Transport vehicle anti-collision early warning and braking control system based on AI visual fusion

Abstract

According to the transport vehicle anti-collision early warning and braking control system based on AI visual fusion, full-dimension AI visual data collection is carried out on a transport vehicle through a deployed multi-modal sensing device, multi-modal data are obtained, image, speed, thermal radiation and vehicle load data can be accurately collected, the environment and vehicle state can be captured in a full-dimension mode, and the vehicle anti-collision early warning and braking control system based on AI visual fusion is provided. A foundation is built for subsequent processing, asynchronous fusion is performed on multi-modal data to obtain multi-source fusion data, data integration can be rapidly completed, synchronous fusion delay is reduced, meanwhile, the risk of privacy disclosure during data uploading is avoided, a dynamic risk field is established based on the multi-source fusion data, risk assessment is performed on the vehicle state in the dynamic risk field, and the risk assessment accuracy is improved. Graded early warning and braking decisions are obtained according to a risk assessment result, dynamic risk assessment is combined with road conditions and load adjustment grades, graded early warning and braking decisions adapt to different scenes, and safety and operation efficiency are balanced.

Description

Technical Field

[0001] This invention relates to the field of vehicle control technology, and in particular to a collision avoidance warning and braking control system for transport vehicles based on AI vision fusion. Background Technology

[0002] With the large-scale development of the logistics industry, the driving safety of transport vehicles is directly related to the lives of people, property, and road traffic efficiency. Existing anti-collision systems for transport vehicles are mostly based on single vision sensors or fusion solutions of LiDAR and vision, which have many limitations: First, they have poor adaptability to complex environments, and are prone to misidentification or missed identification of obstacles in adverse weather conditions such as strong light, shadows, rain, and fog. In particular, they lack the ability to distinguish virtual obstacles (such as road surface reflections and sign shadows), leading to false warnings or braking. Second, the dynamic target trajectory prediction is simplistic, judging collision risk only based on distance and speed parameters, without considering the braking delay caused by the heavy load characteristics of logistics vehicles and the inertia of cargo, resulting in low risk assessment accuracy. Third, the braking control strategy is rigid, adopting a one-size-fits-all approach of warning and emergency braking, which is not adapted to different road conditions (highways, national highways, factory areas) and cargo status, and is prone to secondary risks such as cargo tipping and vehicle rollover. Fourth, they lack adaptive learning capabilities, cannot optimize model parameters based on historical data of different transportation scenarios, and have weak system generalization ability.

[0003] To address the aforementioned issues, there is an urgent need for a collision avoidance warning and braking control system that integrates multimodal visual information, adapts to the characteristics of logistics vehicles, and possesses dynamic decision-making and self-optimization capabilities. This system would break through the conventional framework of existing technologies and improve the reliability of safety protection in complex scenarios. Summary of the Invention

[0004] This invention provides a collision avoidance warning and braking control system for transport vehicles based on AI vision fusion, which is used to solve the problems mentioned in the background art.

[0005] A collision avoidance warning and braking control system for transport vehicles based on AI vision fusion, comprising: The perception layer is used to collect multi-dimensional AI visual data from transport vehicles based on deployed multi-modal perception devices to obtain multi-modal data. The edge computing layer is used to asynchronously fuse multimodal data to obtain multi-source fused data; The decision control layer is used to establish a dynamic risk field based on multi-source fusion data, conduct risk assessment of vehicle status in the dynamic risk field, and obtain graded warnings and braking decisions based on the risk assessment results. The execution layer is used to control the braking system to work in coordination according to the braking decision, and to perform collision prediction in real time. When it is determined that a collision cannot be avoided, it provides steering assistance parameters.

[0006] Preferably, it also includes a self-learning optimization module, which is used to summarize the driving data of transport vehicles and periodically update the asynchronous fusion, dynamic risk assessment mechanism, and auxiliary parameters based on the driving data.

[0007] Preferably, the sensing layer includes: A vision camera is used to acquire two-dimensional images of the scene in front of the transport vehicle and output obstacle outlines, texture features, and relative distance data. Millimeter radar waves are used to collect data on the speed, azimuth, and trajectory of obstacles. Infrared imagers are used to activate in low-light conditions and capture the thermal radiation characteristics of obstacles. Environmental sensors are used to collect data on road surface adhesion coefficient, light intensity, rainfall, and real-time vehicle status.

[0008] Preferably, the edge computing layer includes: A fast processing unit is used to perform lightweight neural network processing on single-channel visual images in multimodal data to obtain two-dimensional processed data; The precision processing unit is used to perform temporal and spatial alignment of binocular vision images, infrared thermal imaging images, and lidar point clouds to obtain three-dimensional processed data. The asynchronous fusion unit is used to correlate two-dimensional and three-dimensional processed data. When the three-dimensional processed data is delayed, the two-dimensional processed data is used for extrapolation, and the extrapolation result is updated immediately after the three-dimensional processed data is acquired.

