed, distance, and angle are input into the innovative safety potential energy field model to calculate a normalized comprehensive safety coefficient, which triggers personalized, low-interference early warnings. This method aims to fundamentally solve the problems of existing systems, such as single-sensory perception, model distortion, and significant early warning interference.

[0055] like Figure 3 As shown in the diagram, this system illustrates a complete early warning system, from "perception" to "decision-making" and then to "execution." It is based on a Raspberry Pi and integrates the YOLOv8 object detection algorithm and other risk models to process video streams and issue warnings.

[0056] S1: Synchronous Acquisition and Preprocessing of Multi-Source Heterogeneous Data This step aims to provide the system with high-quality, spatiotemporally strictly aligned raw image data, which is the foundation for ensuring the accuracy of subsequent 3D perception.

[0057] Hardware synchronization trigger: An exposure control mechanism based on hardware trigger signals (such as an external synchronization pulse generator) is employed to ensure that the left and right imaging sensors of the binocular camera are synchronously exposed and acquire images within the same microsecond time window.

[0058] Stereo correction and epipolar constraint alignment: After system installation, offline binocular camera calibration is performed using a high-precision checkerboard calibration board to accurately obtain the intrinsic parameter matrices (including focal lengths fx, fy, principal points cx, cy, and distortion coefficients) and extrinsic parameter matrices (rotation matrix R and translation vector T describing the position of the right camera relative to the left camera) of the left and right cameras.

[0059] This transformation simplifies stereo matching from a two-dimensional full-image search to a one-dimensional search along a horizontal line, greatly improving computational efficiency and matching accuracy, and is the cornerstone of subsequent high-precision three-dimensional reconstruction.

[0060] S2: Multi-object detection and continuous tracking based on deep learning This step identifies and tracks each non-motorized vehicle target in real time and continuously from the preprocessed image stream, providing target identity and temporal location information for state calculation.

[0061] Lightweight target detection: YOLOv8n (nano version) is used as the core detector. The model is pre-trained on the COCO dataset and fine-tuned using a dedicated non-motorized vehicle dataset containing categories such as "bicycle", "electric bicycle", and "tricycle" to improve detection accuracy and recall in the target scene.

[0062] For each frame of the corrected image, the model outputs a series of detection boxes, each containing the target category, confidence level, and normalized location (x, y, w, h) in the image.

[0063] The output assigns a unique track_id to each target and records its continuous position coordinates in the image sequence.

[0064] S3: Precise calculation of multi-dimensional interaction state parameters This step, based on the tracking results of S2, calculates in parallel the three core physical quantities that determine the collision risk between the main vehicle and each neighboring target: relative distance, relative speed, and relative angle.

[0065] High-precision relative distance measurement based on binocular vision: For each tracked target, ORB feature points are extracted within its detection bounding box. Using the epipolar constraints established in S1, a one-dimensional search is performed on the corresponding horizontal line in the right figure to find matching points for each ORB feature point.

[0066] Calculate the disparity of all successfully matched point pairs. Median filtering is used to remove outliers that mismatch, thus obtaining the representative disparity of the target. .

[0067] Based on the principle of binocular triangulation: Where B is the baseline distance and f is the corrected equivalent focal length (in pixels), the three-dimensional coordinates (X, Y, Z) of the target in the left camera coordinate system are calculated.

[0068] Calculate the three-dimensional coordinates of the driver's position and the surrounding vehicles, and the Euclidean distance between them. This is the precise relative distance. The process is as follows: Figure 4 As shown in the schematic diagram.

[0069] Relative velocity calculation based on 3D trajectory: For a stable trajectory, take its three-dimensional position in the world coordinate system (which can be transformed from the camera coordinate system) between the current frame t and the previous frame t-1. .

[0070] Calculate displacement vector Divide this by the precise frame interval time Δt (obtained from a high-precision timestamp) to obtain the instantaneous velocity vector of the target. And the speed magnitude v.

