Intelligent safety belt hook detection method and device, intelligent safety belt and medium

Abstract

The invention discloses an intelligent safety belt hook detection method and device based on artificial intelligence, an intelligent safety belt and a medium, and relates to the technical field of intelligent safety belts. The distance parameter is a distance value of the distance measuring sensor relative to the measured object, and the object category identification parameter comprises a parameter used for identifying the material of the measured object; when the distance parameter is smaller than a preset distance threshold value, inputting the distance parameter and the object category recognition parameter into a pre-trained object category recognition model, so that the object category recognition model outputs a material recognition result of the measured object; and determining whether the hook of the intelligent safety belt is hung back to the human body or not according to the material identification result and the preset material category. The real-time response speed and the detection real-time performance are effectively improved, whether the hook of the intelligent safety belt is correctly hung back to the human body or not can be effectively judged, and the detection accuracy is improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent seat belt technology, and in particular to an artificial intelligence-based intelligent seat belt hook detection method, detection device, intelligent seat belt, and medium. Background Technology

[0002] In high-altitude operations, the hook of the smart safety belt is a key component to ensure the safety of workers. Due to insufficient safety awareness or operating habits, some workers may mistakenly hook the hook to themselves instead of the designated suspension point, causing the hook to lose its due safety protection function and exposing workers to significant safety risks.

[0003] Currently, most common smart safety belt detection solutions on the market rely on visual recognition technology. This involves installing a camera on the hook, capturing images of the hook, and performing real-time image processing to determine whether the hook is attached to the suspension point or hooked back onto the worker. However, this type of solution requires a large amount of data input and processing, resulting in insufficient real-time detection capabilities.

[0004] Therefore, improving the real-time detection capability of smart seat belts is a technical problem that urgently needs to be solved. Summary of the Invention

[0005] This invention provides an artificial intelligence-based intelligent seat belt hook detection method, detection device, intelligent seat belt, and medium to solve the problem that the large amount of data input and data processing in existing intelligent seat belt detection schemes leads to insufficient real-time detection.

[0006] Firstly, an artificial intelligence-based intelligent seatbelt hook detection method is provided, including: The measurement parameters returned by the ranging sensor are obtained. The measurement parameters include distance parameters and object category identification parameters. The distance parameters are the distance values ​​of the ranging sensor relative to the measured object. The object category identification parameters include parameters for identifying the material of the measured object. When the distance parameter is less than a preset distance threshold, the distance parameter and the object category recognition parameter are input into a pre-trained object category recognition model so that the object category recognition model outputs the material recognition result of the tested object. Based on the material identification results and the preset material category, determine whether the hook of the smart safety belt should be hooked back onto the human body.

[0007] In a second aspect, an intelligent seat belt detection device is provided, comprising a distance sensor and a controller, wherein the controller is connected to the distance sensor and is used to implement the steps of the artificial intelligence-based intelligent seat belt hook detection method described in the first aspect above.

[0008] Thirdly, a smart seat belt is provided, the smart seat belt including a smart seat belt detection device and a seat belt hook as described in the second aspect, the smart seat belt detection device being disposed on the seat belt hook.

[0009] Fourthly, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the artificial intelligence-based intelligent seatbelt hook detection method described in the first aspect above.