[0009] Preferably, the decision control layer includes: The risk field establishment unit is used to establish a static obstacle risk field, a dynamic target prediction risk field, and a road structure risk field based on multi-source fusion data. The static obstacle risk field, the dynamic target prediction risk field, and the road structure risk field are then superimposed to obtain the dynamic risk field. The trajectory analysis unit is used to predict multiple short-term trajectories of the transport vehicle based on the current vehicle speed, steering wheel angle, and driver intention. Based on the dynamic risk field, it obtains the longitudinal predicted position of each short-term trajectory along the road centerline and the lateral predicted position perpendicular to the road centerline at any time, and determines the risk concentration value. For each short-term trajectory within the future time window, it multiplies the risk concentration value at each instant by the instantaneous vehicle speed, and continuously accumulates all instantaneous results to obtain the cumulative risk exposure. It also extracts the maximum risk concentration value and the time point of occurrence for each short-term trajectory. The vector building unit is used to obtain the rate of change of cumulative risk exposure and maximum concentration over time. It integrates driver status, vehicle status, multiple short-term trajectories, cumulative risk exposure, maximum concentration, as well as the time of occurrence and rate of change to obtain a risk status vector. The graded early warning unit is used to compare the risk status vector with the early warning rule threshold and determine the graded early warning based on the comparison result. The braking decision unit is used to make braking decisions based on a gradient braking control mechanism that matches the dynamic risk field when a graded warning indicates that the driver's intervention is insufficient or that there is not enough time to react.

[0010] Preferably, the risk field establishment unit includes: The static risk field establishment unit is used to assign basic hazard weights based on the physical properties of obstacles, assign dynamic hazard weights based on the specific characteristics of obstacles that change with environmental parameters, generate single obstacle risk impact areas based on basic hazard weights, dynamic hazard weights and obstacle characteristics, and superimpose all single obstacle risk impact areas to obtain the static obstacle risk field. The target risk field establishment unit is used to predict multiple future motion trajectories based on the trajectory of a dynamic target through a deep learning model, determine the probability of each future motion trajectory, establish a probability cloud of the dynamic target over time, generate a virtual repulsive force vector field proportional to the predicted kinetic energy of the dynamic target, convolve the virtual repulsive force vector field with the probability cloud, and combine it with the type gain of the dynamic target to generate a predicted risk field of the dynamic target. The road risk field establishment unit is used to obtain the road curvature and the real-time estimated vehicle lateral adhesion coefficient, determine the lateral instability risk, and generate an additional risk gradient on the outside of the curve that is proportional to the curvature and the square of the speed. Based on the lateral instability risk and the additional risk gradient, a quadratic attractive potential well is set on the lane centerline, and an exponential repulsive potential barrier is set on the lane line and the road boundary to obtain the road structure risk field. The risk field integration unit is used to define longitudinal distance, lateral offset and time with the center line of the road ahead as a reference, establish a risk field coordinate system, and configure the static obstacle risk field, dynamic target prediction risk field and road structure risk field into the risk field coordinate system to obtain the dynamic risk field.

[0011] Preferably, it also includes a risk field calibration unit, which performs real-time calibration of the dynamic risk field based on a three-level adaptive calibration model; The first layer of the three-level adaptive calibration model is based on vehicle physical state calibration: The vehicle load is acquired, and the dynamic risk field is calibrated in real time according to the rule that the risk area of ​​static obstacles expands but the remote influence range shrinks under heavy load. Obtain the road surface adhesion coefficient and, according to the rule that the same curve is more dangerous under low adhesion, calibrate the dynamic risk field in real time. In the three-level adaptive calibration model, the second level is calibration based on driver behavior profiles; Based on the driver's risk preference coefficient, reaction time model, and trust decay factor, the dynamic risk field is calibrated in real time. In the three-level adaptive calibration model, the third level is scenario-based dynamic decision calibration. Establish dynamic early warning threshold surface and dynamic braking threshold surface to perform real-time calibration of dynamic risk field.

[0012] Preferably, the graded early warning unit includes: The Level 1 warning unit is used to determine a Level 1 warning when the cumulative risk exposure exceeds the dynamic baseline threshold determined based on driver profiles and road conditions, and the maximum concentration value is not in the emergency area. The Level 2 warning unit is used to determine a Level 2 warning when the maximum concentration value is greater than the primary warning threshold and the cumulative risk exposure shows an upward trend. The Level 3 warning unit is used to determine a Level 3 warning when the maximum concentration value exceeds the high-level warning threshold.

[0013] Preferably, the gradient braking control mechanism in the braking decision unit is as follows: Pre-braking phase: Once braking arbitration is initiated, the braking system is immediately pre-filled; Proportional following phase: The target braking force is proportional to the gradient at the current moment and is limited by the current road adhesion coefficient and vehicle load; Full braking phase: If the risk increases sharply, enter full braking mode and apply the maximum safety braking force.

[0014] Preferably, the execution layer includes: The collision prediction unit is used to determine whether braking can avoid a collision based on the real-time vehicle dynamics and dynamic risk field. The parameter determination unit is used to generate steering assistance parameters based on real-time acquired lateral space data when collisions cannot be avoided by braking.