[0071] The relative velocity scalar between vehicle A and adjacent vehicle B is the magnitude of the difference between their velocity vectors: Relative approximation speed The relative velocity vector is the projection of the two vehicles onto the line connecting them, and is a key indicator for assessing the urgency of a collision.

[0072] Driving direction classification and relative angle calculation based on a dedicated CNN model (key innovation): Train a dedicated lightweight CNN orientation classification model (e.g., based on the MobileNetV2 architecture). The model input is a cropped image of a single non-motorized vehicle target (resized to 64x64), and the output layer is designed with 7 neurons, representing seven discrete orientation angles (relative to the horizontal axis of the image): 0°, 5°, 10°, 15°, 20°, 25°, and 30°.

[0073] The collected cropped images of non-motorized vehicles from different angles are input into the model to obtain the driving direction angle. The relative angle can be denoted as... ,in ,so .

[0074] Disruptive Risk Range Definition: Based on accident statistical analysis, this invention redefines the risk level of relative angles, such as... Figure 5 As shown: High-risk area (small angle: ): This corresponds to approaching, rear-end collisions, or head-on collisions in the same or opposite directions, with highly overlapping paths and extremely short reaction times.

[0075] Medium danger zone (medium angle: ): This corresponds to scenarios such as intersections, where there is a possibility of path crossing, but the angle has already provided some predictive clues.

[0076] Low-risk zone (wide angle): ): This corresponds to obvious lane changes or turns, with a small component of the motion vector in the collision direction, and relatively sufficient time for avoidance.

[0077] like Figure 6 As shown in the figure, this diagram visually illustrates the quantitative relationship between the safety factor and distance in the non-motorized vehicle collision risk warning model. Taking the direction directly in front (0° direction) as an example, the figure plots the corresponding curves between the safety factor (range 0.0-1.0) and the target distance (10-50 meters), reflecting the dynamic assessment logic of the system on the risk level during the perception and decision-making stages.

[0078] S4: Comprehensive Risk Quantification and Decision-Making Based on Security Potential Field This step is the core of the method, aiming to construct a risk quantification model with clear physical meaning, and to integrate the heterogeneous parameters calculated by S3 into a normalized comprehensive safety factor (SF).

[0079] Virtual quality calculation (S41): Definition .in, To calculate the average weight (e.g., 80kg for an electric bicycle); Risk coefficient for vehicle model; Velocity effect function (preferred) ), which amplifies the contribution of speed to the "dangerous equivalent" by a high power.

[0080] Calculation of safety factors for each dimension: Basic safety factor ( ): Simulates the decay of the risk field with distance. Directional safety factor ( Key innovation): To reflect the principle that "the smaller the angle, the higher the risk," a directional risk function is designed. (θ is in radians), preferably A=0.1, B=0.9, C=3.0. The function for calculating the directional safety factor is: .

[0081] Speed ​​safety factor ( ): Using piecewise linear functions to approximate relative velocities A highly nonlinear mapping is performed, and different speed safety factor judgment functions are applied within different speed ranges.

[0082] Quality and safety factor ( ): Preferred To prevent extreme values ​​from dominating.

[0083] Comprehensive decision-making: Weighted fusion (S47): The weighting reflects the risk priority: .

[0084] Risk Level Determination (S48): Mapping continuous SF to three levels of discrete risk: The threshold is calibrated based on historical conflict data.

[0085] S5: Personalized, low-interference proactive safety warning output When the system determines that the risk reaches the warning or danger level, it will activate personalized, low-interference, multi-level early warnings. The yellow warning level uses a mild alert: a buzzer is used to remind the user that the surrounding area is relatively dangerous and requires extra attention to road conditions. The red danger level activates a strong warning: a vibration is generated by the accompanying motor, and a buzzer responds simultaneously to warn the driver to avoid surrounding vehicles. When a "ghost peek" phenomenon is detected, a high-frequency buzzer sounds continuously for 1 second to warn of the high risk of a "ghost peek" event ahead. Figure 7 , Figure 8 and Figure 9 As shown, non-motorized vehicles are classified, judged, and warned under actual road conditions.

