An artificial intelligence-based in-vehicle occupant safety assessment eCall system and method
The eCall system, through multimodal perception and AI model scheduling, solves the problem that existing in-vehicle eCall systems cannot obtain the status inside the vehicle after an accident, realizes full-cycle safety monitoring and accurate rescue data reporting, and improves the intelligence and accuracy of the safety assessment of people inside the vehicle.
Patent Information
- Application Number
- CN202611162509.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-03
- Publication Date
- 2026-11-03
AI Technical Summary
Existing in-vehicle eCall systems are unable to obtain the actual status of people inside the vehicle, the extent of their injuries, and environmental risks after an accident. Their perception modes are limited and their adaptability is poor. They are unable to record driving behavior risks and in-vehicle safety status data before the accident, resulting in mismatched rescue plans and insufficient accuracy.
The system employs a multimodal perception unit, an AI model scheduling and safety assessment unit, an eCall communication unit, and an in-vehicle control unit to achieve multimodal perception, dynamic computing power scheduling, full-cycle safety monitoring and data reporting, generate standardized safety scores, and prioritize writing them into the extended fields of the MSD message.
It enables comprehensive and high-precision perception of the status of people inside the vehicle, improving the accuracy and targeting of rescue plans and significantly increasing the success rate of accident rescue.
Smart Images
Figure CN123025731A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of vehicle emergency call technology, specifically to an artificial intelligence-based eCall system and method for assessing the safety of occupants in a vehicle. Background Technology
[0002] The in-vehicle eCall emergency call system is a core component of a vehicle's passive safety protection system. Its core function is to automatically trigger the emergency call link after a collision and upload a minimum set of data (MSD) to a remote rescue service platform, providing basic data support for emergency rescue. With the rapid development of intelligent connected vehicle technology, the level of in-vehicle perception capabilities and artificial intelligence computing power has continuously improved, providing a technological foundation for the intelligent upgrade of eCall systems. However, existing mass-produced eCall systems still have many technical shortcomings in terms of perception capabilities, computing power scheduling, and the dimensions of rescue data, making it difficult to meet the industry's needs for refined and intelligent emergency rescue.
[0003] First, the existing eCall system's rescue information interaction heavily relies on post-accident human voice communication. When occupants are unconscious, in shock, mute, or trapped and unable to respond due to collision injuries, the rescue platform cannot obtain the actual status of the people inside the vehicle, the extent of their injuries, the presence of child occupants, or the risks posed by the in-vehicle environment. For example, when a severe collision causes the driver to lose consciousness, the existing system can only report basic data such as the vehicle's location and the number of occupants. The rescue platform cannot know if there are child occupants requiring priority rescue, nor can it determine if there are life-threatening secondary risks such as fire or dense smoke inside the vehicle. This can easily lead to mismatched rescue plans, delays in the golden rescue time, and a significant reduction in the success rate of accident rescue.
[0004] Secondly, adopting a multimodal approach can improve the stability and accuracy of single-sensor modes in conditions such as insufficient lighting, light pollution, camera obstruction, radar interference, and abnormal occupant posture. The multimodal fusion decision mechanism can dynamically allocate computing resources based on different sensing units and vehicle status to meet the diverse mass production needs of high-end models with multi-sensor fusion and low-end models with single-sensor deployment.
[0005] Furthermore, existing eCall technology only passively reports data after an accident occurs, and cannot record and store data on driving behavior risks and in-vehicle safety status before the accident. It lacks core pre-accelerated data for tracing the causes of accidents and predicting the injuries of the injured, resulting in a single dimension and insufficient accuracy in rescue assessment. It cannot provide rescue platforms with correlation analysis of risk factors such as driver fatigue and distracted driving before the accident. Summary of the Invention
[0006] The purpose of this application is to provide an AI-based eCall system and method for assessing in-vehicle occupant safety. This system addresses the technical problems of existing eCall systems, such as high reliance on voice interaction, unknown in-vehicle status during accidents, lack of prior risk data support, rigid computing power scheduling mechanisms, and single perception modes with poor adaptability. The system aims to achieve full-cycle in-vehicle safety monitoring, quantitative scoring, graded early warning, and automated reporting of multi-dimensional emergency data related to accidents, thereby improving the intelligence and precision of in-vehicle active safety control capabilities and passive emergency rescue.