[0015] Compared with the prior art, the present invention has achieved the following beneficial effects: By deploying multimodal sensing devices, comprehensive AI visual data is collected from transport vehicles, yielding multimodal data. This data can accurately capture images, speed, thermal radiation, and vehicle load data in various scenarios, including high-light conditions on highways, rain and fog on national roads, and low-light conditions in factory areas. It comprehensively captures environmental and vehicle status, solving the problem of false or missed identifications in harsh environments caused by single sensors, thus laying a solid foundation for subsequent processing. Asynchronous fusion of multimodal data yields multi-source fused data, which can be processed locally at the edge without relying on the cloud. In sudden scenarios such as emergency braking by a vehicle ahead on a highway, data integration can be completed quickly, reducing synchronous fusion latency and mitigating the risk of privacy leaks during data upload. A dynamic risk field is established based on the multi-source fused data, and vehicle status is assessed within this dynamic risk field. The results provide tiered early warning and braking decisions. Dynamic risk assessment, combined with road conditions and load adjustment levels, allows for tiered early warning and braking decisions adapted to the transportation of different goods such as fragile items and building materials. In heavy-load scenarios on national highways, it can prevent cargo tipping due to sudden braking, while in low-speed scenarios in factory areas, it reduces excessive early warning interference, balancing safety and operational efficiency. Based on braking decisions, the braking system works collaboratively and performs real-time collision prediction. When it is determined that a collision cannot be avoided, steering assist parameters are provided. The braking system works collaboratively, adapting to pneumatic and hydraulic braking architectures, to precisely control the braking pressure of heavy-load vehicles. In high-speed obstacle avoidance and narrow factory road scenarios, the steering assist parameters provide safe steering references, compensating for the lack of maneuverability in heavy-load vehicles, ensuring safety and preventing cargo damage, and solving the rigid braking defects of conventional systems.

[0016] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in this application.

[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a structural diagram of a collision avoidance warning and braking control system for transport vehicles based on AI vision fusion, as described in an embodiment of the present invention. Figure 2 This is a structural diagram of the edge computing layer in an embodiment of the present invention. Detailed Implementation

[0020] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0021] Example 1: This embodiment of the invention provides a collision avoidance warning and braking control system for transport vehicles based on AI vision fusion, such as... Figure 1 As shown, it includes: The perception layer is used to collect multi-dimensional AI visual data from transport vehicles based on deployed multi-modal perception devices to obtain multi-modal data. The edge computing layer is used to asynchronously fuse multimodal data to obtain multi-source fused data; The decision control layer is used to establish a dynamic risk field based on multi-source fusion data, conduct risk assessment of vehicle status in the dynamic risk field, and obtain graded warnings and braking decisions based on the risk assessment results. The execution layer is used to control the braking system to work in coordination according to the braking decision, and to perform collision prediction in real time. When it is determined that a collision cannot be avoided, it provides steering assistance parameters.

[0022] In this embodiment, the multimodal sensing device is a combination of various types of sensing devices, which achieves full-dimensional data acquisition through complementary acquisition, with core coverage of visual, motion, and environmental dimensions.

[0023] In this embodiment, asynchronous fusion addresses the differences in data transmission rates and latency among multiple sensors without forcing the synchronization of timestamps. Instead, it uses algorithms to align data timing and feature dimensions to achieve dynamic fusion.

[0024] In this embodiment, the tiered warning system includes different risk levels and adopts a differentiated warning approach to gradually enhance the driver's perception and avoid over-warning.

[0025] In this embodiment, the auxiliary parameters are calculated based on the available space around the vehicle and the driving speed, such as the safe steering angle and torque, to help the driver avoid obstacles.

[0026] The beneficial effects of the above design scheme are as follows: By deploying multimodal sensing devices, comprehensive AI visual data collection of transport vehicles is achieved, resulting in multimodal data. This allows for accurate acquisition of image, speed, thermal radiation, and vehicle load data in scenarios such as high-speed strong light, national highway rain and fog, and low-light conditions in factory areas. It comprehensively captures environmental and vehicle status, solving the problem of false or missed identifications in harsh environments using a single sensor, thus laying a solid foundation for subsequent processing. Asynchronous fusion of multimodal data yields multi-source fused data, enabling local processing at the edge without relying on the cloud. In sudden scenarios such as emergency braking by a vehicle ahead on a highway, data integration can be completed quickly, reducing synchronous fusion latency and mitigating the risk of privacy leaks during data upload. Furthermore, a dynamic risk field is established based on the multi-source fused data, allowing for risk assessment of vehicle status within this dynamic risk field. Based on risk assessment results, tiered early warnings and braking decisions are generated. Dynamic risk assessments are combined with road conditions and load to adjust the level. Tiered early warnings and braking decisions are adapted to the transportation of different goods such as fragile items and building materials. In heavy-load scenarios on national highways, sudden braking can avoid cargo tipping. In low-speed scenarios in factory areas, excessive early warning interference is reduced, balancing safety and operational efficiency. According to the braking decision, the braking system is controlled to work in coordination and collision prediction is performed in real time. When it is determined that a collision cannot be avoided, steering assist parameters are provided. The braking system works in coordination to adapt to pneumatic and hydraulic braking architectures, accurately controlling the braking pressure of heavy-load vehicles. In high-speed obstacle avoidance and narrow factory road scenarios, steering assist parameters provide safe steering references, making up for the lack of maneuverability of heavy-load vehicles, ensuring safety and avoiding cargo damage, and solving the rigid braking defects of conventional systems.

[0027] Example 2: Based on Example 1, this embodiment of the invention provides a collision avoidance warning and braking control system for transport vehicles based on AI vision fusion, which also includes a self-learning optimization module for summarizing transport vehicle driving data and periodically updating asynchronous fusion, dynamic risk assessment mechanism and auxiliary parameters based on driving data.