[0086] It should be noted that the above content merely illustrates the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, various improvements and modifications can be made without departing from the principle of the present invention, and all such improvements and modifications fall within the scope of protection of the claims of the present invention.

Claims

1. An intelligent early warning method for real-time assessment of non-motorized vehicle collision risk, characterized in that, The method includes the following steps: Step S1: Synchronous acquisition and preprocessing of multi-source heterogeneous data; 1a) Hardware synchronization: A hardware trigger signal or a high-precision software synchronization mechanism is used to ensure that the left and right imaging units of the binocular camera complete the exposure within the same microsecond time window; 1b) Stereo Correction and Epipolar Constraint: Using the camera intrinsic and extrinsic parameter matrices obtained from offline calibration, stereo correction is performed on the original left and right images so that the corrected images satisfy the epipolar constraints; Step S2: Multi-target detection and continuous tracking stage based on deep learning; 2a) Object detection and classification: YOLOv8n, a lightweight variant of the YOLOv8 series, is used as the core detection network. For each input image frame, the output is a detection box containing the categories of "bicycle", "electric bicycle" and "tricycle". 2b) Multi-target tracking and trajectory management: The ByteTrack algorithm is used for multi-target tracking. Each successfully tracked target is assigned a unique trajectory ID and its two-dimensional image coordinates over time are recorded. Step S3: Accurate calculation stage of multi-dimensional interaction state parameters; 3a) Relative velocity calculation: For the tracked target, the instantaneous velocity vector of the target is calculated based on its three-dimensional position in the world coordinate system in the current frame and the previous frame, and then the relative velocity scalar between the master vehicle and the neighboring vehicle is calculated. 3b) Relative distance measurement: Based on epipolar correction, for each tracked target, ORB feature points are extracted within its bounding box and one-dimensional search matching is performed. The three-dimensional coordinates of the target are calculated according to the binocular vision triangulation formula, and then the Euclidean distance between the main vehicle and the neighboring vehicle is calculated as the relative distance. 3c) Classification of driving direction and calculation of relative angle: Using a dedicated lightweight CNN model, the cropped image of the non-motorized vehicle target is input, and seven discrete driving direction angles of 0°, 5°, 10°, 15°, 20°, 25° and 30° are output. The absolute difference between the driving direction angles of the main vehicle and the adjacent vehicle is calculated as the relative angle. Step S4: Comprehensive risk quantification and decision-making stage based on the safety potential energy field; 4a) Virtual mass calculation: Based on the virtual mass calculation formula Calculate the virtual mass of the target; To calculate the average weight, we take 80 kg. For type coefficients, electric bicycles are assigned a coefficient of 1.0-1.2, and bicycles are assigned a coefficient of 0.8-1.

0. For the velocity effect function, the parameters are taken as follows: β=7.0, γ=0.5; 4b) Safety factor calculation: Based on the relative distance, relative angle, relative approach speed and virtual mass, calculate the basic safety factor, directional safety factor, speed safety factor and mass safety factor respectively; 4c) Comprehensive decision-making: The four safety factors mentioned above are weighted and summed to obtain a normalized comprehensive safety factor, and the risk level is determined as safe, alert, or dangerous based on the value of the comprehensive safety factor; Step S5: Personalized, low-interference proactive safety warning output stage: When the system determines that the risk level has reached the warning threshold, it will issue a warning to the rider through the human-computer interaction module of the smart helmet, either by tactile vibration or voice.

2. The intelligent early warning method for real-time assessment of non-motorized vehicle collision risk according to claim 1, characterized in that, In step S3b), the binocular visual triangulation formula is: Where Z is the target depth, B is the baseline distance, and f is the corrected equivalent focal length. The target is representative of the parallax.

3. The intelligent early warning method for real-time assessment of non-motorized vehicle collision risk according to claim 1, characterized in that, In step S3c), the relative angle The numerical range is divided into: the high-risk range is 0°≤ <10°, the medium-risk zone is 10°≤ <20°, the low-risk zone is 20°≤ ≤30°.