[0007] This application provides an AI-based eCall system for assessing in-vehicle occupant safety, comprising: a multimodal perception unit, an AI model scheduling and safety assessment unit, an eCall communication unit, and an in-vehicle control unit;
[0008] The multimodal perception unit supports two working modes: single-modal independent perception and multimodal fusion perception. It is used to collect in-vehicle perception data and output preprocessed standardized perception data.
[0009] The AI model scheduling and safety assessment unit is used to dynamically schedule AI models and allocate computing resources according to the real-time operating conditions of the vehicle, quantify and generate in-vehicle safety scores based on the standardized perception data, and cyclically cache the safety score time series data within a preset time before the accident.
[0010] The eCall communication unit is used to write the pre-accident safety score time-series data and post-accident occupant status data into the extended field of the MSD minimum dataset message when the eCall is triggered by a vehicle collision, and to establish a voice call link after completing the data reporting first.
[0011] The in-vehicle controller unit is used to receive safety scores and graded warning instructions output by the AI model scheduling and safety assessment unit, and drive the in-vehicle actuators to achieve multi-level safety reminders.
[0012] In one possible implementation, the multimodal perception unit includes at least one of an in-cabin visual acquisition module, an in-vehicle radar module, and an audioprint acquisition module;
[0013] In single-modal independent sensing mode, any of the modules can be called independently to complete the sensing data acquisition.
[0014] In the multimodal fusion sensing mode, at least two of the modules can be linked to synchronously collect data and fuse and process it.
[0015] In one possible implementation, the AI model scheduling and safety assessment unit includes:
[0016] The model scheduling engine is used to execute corresponding computing resource allocation strategies and schedule corresponding AI models based on four types of operating conditions: vehicle power-on initialization, occupant change events, normal driving, and collision accidents.
[0017] The AI algorithm model group includes at least a driver status monitoring model and a passenger attribute monitoring model;
[0018] The safety assessment module is used to integrate the identification results of the driver state monitoring model and the occupant attribute monitoring model, and output the in-vehicle safety score according to a preset weighted algorithm;
[0019] The time-series caching module is used to store security score data within a preset time period in a cyclic overwrite manner.
[0020] In one possible implementation, the computing resource allocation strategy of the model scheduling engine includes:
[0021] When the vehicle is powered on and initialized, the occupant attribute monitoring model is scheduled to complete the archiving of information for all occupants.
[0022] When a crew change event is detected, the crew attribute monitoring model is triggered to perform full inference to update the crew data;
[0023] When the vehicle speed reaches a preset threshold, computing power is shifted to the driver state monitoring model, and the occupant attribute monitoring model operates at a reduced frequency.
[0024] Upon receiving a collision signal, all AI models are fully activated and all computing resources are released.
[0025] In one possible implementation, the AI algorithm model group further includes an in-vehicle environment recognition model and a voiceprint state analysis model;
[0026] The in-vehicle environment recognition model is used to identify the risks of fire, smoke, and foreign object intrusion inside the vehicle.
[0027] The voiceprint state analysis model is used to identify the occupant's cries for help, groans, and breathing state based on in-vehicle sound signals, and to help determine the occupant's state of consciousness and degree of injury.
[0028] In one possible implementation, the safety assessment module uses a multi-factor weighted algorithm, taking driver fatigue coefficient, driver distraction coefficient, occupant seat belt compliance rate, number of dangerous occupant sitting postures, and in-vehicle environment risk level as input factors to generate a standardized safety score, and outputs graded warning instructions based on preset scoring thresholds.
[0029] In one possible implementation, the data written by the eCall communication unit into the extended field of the MSD message includes at least: the average safety score before the accident, the attribute labels of each seat occupant, the occupant awareness level, the child occupant identification, the seat belt wearing status, the fire and smoke risk level inside the vehicle, and the risk of child safety seat use.