[0028] The beneficial effects of the above design scheme are: by aggregating fleet driving data, iteratively updating asynchronous fusion, risk assessment and steering parameters, continuously adapting to multiple scenarios, improving the system's generalization ability, and helping the fleet to be promoted on a large scale.

[0029] Example 3: Based on Example 1, this embodiment of the invention provides a collision avoidance warning and braking control system for transport vehicles based on AI vision fusion, wherein the perception layer includes: A vision camera is used to acquire two-dimensional images of the scene in front of the transport vehicle and output obstacle outlines, texture features, and relative distance data. Millimeter radar waves are used to collect data on the speed, azimuth, and trajectory of obstacles. Infrared imagers are used to activate in low-light conditions and capture the thermal radiation characteristics of obstacles. Environmental sensors are used to collect data on road surface adhesion coefficient, light intensity, rainfall, and real-time vehicle status.

[0030] In this embodiment, the real-time vehicle status data includes load, vehicle speed, braking system pressure, etc.

[0031] In this embodiment, the thermal radiation characteristics of obstacles are used to distinguish between living targets (pedestrians, animals) and non-living targets (static obstacles, virtual shadows).

[0032] In this embodiment, the millimeter-wave radar is installed in the middle of the vehicle bumper. It has a strong ability to penetrate rain, fog, and dust, which can supplement the visual sensor's perception shortcomings in bad weather. The detection range is 5-300m.

[0033] In this embodiment, the vision camera is a binocular stereo vision camera with a sampling frequency of 30fps and a depth measurement range of 0.5-200m.

[0034] In this embodiment, all data collected by the visual camera, millimeter-wave radar, infrared imager, and environmental sensors are used as multimodal data.

[0035] The beneficial effects of the above design scheme are as follows: It accurately acquires two-dimensional images of the scene in front, and simultaneously outputs obstacle outlines, texture features, and relative distance data, providing a core image foundation for subsequent AI visual fusion and obstacle recognition. The millimeter-wave radar makes up for the shortcomings of the visual camera in terms of environmental adaptability, and has outstanding ability to penetrate rain, fog, and dust. It can stably acquire obstacle speed, azimuth angle, and motion trajectory data. The infrared imager is specifically adapted to low-light scenes. By capturing the thermal radiation characteristics of obstacles, it distinguishes real targets from virtual shadows, solving the problem of blurred visual recognition in low-light environments. In particular, it can accurately identify live targets such as pedestrians and animals. The environmental sensor collects road surface, environmental, and vehicle data in multiple dimensions. The road surface adhesion coefficient, rainfall, etc., provide environmental basis for braking decisions.

[0036] Example 4: Based on Example 1, this embodiment of the invention provides a collision avoidance warning and braking control system for transport vehicles based on AI vision fusion, such as... Figure 2 As shown, the edge computing layer includes: A fast processing unit is used to perform lightweight neural network processing on single-channel visual images in multimodal data to obtain two-dimensional processed data; The precision processing unit is used to perform temporal and spatial alignment of binocular vision images, infrared thermal imaging images, and lidar point clouds to obtain three-dimensional processed data. The asynchronous fusion unit is used to correlate two-dimensional and three-dimensional processed data. When the three-dimensional processed data is delayed, the two-dimensional processed data is used for extrapolation, and the extrapolation result is updated immediately after the three-dimensional processed data is acquired.

[0037] The beneficial effects of the above design scheme are as follows: It employs a lightweight neural network to process single-channel visual images, achieving efficient and rapid computation at the edge, significantly reducing data processing latency and computational power consumption, avoiding excessive resource consumption by complex models, and accurately outputting two-dimensional processed data to provide foundational support for subsequent fusion. It specifically aligns binocular vision, infrared thermal imaging, and LiDAR point cloud data, achieving dual calibration in both time and space dimensions, eliminating misalignment biases in multi-source data, generating high-precision three-dimensional processed data, accurately restoring the spatial structure of the scene and the relationship between target positions, providing reliable three-dimensional basis for collision risk assessment, and intelligently associating two-dimensional and three-dimensional processed data. When dealing with three-dimensional data latency scenarios, it maintains continuous system operation through two-dimensional data extrapolation, and updates results immediately after acquiring three-dimensional data. This ensures both the real-time nature of data fusion and the final accuracy, avoiding warning interruptions or misjudgments caused by three-dimensional data lag. It is also suitable for scenarios with uneven multimodal data transmission rates in complex road conditions for logistics vehicles, improving system stability.

[0038] Example 5: Based on Example 1, this embodiment of the invention provides a collision avoidance warning and braking control system for transport vehicles based on AI vision fusion. The decision control layer includes: The risk field establishment unit is used to establish a static obstacle risk field, a dynamic target prediction risk field, and a road structure risk field based on multi-source fusion data. The static obstacle risk field, the dynamic target prediction risk field, and the road structure risk field are then superimposed to obtain the dynamic risk field. The trajectory analysis unit is used to predict multiple short-term trajectories of the transport vehicle based on the current vehicle speed, steering wheel angle, and driver intention. Based on the dynamic risk field, it obtains the longitudinal predicted position of each short-term trajectory along the road centerline and the lateral predicted position perpendicular to the road centerline at any time, and determines the risk concentration value. For each short-term trajectory within the future time window, it multiplies the risk concentration value at each instant by the instantaneous vehicle speed, and continuously accumulates all instantaneous results to obtain the cumulative risk exposure. It also extracts the maximum risk concentration value and the time point of occurrence for each short-term trajectory. The vector building unit is used to obtain the rate of change of cumulative risk exposure and maximum concentration over time. It integrates driver status, vehicle status, multiple short-term trajectories, cumulative risk exposure, maximum concentration, as well as the time of occurrence and rate of change to obtain a risk status vector. The graded early warning unit is used to compare the risk status vector with the early warning rule threshold and determine the graded early warning based on the comparison result. The braking decision unit is used to make braking decisions based on a gradient braking control mechanism that matches the dynamic risk field when a graded warning indicates that the driver's intervention is insufficient or that there is not enough time to react.