4. The intelligent early warning method for real-time assessment of non-motorized vehicle collision risk according to claim 1, characterized in that, In step S4b): the basic safety factor The calculation formula is: ,in These are preset parameters. It is a relative distance; The directional safety factor The calculation formula is: , where the directional risk function θ is the relative angle expressed in radians, and A, B, and C are preset constants; The speed safety factor SF_{speed} is calculated using a piecewise linear function: when hour, ; when hour, ; when hour, ; when hour, ; The quality and safety factor The calculation formula is: .

5. The intelligent early warning method for real-time assessment of non-motorized vehicle collision risk according to claim 4, characterized in that, In step S4b), the preset parameters are K=0.1 and k1=1.

3.

6. The intelligent early warning method for real-time assessment of non-motorized vehicle collision risk according to claim 4, characterized in that, In step S4b), the preset constants are A=0.1, B=0.9, and C=3.

0.

7. The intelligent early warning method for real-time assessment of non-motorized vehicle collision risk according to claim 1, characterized in that, In step S4c): The weighted summation formula is as follows: ; The weight allocation is as follows: ; The risk level is determined as follows: when SF≥0.5, it is considered safe; when 0.28≤SF<0.5, it is considered alert; and when SF<0.28, it is considered dangerous.

8. The intelligent early warning method for real-time assessment of non-motorized vehicle collision risk according to claim 1, characterized in that, Specifically, step S5 employs a "ghost peek" risk warning method based on infrared thermal sensing and area determination: S81. Thermal imaging image acquisition and three-zone division: a) Hardware deployment and data source: Sensing unit: An uncooled infrared focal plane array thermal imaging camera deployed at the front of the vehicle; b) Key Region Partitioning Algorithm For each input frame Perform static space partitioning: To obtain image dimensions: Let the image width be W and the height be H; Divide the sample into three equal parts: Left region ; Central region ; Right region ; The three areas are equal in size ; S82. High-temperature target detection in the intermediate area: a) Calculation Median temperature of all pixels This serves as a robust estimate of the current frame's ambient background temperature; then, a dynamic threshold is set. : ； in, A preset temperature rise threshold, typically 2℃-3℃, is used to distinguish between living organisms and the environment; Generate binary image Otherwise, it is 0; apply morphological opening to B(x,y) to mark all connected components. ; b) Target filtering and labeling: For each connected component Calculation feature: pixel area The center position of the circumscribed rectangle Average temperature For all detected high-temperature connected components, the following filtering rules are applied sequentially to remove interfering targets that clearly do not conform to the characteristics of a human body or bicycle: Area filter: Calculates the pixel area of ​​each connected component. ;like or If so, then the connected component is removed; Shape Filter: For connected components filtered by area, calculate the aspect ratio of their minimum bounding rectangle and retain only connected components whose aspect ratio falls within the range of typical human or bicycle targets. Output: Connected components that pass all the above filtering rules are marked by the system as "high temperature targets to be evaluated"; the system will record the minimum bounding rectangle parameter of each such target for use by subsequent tracking or classification modules; S83. Distance Determination and Warning Trigger a) Distance estimation model: for any marked target within the central region Obtain key image parameters: pixel height of the target in the image. ; the ordinate of the bottom edge of the target Applying a simplified monocular ranging model: ； in For the estimated physical distance, The target's true typical height is defined by k2, which is a calibration coefficient, combined with the camera's focal length. Pixel size and installation height Pre-calibration yielded: ; b) Level 1 warning triggering logic: Conditional judgment: If in the current frame of The region contains at least one target that passes the screening. And its estimated distance If it is, it will be triggered immediately; critical threshold The preset length is 5 meters. S84. After receiving the Level 1 warning signal, the warning execution module will drive the miniature vibration motor installed in the non-motorized vehicle helmet to generate strong and continuous vibration. The vibration mode adopts the form of high-frequency pulse, which directly transmits the tactile warning to the wearer's head through the internal structure of the helmet, forming a physical perception signal that cannot be ignored, reminding the rider of the risk of close-range collision ahead, prompting him to take immediate deceleration or avoidance measures.

Images