[0030] In one possible implementation, the system supports two mechanisms for determining the occupant's state of consciousness:
[0031] A single-modal independent determination mechanism can independently complete the determination based on any one of visual features, radar vital signs features, or voiceprint acoustic features.
[0032] The multimodal fusion judgment mechanism integrates features from at least two modalities for weighted decision-making and outputs tiered levels of consciousness, including conscious and comatose states.
[0033] This application also provides an AI-based eCall method for assessing in-vehicle occupant safety, applied to any of the systems described above, including:
[0034] After the vehicle is powered on and initialized, the occupant attribute monitoring model is invoked to complete the full identification of all occupants and establish an initial occupant status profile.
[0035] When a change in occupant status is detected, the occupant attribute monitoring model is triggered to update the occupant status data; when the vehicle speed reaches a preset threshold, the preset core computing power is dynamically allocated to the driver status monitoring model, and the occupant attribute monitoring model switches to inspection mode.
[0036] Real-time calculation of in-vehicle safety score; triggers graded safety warning when the score falls below the corresponding threshold.
[0037] When a collision signal is detected and an eCall is triggered, all AI models are activated to identify the occupant status and in-vehicle environmental risks after the collision. Safety score data cached before the accident is retrieved, fused to generate a rescue dataset, and written into the extended field of the MSD message. The MSD message is sent to the rescue platform first, and then a voice call link is established.
[0038] In one possible implementation, the information identified after a collision triggers an eCall includes the occupant's state of consciousness, the occupant's trapped state, and the risk of fire, smoke, or foreign object intrusion inside the vehicle.
[0039] Compared to existing technologies, the beneficial effects of the AI-based in-vehicle occupant safety assessment eCall system provided in this application are as follows:
[0040] 1. The multimodal sensing unit supports two working modes: single-modal independent sensing and multimodal fusion sensing. It can be flexibly adapted to the actual hardware configuration and working environment of the vehicle. It not only meets the low-cost deployment requirements of single sensor for low-end models, but also overcomes the defects of poor stability and insufficient accuracy of single sensing mode in scenarios such as insufficient lighting, occlusion, and abnormal posture through multimodal fusion, so as to achieve all-round and high-precision perception of the status of people in the vehicle.
[0041] 2. By dynamically scheduling AI models and allocating computing resources based on the real-time operating conditions of the vehicle through the AI model scheduling and safety assessment unit, the defects of the fixed computing power allocation mode of the traditional in-cabin perception system are overcome. It can realize the on-demand allocation of computing resources under different operating conditions of the vehicle, and meet the automotive-grade deployment requirements of normal monitoring accuracy, low power consumption operation and ultra-high reliability in emergency scenarios under the limited computing power of the vehicle embedded platform.
[0042] 3. By quantitatively generating in-vehicle safety scores through safety assessment units and cyclically caching the time-series data of safety scores within a preset time period before an accident, the system solves the technical problem that the existing eCall system cannot record driving behavior risks and in-vehicle safety status data before an accident, providing core pre-data support for tracing the causes of accidents and predicting the injuries of the injured.
[0043] 4. When a vehicle collision triggers an eCall, the eCall communication unit writes the pre-accident safety score timing data and post-accident occupant status data into the MSD message extension field. After prioritizing data reporting, a voice call link is established. This completely solves the industry pain point of traditional eCall systems where rescue information is completely blank in scenarios where the occupant is unconscious, deaf, or unable to answer the call. This allows the rescue platform to fully grasp the core information such as the status of the people in the vehicle, environmental risks, and pre-accident risks before the voice call is connected, greatly improving the accuracy and targeting of the rescue plan and significantly increasing the success rate of accident rescue. Attached Figure Description
[0044] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:
[0045] Figure 1 This paper illustrates a schematic diagram of the structure of an artificial intelligence-based eCall system for assessing in-vehicle occupants safety, provided in an embodiment of this application.
[0046] Figure 2 A flowchart of an AI-based eCall method for assessing in-vehicle occupant safety is shown in an embodiment of this application. Detailed Implementation
[0047] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.