[0039] The beneficial effects of the above design scheme are as follows: By integrating three types of risk fields—static obstacles, dynamic target prediction, and road structure—discrete perceived targets and continuous road geometry are unified and quantified into a dynamic and computable risk terrain map. This allows the system to simultaneously assess risks from different dimensions. The introduction of cumulative risk exposure as a core assessment indicator transforms risk assessment from instantaneous point-in-time judgment to integration over future time periods. By combining risk concentration, exposure time, and vehicle speed for path integration, chronic hazards accumulated from prolonged exposure to moderate-risk environments can be identified earlier, rather than simply responding to sudden, instantaneous high-risk events. Simultaneously predicting multiple trajectories and assessing their risks provides a comparative basis for subsequent optimization decisions. Integrating risk indicators, changing trends, and vehicle and driver states to form a multi-dimensional risk state vector condenses the complex surrounding situation and the driver's own state into a structured and interpretable decision input. This ensures that decisions are based not only on the magnitude of the risk but also on the speed of risk changes and the driver's ability to cope, providing comprehensive data support for subsequent personalized warnings and arbitration. By comparing with dynamic thresholds that integrate driver characteristics and vehicle states, precise and personalized classification of warning levels is achieved, significantly improving the accuracy of warnings and driver acceptance. This avoids frequent false alarms or delayed or missed alarms caused by fixed thresholds. Braking intervention is smoothly and adaptively adjusted according to the evolution trend and form of the risk field, enabling braking response to both gentle early braking in the initial stage of risk accumulation to resolve the crisis and decisive execution in emergency situations, while maximizing ride comfort and vehicle stability. Ultimately, this achieves full-process intelligentization from environmental perception, risk quantification, trend prediction to personalized decision-making, greatly improving the smoothness and trust of human-machine co-driving while enhancing safety.

[0040] Example 6: Based on Example 5, this embodiment of the invention provides a collision avoidance warning and braking control system for transport vehicles based on AI vision fusion. The risk field establishment unit includes: The static risk field establishment unit is used to assign basic hazard weights based on the physical properties of obstacles, assign dynamic hazard weights based on the specific characteristics of obstacles that change with environmental parameters, generate single obstacle risk impact areas based on basic hazard weights, dynamic hazard weights and obstacle characteristics, and superimpose all single obstacle risk impact areas to obtain the static obstacle risk field. The target risk field establishment unit is used to predict multiple future motion trajectories based on the trajectory of a dynamic target through a deep learning model, determine the probability of each future motion trajectory, establish a probability cloud of the dynamic target over time, generate a virtual repulsive force vector field proportional to the predicted kinetic energy of the dynamic target, convolve the virtual repulsive force vector field with the probability cloud, and combine it with the type gain of the dynamic target to generate a predicted risk field of the dynamic target. The road risk field establishment unit is used to obtain the road curvature and the real-time estimated vehicle lateral adhesion coefficient, determine the lateral instability risk, and generate an additional risk gradient on the outside of the curve that is proportional to the curvature and the square of the speed. Based on the lateral instability risk and the additional risk gradient, a quadratic attractive potential well is set on the lane centerline, and an exponential repulsive potential barrier is set on the lane line and the road boundary to obtain the road structure risk field. The risk field integration unit is used to define longitudinal distance, lateral offset and time with the center line of the road ahead as a reference, establish a risk field coordinate system, and configure the static obstacle risk field, dynamic target prediction risk field and road structure risk field into the risk field coordinate system to obtain the dynamic risk field.

[0041] In this embodiment, the probability cloud represents the spatial location that a dynamic target may occupy at different times in the future.

[0042] In this embodiment, a quadratic attractive potential well is set up at the lane centerline based on the lateral instability risk and the additional risk gradient, while an exponential repulsive potential barrier is set up at the lane lines and road boundaries. This can be interpreted as the system setting up a wide and gentle attractive valley to guide vehicles to stay within the core area for safe passage. At the lane edges and boundaries such as shoulders and guardrails, steep repulsive barriers are set up, where the risk of crossing the boundary increases sharply to prevent vehicles from deviating from the drivable area.