[0010] The present invention provides a technical solution with the following advantages: By acquiring the measurement parameters of the object being measured returned by the ranging sensor, i.e., acquiring the distance parameters and object category recognition parameters; then, when the distance parameters are less than a preset distance threshold, the distance parameters and object category recognition parameters are input into a pre-trained object category recognition model to output the material recognition result of the object being measured. That is, the material recognition result of the object being measured is only identified by the object category recognition model when the distance parameters between the ranging sensor and the object being measured are less than the preset distance threshold, effectively reducing unnecessary data processing and improving the resource utilization of the processor; furthermore, the present invention only inputs the distance parameters and object category recognition parameters into the pre-trained object category recognition model for recognition, which significantly reduces the amount of data input and data processing of the object category recognition model. Compared with other solutions, this effectively reduces the processor load, improves the real-time response speed, and thus improves the real-time performance of the detection; next, the material recognition result is matched with the preset material category to determine whether the hook of the smart seat belt is hooked back onto the human body, which can effectively determine whether the hook of the smart seat belt is correctly hooked back onto the human body, effectively avoiding safety hazards caused by the hook being mistakenly hooked to a non-suspension point, and improving the detection accuracy. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a schematic diagram of an intelligent seatbelt hook detection device based on artificial intelligence in one embodiment of the present invention; Figure 2 This is a flowchart of an intelligent seatbelt hook detection method based on artificial intelligence in one embodiment of the present invention; Figure 3 This is a training effect diagram of an object category recognition model in one embodiment of the present invention; Figure 4 This is a data processing flowchart of an object category recognition model according to an embodiment of the present invention; Figure 5 This is another data processing flowchart of the object category recognition model in one embodiment of the present invention; Figure 6 This is a schematic diagram of a smart seat belt according to an embodiment of the present invention. Detailed Implementation

[0013] To make the technical problems solved, the technical solutions, and the beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0014] Existing technology First, in high-altitude operation scenarios, the installation height and angle of the camera change dynamically with the working environment, resulting in unstable image acquisition perspectives. This means that code logic design, artificial intelligence model architecture design, and the construction of training data for artificial intelligence models all need to cover a massive number of complex perspectives. This not only significantly increases the development difficulty, but also leads to low accuracy and poor real-time performance of the artificial intelligence model in judging the hook state due to insufficient data diversity.

[0015] Secondly, visual recognition technology relies on a large amount of image data input (taking 640×480 standard definition resolution as an example, the data volume of a single frame image is about 300,000 pixels), which puts extremely demanding requirements on the controller's computing power, storage capacity and hardware cost. Lightweight devices (such as intelligent seat belt detection devices) often cannot complete the calculation in time under such conditions, making it difficult to meet the real-time alarm requirements in dynamic working environments and restricting the reliability in dynamic working environments.

[0016] Thirdly, in high-altitude working environments, weather conditions (such as sunny days and cloudy days) are highly variable, which can affect the illumination distribution, contrast, and noise level of images, further reducing the environmental robustness of visual recognition and resulting in insufficient reliability of detection results.

[0017] In summary, existing visual recognition technologies not only have high development costs and huge data requirements in high-altitude operation scenarios, but also suffer from a significant drop in accuracy under extreme environments such as low visible light and sudden changes in viewing angle, failing to meet the rigid requirements of smart safety belts for "low power consumption, lightweight, all-weather, and real-time performance".

[0018] Explanation of selecting the range sensor return parameters as the core input As mentioned above, traditional visual recognition technology relies on image data (300,000 pixels per frame), which has core defects such as data explosion, insufficient real-time performance, and poor environmental robustness (e.g., the camera's angle changes frequently and the lighting changes drastically during high-altitude operations, leading to recognition failure). If the visual recognition path is followed, those skilled in the art will fall into a vicious cycle of increasing resolution and computing power, which leads to soaring costs and cannot meet the essential requirements of smart safety belts for low power consumption and real-time alarms.

[0019] Unlike visual recognition, the inventors discovered that the core criteria for determining whether a hook has been reattached to a person can be summarized into two factors: "distance" and "material." For example, when a hook is reattached to a person, the spatial distance between the hook and the person must be less than a preset distance threshold (e.g., less than 30 centimeters), and the material of the nearby objects is clothing. In other words, as long as the material of the object being tested adjacent to the hook is accurately determined, it is possible to accurately determine whether the hook has been reattached to a person without relying on massive amounts of image data.