[0048] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains.
[0049] Furthermore, the terms "first" and "second," etc., are used to distinguish different objects, not to describe a specific order. Additionally, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to those processes, methods, products, or apparatuses.
[0050] Please refer to Figure 1 , Figure 1 This is a schematic diagram of the structure of an artificial intelligence-based in-vehicle occupant safety assessment eCall system provided in an embodiment of this application. The system includes: a multimodal perception unit 100, an AI model scheduling and safety assessment unit 200, an eCall communication unit 300, and an in-vehicle control unit 400.
[0051] The multimodal perception unit 100 supports two working modes: single-modal independent perception and multimodal fusion perception. It is used to collect in-vehicle perception data and output preprocessed standardized perception data.
[0052] Specifically, the multimodal perception unit 100 includes at least one of an in-cabin vision acquisition module, an in-vehicle radar module, and a voiceprint acquisition module. In a single-modal independent perception mode, any one of the modules can be called independently to complete the perception data acquisition. For example, the in-cabin vision acquisition module can be called independently to acquire images of the in-vehicle environment. In a multimodal fusion perception mode, at least two of the modules can be linked to collect data synchronously and process the data. For example, the in-cabin vision acquisition module, the in-vehicle radar module, and the voiceprint acquisition module can be called to collect data, or only the in-cabin vision acquisition module and the voiceprint acquisition module can be called to collect data.
[0053] Furthermore, the system provided in this application supports two mechanisms for determining the occupant's state of consciousness based on two working modes of the multimodal perception unit 100: 1. A single-modal independent determination mechanism, which independently completes the determination based on any one of visual features, radar vital signs features, or voiceprint acoustic features; 2. A multimodal fusion determination mechanism, which integrates at least two modal features for weighted decision-making and outputs multiple levels of consciousness such as conscious and comatose.
[0054] The AI model scheduling and safety assessment unit 200 is used to dynamically schedule AI models and allocate computing resources according to the real-time operating conditions of the vehicle, quantify and generate in-vehicle safety scores based on the standardized perception data, and cyclically cache the safety score time series data within a preset time before the accident.
[0055] Specifically, the AI model scheduling and security assessment unit 200 includes: a model scheduling engine 201, an AI algorithm model group 202, a security assessment module 203, and a time-series cache module 204;
[0056] Model scheduling engine 201 is used to execute corresponding computing resource allocation strategies and schedule corresponding AI models based on four types of operating conditions: vehicle power-on initialization, occupant change events, normal driving, and collision accidents.
[0057] The computing resource allocation strategy of the model scheduling engine includes:
[0058] 1. When the vehicle is powered on and initialized, the occupant attribute monitoring model is scheduled to complete the archiving of all occupant information;
[0059] 2. When a crew change event is detected, the crew attribute monitoring model is triggered to perform full inference to update the crew data;
[0060] 3. When the vehicle speed reaches a preset threshold, computing power is shifted to the driver state monitoring model, and the occupant attribute monitoring model operates at a reduced frequency.
[0061] 4. Upon receiving a collision signal, fully wake up all AI models and release all computing resources.
[0062] AI algorithm model 202 includes at least a driver status monitoring model and a passenger attribute monitoring model;
[0063] In some embodiments, the AI algorithm model group 202 may further include an in-vehicle environment recognition model and a voiceprint state analysis model; the in-vehicle environment recognition model is used to identify the risk of fire, smoke, and foreign object intrusion inside the vehicle; the voiceprint state analysis model is used to identify the occupant's cries for help, groans, and breathing state based on in-vehicle sound signals, and to assist in determining the occupant's state of consciousness and degree of injury.
[0064] The safety assessment module 203 is used to integrate the identification results of the driver state monitoring model and the occupant attribute monitoring model, and output the in-vehicle safety score according to a preset weighted algorithm;
[0065] Specifically, the safety assessment module 203 adopts a multi-factor weighted algorithm, using driver fatigue coefficient, driver distraction coefficient, occupant seat belt compliance rate, number of dangerous occupant sitting postures, and in-vehicle environment risk level as input factors to generate a standardized safety score, and outputs graded warning instructions based on preset scoring thresholds.