[0043] The beneficial effects of the above design scheme are as follows: By assigning both basic and dynamic hazard weights to obstacles, the risk assessment has evolved from fixed attributes to environmental coupling. The system can adjust the shape and range of the risk field in real time according to weather and lighting, making the static risk assessment more in line with the actual perceived uncertainty and the driver's psychological expectations. By combining trajectory probability prediction with a kinetic energy-based virtual repulsive force field, it not only predicts the possible location of dynamic targets, but also quantifies the sense of pressure and urgency of their behavior on the surrounding space. This allows the system to perceive and quantify abstract risks such as aggressive driving styles of vehicles ahead, achieving early prediction of interaction intentions. By combining curvature, speed and real-time lateral adhesion coefficient, the physical law that excessive speed in curves may lead to sideslip is directly and intrinsically reflected in the risk field. The setting of attractive potential traps and repulsive potential barriers provides vehicles with a natural guidance channel that conforms to safe driving standards, making the risk assessment both standardized and physically realistic. By establishing a unified spatiotemporal coordinate system, the three types of heterogeneous risk sources are seamlessly integrated into the same field. This enables the system to perform cross-dimensional risk comparisons and overlays, providing a unique and coherent data foundation for subsequent trajectory risk integration and unified decision-making.

[0044] Example 7: Based on Example 6, this embodiment of the invention provides a collision avoidance warning and braking control system for transport vehicles based on AI vision fusion, which also includes a risk field calibration unit that performs real-time calibration of the dynamic risk field based on a three-level adaptive calibration model; The first layer of the three-level adaptive calibration model is based on vehicle physical state calibration: The vehicle load is acquired, and the dynamic risk field is calibrated in real time according to the rule that the risk area of ​​static obstacles expands but the remote influence range shrinks under heavy load. Obtain the road surface adhesion coefficient and, according to the rule that the same curve is more dangerous under low adhesion, calibrate the dynamic risk field in real time. In the three-level adaptive calibration model, the second level is calibration based on driver behavior profiles; Based on the driver's risk preference coefficient, reaction time model, and trust decay factor, the dynamic risk field is calibrated in real time. In the three-level adaptive calibration model, the third level is scenario-based dynamic decision calibration. Establish dynamic early warning threshold surface and dynamic braking threshold surface to perform real-time calibration of dynamic risk field.

[0045] In this embodiment, the driver's risk preference coefficient is derived through cluster analysis based on statistics such as average following distance, aggressiveness of lane changes, and speed of cornering during historical trips.

[0046] In this embodiment, the reaction time model is a piecewise linear model that describes typical reaction delays under different fatigue indices (based on historical data from steering wheel micro-operations and facial feature analysis from in-vehicle cameras).

[0047] In this embodiment, the trust decay factor records the driver's past response to system warnings (such as whether the driver brakes in time after the system warning). If the response is high, the trust decay factor increases, and the system will be relatively stable near the warning threshold. If the response is low, the trust decay factor decreases, and the system will tend to issue warnings earlier and more clearly to wake up the driver.

[0048] In this embodiment, the dynamic warning threshold surface and the dynamic braking threshold surface make the final intervention decision not based on a fixed threshold, but dynamic. The dynamic warning threshold surface indicates that the warning threshold will automatically decrease when the road surface has low adhesion, when facing a conservative driver, or when the driver's reaction seems to slow down.

[0049] The beneficial effects of the above design scheme are: It directly injects key physical parameters such as vehicle load and road surface adhesion coefficient into the risk field morphology. This ensures that risk assessment strictly adheres to physical laws: the braking performance of heavy-duty vehicles decreases, and the core risk area is reasonably amplified; the risk of low-adhesion road surface curves is significantly enhanced. The system's decisions are thus precisely matched with the vehicle's instantaneous dynamic capabilities, avoiding premature or delayed warnings or braking due to differences in vehicle conditions. By integrating the driver's personalized characteristics and history of trust in the system, risk perception and warning strategies are tailored to the individual driver. The system is more sensitive to conservative drivers to provide a sense of security, and more robust to aggressive drivers to reduce interference. It also intervenes in advance when the driver distrusts the system to rebuild safety redundancy, significantly improving human-machine collaboration efficiency and acceptance. By establishing a dynamic threshold surface linked to multiple parameters, the triggering timing of warnings and braking is no longer a fixed value, but an intelligent surface that evolves in real time with vehicle status, driver profile, and environmental risks. This ensures that the system can automatically find the optimal balance point under different scenario combinations, achieving global adaptive optimization of safety and comfort.

[0050] Example 8: Based on Example 5, this embodiment of the invention provides a collision avoidance warning and braking control system for transport vehicles based on AI vision fusion, wherein the hierarchical warning unit includes: The Level 1 warning unit is used to determine a Level 1 warning when the cumulative risk exposure exceeds the dynamic baseline threshold determined based on driver profiles and road conditions, and the maximum concentration value is not in the emergency area. The Level 2 warning unit is used to determine a Level 2 warning when the maximum concentration value is greater than the primary warning threshold and the cumulative risk exposure shows an upward trend. The Level 3 warning unit is used to determine a Level 3 warning when the maximum concentration value exceeds the high-level warning threshold.

[0051] In this embodiment, during a Level 1 warning, the risk source outline is highlighted in amber semi-transparent on the augmented reality head-up display, and the dashboard displays a risk heat map overview of the road ahead. During a Level 2 warning, the risk source outline flashes orange; accompanied by a short warning sound, the steering wheel or seat belt provides a gentle tactile pulse. During a Level 3 warning, the risk source outline flashes bright red, and an arrow pointing in the escape direction is superimposed; a clear and distinct voice alarm is issued, such as "Brake! Vehicle cutting in from the left!"; the tactile feedback is upgraded to a continuous, high-frequency vibration or pulse sequence, simulating the nodding sensation of emergency braking.