[0020] Based on this idea, the inventors discovered that material can be indirectly characterized through the measurement parameters of a ranging sensor. For example, the material of the object can be accurately characterized by learning the correlation between the measurement parameters and the material of the object through a model based on distance parameters and object category recognition parameters. For example, different materials have different reflectivities (e.g., the reflectivity of skin is approximately 0.45±0.1, and the reflectivity of metal is approximately 0.8±0.05), and the object category recognition model can accurately analyze the material of the object based on the distance parameters and object category recognition parameters corresponding to different objects. For example, the object category recognition model can analyze the reflectivity of the object from the distance parameters and object category recognition parameters, and then determine the material of the object based on the reflectivity. It should be understood that this is only an example and does not limit the specific model implementation process. Furthermore, the inventors proved through experiments that the above measurement parameters are sufficient to distinguish materials, breaking the technical prejudice in the field that materials cannot be identified without images, overturning the conventional understanding that detection must rely on vision, and opening up a new path for parameter-driven detection. This is also the logical starting point that distinguishes this invention from the prior art.

[0021] Terminology Explanation In this invention, the object category recognition model can be a parameter-driven classification model based on artificial intelligence, such as a model that takes measurement parameters as input data and material (skin, clothing, metal, etc.) as output results. The size of the object category recognition model can be controlled to be small (e.g., 10 Kbytes), and it can be directly deployed on the controller (MCU) of the smart seat belt detection device without the need for a high-performance GPU. It can complete the judgment in a very short time (millisecond level), and has real-time performance.

[0022] In this invention, the distance parameter refers to the spatial distance between the ranging sensor and the object being measured, and the unit is millimeters (mm).

[0023] In this invention, ambient rate refers to the rate of ambient signal caused by noise, and is used to assess external optical interference.

[0024] In this invention, the number of reflected photons (c.signal) refers to the total number of photons in the laser echo signal collected by the ranging sensor, which is used to reflect the surface reflection characteristics of the object being measured.

[0025] In this invention, the number of SPADs enabled refers to the number of SPADs currently enabled by the ranging sensor and participating in laser echo signal detection, which is the photon emitting device of the laser ranging sensor.

[0026] In this invention, the distance measurement noise estimate (sigma) refers to the estimated value of noise in the distance of the measured object.

[0027] Instructions for Smart Seatbelt Detection Device In this invention, such as Figure 1 As shown, the intelligent seatbelt detection device 100 includes a ranging sensor 101 and a controller 102 (MCU, Micro Controller Unit). The ranging sensor 101 and the controller 102 are mounted on a printed circuit board (PCB). The ranging sensor 101 is connected to the controller 102, and the controller 102 stores and deploys an object category recognition model to implement the steps of the artificial intelligence-based intelligent seatbelt hook detection method. The object category recognition model is a model pre-trained on a server. The ranging sensor 101 includes, but is not limited to, laser ranging sensors and other ranging sensors.

[0028] This invention provides an artificial intelligence-based intelligent seatbelt hook detection method, comprising: acquiring measurement parameters returned by a ranging sensor, the measurement parameters including distance parameters and object category recognition parameters, the distance parameters being the distance value of the ranging sensor relative to the measured object, and the object category recognition parameters including parameters for identifying the material of the measured object; when the distance parameters are less than a preset distance threshold, inputting the distance parameters and object category recognition parameters into a pre-trained object category recognition model, so that the object category recognition model outputs the material recognition result of the measured object; and determining whether the hook of the intelligent seatbelt is hooked back onto the human body based on the material recognition result and the preset material category.

[0029] Compared to the closest prior art, the present invention achieves the following results: by acquiring the measurement parameters of the object being measured returned by the ranging sensor, i.e., acquiring the distance parameters and object category recognition parameters; then, when the distance parameters are less than a preset distance threshold, the distance parameters and object category recognition parameters are input into a pre-trained object category recognition model to output the material recognition result of the object being measured. In other words, the material recognition result of the object being measured is only identified by the object category recognition model when the distance parameters between the ranging sensor and the object being measured are less than the preset distance threshold, effectively reducing unnecessary data processing and improving processor resource utilization. Furthermore, this invention only inputs the distance parameters and object category... The identification parameters are input into a pre-trained object category recognition model, which significantly reduces the amount of data input and processing required for the object category recognition model. Compared with other solutions, this effectively reduces the processor load and improves the real-time response speed, thereby improving the real-time performance of the detection. Next, the material identification results are matched with the preset material categories to determine whether the hook of the smart safety belt is back on the human body. This can effectively determine whether the hook of the smart safety belt is correctly back on the human body, effectively avoiding safety hazards caused by the hook being mistakenly attached to a non-suspension point, improving detection accuracy, and meeting the rigid requirements of smart safety belts for "low power consumption, lightweight, all-weather, and real-time performance".