[0066] The time-series cache module 204 is used to store security score data within a preset time period in a cyclic overwrite manner.
[0067] The eCall communication unit 300 is used to write the pre-accident safety score time-series data and post-accident occupant status data into the extended field of the MSD minimum dataset message when the eCall is triggered by a vehicle collision, and to establish a voice call link after completing the data reporting first.
[0068] Specifically, the data written by the eCall communication unit 300 into the extended field of the MSD message includes at least: the average safety score before the accident, the attribute labels of each seat occupant, the occupant awareness level, the child occupant identification, the seat belt wearing status, the fire and smoke risk level inside the vehicle, and the risk of child safety seat use.
[0069] The in-vehicle controller unit 400 is used to receive safety scores and graded warning instructions output by the AI model scheduling and safety assessment unit 200, and drive the in-vehicle actuators to realize multi-level safety reminders.
[0070] This application provides an AI-based eCall method for assessing in-vehicle occupant safety, applied to the aforementioned AI-based eCall system for assessing in-vehicle occupant safety. Figure 2 As shown, the method includes the following steps:
[0071] S101. After the vehicle is powered on and initialized, the occupant attribute monitoring model is called to complete the full identification of all occupants in the vehicle and establish an initial occupant status file.
[0072] S102. When an occupant change event is detected, the occupant attribute monitoring model is triggered to update the occupant status data;
[0073] S103. After the vehicle speed reaches the preset threshold, the preset core computing power is allocated to the driver status monitoring model, and the occupant attribute monitoring model is switched to inspection mode.
[0074] S104. Calculate the in-vehicle safety score in real time, and trigger a graded safety warning when the score is lower than the corresponding threshold.
[0075] S105. When a collision signal is detected and eCall is triggered, all AI models are activated to identify the occupant status and in-vehicle environmental risks after the collision. Safety score data cached before the accident is retrieved, fused to generate a rescue dataset and written into the extended field of the MSD message. The MSD message is sent to the rescue platform first before establishing a voice call link.
[0076] In some embodiments, the information identified after a collision triggers an eCall includes the occupant's state of consciousness, the occupant's state of being trapped in a crush, and the risk of fire, smoke, or foreign object intrusion inside the vehicle.
[0077] The above system and method are further described below with a specific embodiment.
[0078] In this embodiment, the system hardware is deployed on an automotive-grade AI domain controller, equipped with an in-cabin infrared high-definition camera acquisition device, an in-vehicle high-precision millimeter-wave radar, and an on-board noise-canceling microphone array. The communication end is adapted to a 4G / 5G on-board eCall communication module, and is compatible with the onboard bus architectures of all mainstream passenger and commercial vehicles. The system supports two perception modes: single-modal independent operation and multi-modal fusion operation, which can adaptively switch according to the actual hardware configuration of the vehicle.
[0079] After the vehicle is powered on (ACC initialization), the system automatically completes sensor self-checks, parameter calibration, AI model preloading, and vehicle bus signal integration. The model scheduling engine 201 prioritizes the occupant attribute monitoring model to run at full speed, performing comprehensive identification of each seat in the vehicle, including occupant count, adult and child attribute differentiation, child safety seat installation status recognition, and seat belt wearing status recognition. For example, if the system detects that a child safety seat is installed in the left rear seat and a child of approximately 3 years old is sitting there, and that an adult is sitting in the front passenger seat wearing a seat belt, it generates an initial occupant profile and stores it locally with encryption.
[0080] During vehicle operation, the system monitors the vehicle's CAN bus signals in real time. When it detects any door changing from open to closed, a change in seatbelt locking / unlocking status, or a change in seat occupancy sensor signals, it immediately triggers the occupant attribute monitoring model to perform a full inference, updating the occupant status data in real time to ensure the real-time nature and accuracy of occupant information. For example, if the vehicle stops to pick up someone, and the rear right-side door opens and then closes, the system detects this door closure event and immediately triggers the occupant attribute monitoring model to perform a full identification. If it finds a new adult occupant in the rear right-side seat who is not wearing a seatbelt, the system updates this information to the occupant profile and inputs the risk factor of not wearing a seatbelt into the safety assessment module 203.