[0052] The beneficial effects of the above design scheme are: the three-level warning logic is clear, the criteria are complementary (taking into account both cumulative and instantaneous amounts), and the thresholds are dynamically adjustable. This forms a smooth and progressive warning gradient, which not only avoids overreaction or delayed response caused by a single threshold, but also significantly improves the driver's trust and acceptance of the system through the interpretability of the tiered system.

[0053] Example 9: Based on Example 5, this embodiment of the invention provides a collision avoidance warning and braking control system for transport vehicles based on AI vision fusion. The gradient braking control mechanism in the braking decision unit is specifically as follows: Pre-braking phase: Once braking arbitration is initiated, the braking system is immediately pre-filled; Proportional following phase: The target braking force is proportional to the gradient at the current moment and is limited by the current road adhesion coefficient and vehicle load; Full braking phase: If the risk increases sharply, enter full braking mode and apply the maximum safety braking force.

[0054] In this embodiment, the pre-braking phase is to eliminate mechanical backlash.

[0055] In this embodiment, the proportional following stage achieves a linear and smooth increase in braking force as the risk increases.

[0056] The beneficial effects of the above design scheme are as follows: In the pre-braking stage, by eliminating mechanical backlash, the response delay of the braking system is minimized, gaining crucial tens of milliseconds for emergency intervention and laying the physical foundation for rapid response. In the proportional following stage, the braking force is dynamically bound to the real-time risk gradient, realizing a smooth and linear increase in braking force as the risk increases, greatly improving comfort and driving confidence. At the same time, the braking force is limited in real time according to the load and adhesion, ensuring braking efficiency and stability. In the full braking stage, as the final safety redundancy, the maximum and controlled braking force is provided when the risk deteriorates sharply, seamlessly connecting with the previous gentle braking, forming a complete and coherent safety closed loop from warning to full braking.

[0057] Example 10: Based on Example 1, this embodiment of the invention provides a collision avoidance warning and braking control system for transport vehicles based on AI vision fusion, wherein the execution layer includes: The collision prediction unit is used to determine whether braking can avoid a collision based on the real-time vehicle dynamics and dynamic risk field. The parameter determination unit is used to generate steering assistance parameters based on real-time acquired lateral space data when collisions cannot be avoided by braking.

[0058] The beneficial effects of the above design scheme are: by coupling the evolution of real-time vehicle dynamics and dynamic risk field, the accuracy of collision warning is greatly improved. When the braking capacity reaches its limit, lateral space is introduced and auxiliary parameters are generated according to the real-time environment, which ensures the coordination of human-machine co-driving, providing a final chance to avoid collision while respecting the driver's control.

[0059] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of this application and its equivalents, this invention also intends to include these modifications and variations.

Claims

1. A collision avoidance warning and braking control system for transport vehicles based on AI vision fusion, characterized in that, include: The perception layer is used to collect multi-dimensional AI visual data from transport vehicles based on deployed multi-modal perception devices to obtain multi-modal data. The edge computing layer is used to asynchronously fuse multimodal data to obtain multi-source fused data; The decision control layer is used to establish a dynamic risk field based on multi-source fusion data, conduct risk assessment of vehicle status in the dynamic risk field, and obtain graded warnings and braking decisions based on the risk assessment results. The execution layer is used to control the braking system to work in coordination according to the braking decision, and to perform collision prediction in real time. When it is determined that a collision cannot be avoided, it provides steering assistance parameters.

2. The collision avoidance warning and braking control system for transport vehicles based on AI vision fusion according to claim 1, characterized in that, It also includes a self-learning optimization module, which is used to aggregate the driving data of transport vehicles and periodically update the asynchronous fusion, dynamic risk assessment mechanism, and auxiliary parameters based on the driving data.

3. The collision avoidance warning and braking control system for transport vehicles based on AI vision fusion according to claim 1, characterized in that, The perception layer includes: A vision camera is used to acquire two-dimensional images of the scene in front of the transport vehicle and output obstacle outlines, texture features, and relative distance data. Millimeter radar waves are used to collect data on the speed, azimuth, and trajectory of obstacles. Infrared imagers are used to activate in low-light conditions and capture the thermal radiation characteristics of obstacles. Environmental sensors are used to collect data on road surface adhesion coefficient, light intensity, rainfall, and real-time vehicle status.

4. The collision avoidance warning and braking control system for transport vehicles based on AI vision fusion according to claim 1, characterized in that, The edge computing layer includes: A fast processing unit is used to perform lightweight neural network processing on single-channel visual images in multimodal data to obtain two-dimensional processed data; The precision processing unit is used to perform temporal and spatial alignment of binocular vision images, infrared thermal imaging images, and lidar point clouds to obtain three-dimensional processed data. The asynchronous fusion unit is used to correlate two-dimensional and three-dimensional processed data. When the three-dimensional processed data is delayed, the two-dimensional processed data is used for extrapolation, and the extrapolation result is updated immediately after the three-dimensional processed data is acquired.