[0030] Please refer to the following as well. Figures 1 to 5 This invention provides an intelligent seatbelt hook detection method based on artificial intelligence.

[0031] Example 1 Firstly, such as Figure 2 As shown, an artificial intelligence-based intelligent seatbelt hook detection method is provided for application in [the following context is missing from the original text]. Figure 1 Taking controller 102 as an example, the steps include the following: S201: Obtain the measurement parameters returned by the ranging sensor. The measurement parameters include distance parameters and object category recognition parameters. The distance parameters are the distance values ​​between the ranging sensor and the measured object. The object category recognition parameters include parameters used to identify the material of the measured object.

[0032] In this embodiment, the distance parameter is the distance value between the ranging sensor and the object being measured.

[0033] In this embodiment, the object category recognition parameters include parameters for identifying the material of the object being tested, which are used for input analysis of the subsequent object category recognition model.

[0034] As an example, when the intelligent seatbelt detection device is determined to be at high altitude, the ranging sensor can be controlled to enter working mode, so that the ranging sensor automatically acquires the number of activated single-photon avalanche diodes and emits a laser; then, the laser echo signal is received and the flight time from the emission of the test laser to the receipt of the laser echo signal is recorded; wherein, the laser echo signal is the signal reflected by the test laser upon contact with the measured object; then, the distance parameter is calculated based on the flight time, for example, the corresponding distance parameter is calculated according to the formula, namely: ;in, For the speed of light ( ),For example, When it is 100ns, The distance is 15cm; then, the ranging sensor extracts the environmental noise signal from the laser echo signal, and obtains the environmental noise signal rate by counting the number of noise signals per unit time (i.e., the frequency of noise pulse occurrence). For example, if 30 noise pulses are detected within 100ms, the environmental signal rate is... Next, the number of effectively reflected photons in the laser echo signal is counted to obtain the number of reflected photons. For example, if 500 effective photons are extracted from the laser echo signal using a photon counting algorithm, then the number of reflected photons is 500. Then, based on the intensity distribution characteristics of the laser echo signal, the theoretical distance measurement variance is calculated. Based on the theoretical distance measurement variance and the number of single-photon avalanche diodes, the distance measurement noise estimate is obtained. For example, the intensity distribution of the laser echo signal is fitted with a Gaussian to obtain the fitting standard deviation. (corresponding to time noise), and then convert the fitting standard deviation (time noise) into the theoretical distance variance based on the speed of light. For example, the speed of light for The variance of the theoretical distance measurement is: =144 Finally, the variance of the theoretical distance is adjusted by combining the number of single-photon avalanche diodes and the pre-calibrated coefficients. For example, if the number of single-photon avalanche diodes activated is... Pre-calibrated calibration coefficients 0.2 If the distance measurement noise is 500, then the estimated value is: = Finally, the number of single-photon avalanche diodes, the ambient signal rate, the number of reflected photons, and the distance measurement noise estimate are used as object category identification parameters. These distance parameters and object category identification parameters are then returned to the controller as measurement parameters for the measured object, thus obtaining the measurement parameters returned by the ranging sensor. It should be understood that the above is merely an example, and the specific process of the ranging sensor acquiring measurement parameters does not constitute a limitation of the present invention. The present invention only needs to obtain the aforementioned measurement data.

[0035] The term "high altitude" can be understood as a state at a certain distance from the ground or a plane. For example, a distance of 3 meters or more from the ground or a plane is considered a high altitude.

[0036] It should be understood that the high-altitude state can be determined in the following ways: by measuring the distance relative to the ground or plane using a distance measuring sensor to determine whether the person is in a high-altitude state; or by obtaining the three-dimensional positioning coordinates of the operator in the target scene; and by determining whether the operator is in a high-altitude state based on the three-dimensional positioning coordinates. This is only an example and does not constitute a limitation of the present invention.