[0081] When the vehicle speed reaches a preset threshold (e.g., 10km / h), the model scheduling engine 201 allocates the main NPU computing resources to the driver status monitoring model to continuously monitor dangerous states such as driver fatigue, distraction, looking down, and line of sight deviation; the occupant attribute monitoring model switches to inspection mode, for example, reducing the detection frequency from every frame to once every 30 seconds, taking into account both the overall vehicle driving safety and the low power consumption operation requirements of the equipment.
[0082] The safety assessment module 203 collects output data from the driver status monitoring model, occupant attribute monitoring model, and in-vehicle environment recognition model in real time, and uses a multi-factor weighted algorithm to calculate a standardized in-vehicle safety score in real time. For example, the weights of each factor are set as follows: driver fatigue coefficient 30%, driver distraction coefficient 25%, occupant seat belt compliance rate 20%, number of dangerous occupant sitting postures 15%, and in-vehicle environment risk level 10%, to comprehensively calculate an in-vehicle safety score from 0 to 100. The system sets up a three-level warning mechanism: a score of 80 or above is a safe state with no warning output; a score of 60 to 80 is a level 1 minor risk, triggering a text prompt on the vehicle screen, such as displaying "Please stay focused on driving" on the central control screen; a score of 40 to 60 is a level 2 moderate risk, triggering a voice warning, such as playing a "Please pay attention to the road conditions ahead" prompt through the vehicle audio system; a score below 40 is a level 3 high risk, simultaneously triggering a voice warning and a driver's seat vibration reminder, realizing tiered safety intervention. Meanwhile, the time-series caching module 204 continuously caches the security score data and peak risk events within 30 seconds before the collision in a cyclic overwrite manner.
[0083] When a vehicle collision occurs and the airbags are deployed, the vehicle collision detection unit outputs a collision signal to activate the eCall system, and the system immediately enters emergency mode. The model scheduling engine 201 releases all computing resources, fully waking up all AI models to perform high-precision recognition and inference. The system adaptively selects the perception mode based on the vehicle's hardware configuration: if the vehicle is equipped with complete multimodal perception hardware, a multimodal fusion judgment mechanism is activated, integrating visual, radar, and voiceprint features for weighted decision-making to determine the consciousness status of each occupant, outputting multiple status levels such as conscious and unconscious; if the vehicle is equipped with only a single visual perception module, a single-modal independent judgment mechanism is activated, independently determining the consciousness status based on visual facial and limb features. Simultaneously, the in-vehicle environment recognition model frequently detects whether there are dangerous environments such as flames, thick smoke, or foreign object intrusion inside the vehicle, while the occupant attribute monitoring model identifies whether occupants are trapped or crushed.
[0084] The system retrieves the average safety score and peak risk event data for the 30 seconds prior to the collision stored in the timing cache module 204. This data is then fused with real-time post-collision data on occupant consciousness, occupant entrapment, and in-vehicle environmental risks to generate a standardized, structured rescue auxiliary dataset. The eCall communication unit 300 writes this rescue dataset as an extended field into an MSD message conforming to national or European standards. For example, the extended field might include the following data: an average safety score of 45 points for the 30 seconds prior to the accident; the driver's seat occupant's consciousness is impaired after the accident; the front passenger seat occupant is conscious; the child occupant on the left rear seat is conscious and their child safety seat is properly installed; there are no occupants on the right rear seat; and there is no fire or smoke inside the vehicle. After message encapsulation, the eCall communication unit 300 prioritizes transmitting this MSD message carrying the extended data to the rescue service platform. After message transmission, an emergency voice call link is automatically established. This allows rescue personnel to fully grasp the occupant status, environmental risks, and pre-accident risks before the voice call is connected, enabling precise rescue deployment.