5. The collision avoidance warning and braking control system for transport vehicles based on AI vision fusion according to claim 1, characterized in that, The decision control layer includes: The risk field establishment unit is used to establish a static obstacle risk field, a dynamic target prediction risk field, and a road structure risk field based on multi-source fusion data. The static obstacle risk field, the dynamic target prediction risk field, and the road structure risk field are then superimposed to obtain the dynamic risk field. The trajectory analysis unit is used to predict multiple short-term trajectories of the transport vehicle based on the current vehicle speed, steering wheel angle, and driver intention. Based on the dynamic risk field, it obtains the longitudinal predicted position of each short-term trajectory along the road centerline and the lateral predicted position perpendicular to the road centerline at any time, and determines the risk concentration value. For each short-term trajectory within the future time window, it multiplies the risk concentration value at each instant by the instantaneous vehicle speed, and continuously accumulates all instantaneous results to obtain the cumulative risk exposure. It also extracts the maximum risk concentration value and the time point of occurrence for each short-term trajectory. The vector building unit is used to obtain the rate of change of cumulative risk exposure and maximum concentration over time. It integrates driver status, vehicle status, multiple short-term trajectories, cumulative risk exposure, maximum concentration, as well as the time of occurrence and rate of change to obtain a risk status vector. The graded early warning unit is used to compare the risk status vector with the early warning rule threshold and determine the graded early warning based on the comparison result. The braking decision unit is used to make braking decisions based on a gradient braking control mechanism that matches the dynamic risk field when a graded warning indicates that the driver's intervention is insufficient or that there is not enough time to react.

6. The AI-based visual fusion-based collision avoidance warning and braking control system for transport vehicles according to claim 5, characterized in that, The risk field establishment unit includes: The static risk field establishment unit is used to assign basic hazard weights based on the physical properties of obstacles, assign dynamic hazard weights based on the specific characteristics of obstacles that change with environmental parameters, generate single obstacle risk impact areas based on basic hazard weights, dynamic hazard weights and obstacle characteristics, and superimpose all single obstacle risk impact areas to obtain the static obstacle risk field. The target risk field establishment unit is used to predict multiple future motion trajectories based on the trajectory of a dynamic target through a deep learning model, determine the probability of each future motion trajectory, establish a probability cloud of the dynamic target over time, generate a virtual repulsive force vector field proportional to the predicted kinetic energy of the dynamic target, convolve the virtual repulsive force vector field with the probability cloud, and combine it with the type gain of the dynamic target to generate a predicted risk field of the dynamic target. The road risk field establishment unit is used to obtain the road curvature and the real-time estimated vehicle lateral adhesion coefficient, determine the lateral instability risk, and generate an additional risk gradient on the outside of the curve that is proportional to the curvature and the square of the speed. Based on the lateral instability risk and the additional risk gradient, a quadratic attractive potential well is set on the lane centerline, and an exponential repulsive potential barrier is set on the lane line and the road boundary to obtain the road structure risk field. The risk field integration unit is used to define longitudinal distance, lateral offset and time with the center line of the road ahead as a reference, establish a risk field coordinate system, and configure the static obstacle risk field, dynamic target prediction risk field and road structure risk field into the risk field coordinate system to obtain the dynamic risk field.

7. A collision avoidance warning and braking control system for transport vehicles based on AI vision fusion according to claim 6, characterized in that, It also includes a risk field calibration unit, which performs real-time calibration of the dynamic risk field based on a three-level adaptive calibration model; The first layer of the three-level adaptive calibration model is based on vehicle physical state calibration: The vehicle load is acquired, and the dynamic risk field is calibrated in real time according to the rule that the risk area of ​​static obstacles expands but the remote influence range shrinks under heavy load. Obtain the road surface adhesion coefficient and, according to the rule that the same curve is more dangerous under low adhesion, calibrate the dynamic risk field in real time. In the three-level adaptive calibration model, the second level is calibration based on driver behavior profiles; Based on the driver's risk preference coefficient, reaction time model, and trust decay factor, the dynamic risk field is calibrated in real time. In the three-level adaptive calibration model, the third level is scenario-based dynamic decision calibration. Establish dynamic early warning threshold surface and dynamic braking threshold surface to perform real-time calibration of dynamic risk field.

8. A collision avoidance warning and braking control system for transport vehicles based on AI vision fusion according to claim 5, characterized in that, The tiered early warning unit includes: The Level 1 warning unit is used to determine a Level 1 warning when the cumulative risk exposure exceeds the dynamic baseline threshold determined based on driver profiles and road conditions, and the maximum concentration value is not in the emergency area. The Level 2 warning unit is used to determine a Level 2 warning when the maximum concentration value is greater than the primary warning threshold and the cumulative risk exposure shows an upward trend. The Level 3 warning unit is used to determine a Level 3 warning when the maximum concentration value exceeds the high-level warning threshold.

9. A collision avoidance warning and braking control system for transport vehicles based on AI vision fusion according to claim 5, characterized in that, The gradient braking control mechanism in the braking decision unit is specifically as follows: Pre-braking phase: Once braking arbitration is initiated, the braking system is immediately pre-filled; Proportional following phase: The target braking force is proportional to the gradient at the current moment and is limited by the current road adhesion coefficient and vehicle load; Full braking phase: If the risk increases sharply, enter full braking mode and apply the maximum safety braking force.

10. A collision avoidance warning and braking control system for transport vehicles based on AI vision fusion according to claim 1, characterized in that, The execution layer includes: The collision prediction unit is used to determine whether braking can avoid a collision based on the real-time vehicle dynamics and dynamic risk field. The parameter determination unit is used to generate steering assistance parameters based on real-time acquired lateral space data when collisions cannot be avoided by braking.

Images