[0085] It is worth mentioning that the system and method of the present invention can flexibly adapt the perception mode according to the actual hardware configuration of the vehicle, taking into account the high precision requirements of multi-sensor fusion for high-end models and the low-cost deployment requirements of single-sensor deployment for low-end models, and has excellent automotive-grade engineering feasibility.
[0086] The beneficial effects of the AI-based eCall system and method for in-vehicle occupant safety assessment provided in this application are as follows:
[0087] 1. This invention supports two working modes through a multimodal sensing unit: single-modal independent sensing and multimodal fusion sensing. It can be flexibly adapted to the actual hardware configuration and working environment of the vehicle. It not only meets the low-cost deployment requirements of single sensors for low-end models, but also overcomes the defects of poor stability and insufficient accuracy of single sensing modes in scenarios such as insufficient lighting, occlusion, and abnormal posture through multimodal fusion, so as to achieve all-round and high-precision perception of the status of people in the vehicle.
[0088] 2. This invention uses an AI model scheduling and safety assessment unit to dynamically schedule AI models and allocate computing resources according to the real-time operating conditions of the vehicle. This overcomes the shortcomings of the fixed computing power allocation mode of traditional in-cabin perception systems. It can realize the on-demand allocation of computing resources under different operating conditions of the vehicle. Under the limited computing power of the vehicle embedded platform, it can meet the automotive-grade deployment requirements of normal monitoring accuracy, low power consumption operation and ultra-high reliability in emergency scenarios.
[0089] 3. This invention generates in-vehicle safety scores through AI model scheduling and safety assessment units, and cyclically caches the time-series data of safety scores within a preset time period before an accident. This solves the technical problem that the existing eCall system cannot record driving behavior risks and in-vehicle safety status data before an accident, and provides core pre-data support for tracing the causes of accidents and predicting the injuries of the injured.
[0090] 4. This invention, through the eCall communication unit, writes the pre-accident safety score timing data and post-accident occupant status data into the MSD message extension field when the eCall is triggered by a vehicle collision. After prioritizing data reporting, a voice call link is established. This completely solves the industry pain point of traditional eCall systems where rescue information is completely blank in scenarios where the occupant is unconscious, deaf, or unable to answer the call. This allows the rescue platform to fully grasp the core information such as the status of the people in the vehicle, environmental risks, and pre-accident risks before the voice connection is established, greatly improving the accuracy and targeting of the rescue plan and significantly increasing the success rate of accident rescue.
[0091] 5. This invention receives safety scores and graded early warning instructions through the in-vehicle controller unit and drives the in-vehicle actuators to achieve multi-level safety reminders. It constructs a fully closed-loop vehicle safety system from routine risk monitoring and in-vehicle graded early warning to intelligent accident assessment and precise rescue reporting. It simultaneously strengthens the vehicle's active safety risk prevention and control capabilities and passive emergency rescue capabilities, and comprehensively improves the intelligence and standardization level of the vehicle safety system.
[0092] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application, and they should all be covered within the scope of the claims and specification of this application.
Claims
1. An AI-based eCall system for assessing in-vehicle occupant safety, characterized in that: include: Multimodal perception unit, AI model scheduling and safety assessment unit, eCall communication unit and in-vehicle control unit; The multimodal perception unit supports two working modes: single-modal independent perception and multimodal fusion perception. It is used to collect in-vehicle perception data and output preprocessed standardized perception data. The AI model scheduling and safety assessment unit is used to dynamically schedule AI models and allocate computing resources according to the real-time operating conditions of the vehicle, quantify and generate in-vehicle safety scores based on the standardized perception data, and cyclically cache the safety score time series data within a preset time before the accident. The eCall communication unit is used to write the pre-accident safety score time-series data and post-accident occupant status data into the extended field of the MSD minimum dataset message when the eCall is triggered by a vehicle collision, and to establish a voice call link after completing the data reporting first. The in-vehicle controller unit is used to receive safety scores and graded warning instructions output by the AI model scheduling and safety assessment unit, and drive the in-vehicle actuators to achieve multi-level safety reminders.
2. The system according to claim 1, characterized in that, The multimodal perception unit includes at least one of an in-cabin visual acquisition module, an in-vehicle radar module, and an audioprint acquisition module. In single-modal independent sensing mode, any of the modules can be called independently to complete the sensing data acquisition. In the multimodal fusion sensing mode, at least two of the modules can be linked to synchronously collect data and fuse and process it.
3. The system according to claim 1, characterized in that, The AI model scheduling and security assessment unit includes: The model scheduling engine is used to execute corresponding computing resource allocation strategies and schedule corresponding AI models based on four types of operating conditions: vehicle power-on initialization, occupant change events, normal driving, and collision accidents. The AI algorithm model group includes at least a driver status monitoring model and a passenger attribute monitoring model; The safety assessment module is used to integrate the identification results of the driver state monitoring model and the occupant attribute monitoring model, and output the in-vehicle safety score according to a preset weighted algorithm; The time-series caching module is used to store security score data within a preset time period in a cyclic overwrite manner.
4. The system according to claim 3, characterized in that, The computing resource allocation strategy of the model scheduling engine includes: When the vehicle is powered on and initialized, the occupant attribute monitoring model is scheduled to complete the archiving of information for all occupants. When a crew change event is detected, the crew attribute monitoring model is triggered to perform full inference to update the crew data; When the vehicle speed reaches a preset threshold, computing power is shifted to the driver state monitoring model, and the occupant attribute monitoring model operates at a reduced frequency. Upon receiving a collision signal, all AI models are fully activated and all computing resources are released.
5. The system according to claim 3, characterized in that, The AI algorithm model group also includes an in-vehicle environment recognition model and a voiceprint status analysis model; The in-vehicle environment recognition model is used to identify the risks of fire, smoke, and foreign object intrusion inside the vehicle. The voiceprint state analysis model is used to identify the occupant's cries for help, groans, and breathing state based on in-vehicle sound signals, and to help determine the occupant's state of consciousness and degree of injury.
6. The system according to claim 3, characterized in that, The safety assessment module uses a multi-factor weighted algorithm, taking driver fatigue coefficient, driver distraction coefficient, occupant seat belt compliance rate, number of dangerous occupant sitting postures, and in-vehicle environment risk level as input factors to generate a standardized safety score, and outputs graded warning instructions based on preset scoring thresholds.
7. The system according to claim 1, characterized in that, The data written into the extended field of the MSD message by the eCall communication unit includes at least: the average safety score before the accident, the attribute labels of each occupant, the occupant's awareness level, the child occupant identification, the seat belt wearing status, the risk level of fire and smoke inside the vehicle, and the risk of child safety seat use.
8. The system according to any one of claims 1 to 7, characterized in that, The system supports two mechanisms for determining the occupant's state of consciousness: A single-modal independent determination mechanism can independently complete the determination based on any one of visual features, radar vital signs features, or voiceprint acoustic features. The multimodal fusion judgment mechanism integrates features from at least two modalities for weighted decision-making and outputs graded levels of consciousness, including conscious and comatose states.
9. An AI-based eCall method for assessing in-vehicle occupant safety, applied to the system described in any one of claims 1 to 8, characterized in that, include: After the vehicle is powered on and initialized, the occupant attribute monitoring model is invoked to complete the full identification of all occupants and establish an initial occupant status profile. When an occupant change event is detected, the occupant attribute monitoring model is triggered to update the occupant status data; Once the vehicle speed reaches a preset threshold, the preset core computing power is dynamically allocated to the driver status monitoring model, and the occupant attribute monitoring model is switched to inspection mode. Real-time calculation of in-vehicle safety score; triggers graded safety warning when the score falls below the corresponding threshold. When a collision signal is detected and an eCall is triggered, all AI models are activated to identify the occupant status and in-vehicle environmental risks after the collision. Safety score data cached before the accident is retrieved, fused to generate a rescue dataset, and written into the extended field of the MSD message. The MSD message is sent to the rescue platform first, and then a voice call link is established.
10. The method according to claim 9, characterized in that, The information identified after a collision triggers an eCall includes the occupants' state of consciousness, their trapped or crushed state, and the risks of fire, smoke, and foreign object intrusion inside the vehicle.