[0037] S202: When the distance parameter is less than the preset distance threshold, the distance parameter and the object category recognition parameter are input into the pre-trained object category recognition model so that the object category recognition model outputs the material recognition result of the tested object.

[0038] In this embodiment, the preset distance threshold refers to a maximum distance value set when determining whether the hook has been reattached to the human body, used to limit the maximum effective judgment range between the ranging sensor and the object being measured. For example, the preset distance threshold can be set to 30 cm, 35 cm, or 40 cm, etc., and is not limited here.

[0039] In this embodiment, the material of the object being tested includes, but is not limited to, skin, clothing, and metal.

[0040] As an example, the distance parameter and object category recognition parameter are only input into the pre-trained object category recognition model when the distance parameter is below a preset distance threshold. This means that the object category recognition model only performs material recognition when the hook and the object being tested are at close range. This effectively avoids invalid recognition of irrelevant objects at a distance, simplifies the scenario, reduces the amount of data processing, speeds up the response, and improves the accuracy of the recognition results. On the other hand, by inputting the distance parameter and object category recognition parameter into the pre-trained object category recognition model, which is a pre-trained model already deployed on the controller, the model can quickly identify the material of the object being tested based on the mapping relationship between the input distance parameter and object category recognition parameter and the material of the object being tested after receiving the distance parameter and object category recognition parameter. Thus, even if the input data only includes distance parameters and object category recognition parameters, it can still learn general rules to accurately identify the material of the object being tested, such as skin, metal, or clothing. This not only improves recognition efficiency and accuracy but also achieves intelligent seat belt hook determination with low data volume, high precision, and high robustness. It does not require a high-powered controller or processor, nor does it require massive data storage. It effectively overcomes the shortcomings of traditional visual recognition, which relies on massive image data, consumes large computing resources, and has poor environmental adaptability. It significantly improves the real-time performance of detection and meets the application requirements of real-time alarms for intelligent seat belts.

[0041] In one embodiment, the object category recognition model is a model determined by training a pre-collected set of training measurement parameters, which includes a sequence of training distance parameters and a sequence of training object category recognition parameters corresponding to different tested objects.

[0042] In this embodiment, the training measurement parameter set includes training distance parameter sequences and training object category recognition parameter sequences corresponding to different measured objects pre-collected by the ranging sensor. For example, the training distance parameter sequences and training object category recognition parameter sequences corresponding to various common materials (jeans, suit pants, metal sheets, palms, arms, etc.) collected by the same ranging sensor under different light intensities and angles pre-collected.

[0043] As an example, by labeling the material category of the object being tested for each training measurement parameter in the training measurement parameter set, a training measurement parameter set and a validation measurement parameter set are formed. That is, by labeling the material category corresponding to each set of training distance parameters and training object category recognition parameters, a training measurement parameter set and a validation measurement parameter set are formed. Here, the training measurement parameter set is considered a feature, and the material category of the object being tested is considered a label. Next, the formed training measurement parameter set and validation measurement parameter set, along with the corresponding material category (label), are input into the object category recognition model in batches (epochs), and the parameters of the object category recognition model are updated using the backpropagation method. During training, when the object category recognition model can output an accurate material recognition result after inputting new training distance parameters and training object category recognition parameters—that is, outputting the accurate material category corresponding to the new training distance parameters and training object category recognition parameters—the model training is considered complete, forming a pre-trained object category recognition model.

[0044] like Figure 3 As shown in the experiment, the object category recognition model, after approximately 400 epochs of training, achieves an accuracy exceeding 95% for the training measurement parameter set and an accuracy exceeding 93% for the validation measurement parameter set. The loss for the training measurement parameter set is as low as 0.12, and the loss for the validation measurement parameter set is as low as 0.15. This fully demonstrates that the pre-trained object category recognition model possesses high recognition accuracy and good generalization ability. In other words, it proves that the present invention can achieve a high recognition accuracy rate using the aforementioned measurement parameters.

[0045] S203: Based on the material identification results and the preset material category, determine whether the hook of the smart safety belt should be hooked back onto the human body.

[0046] In this embodiment, the preset material category can be set according to the actual scenario. For example, the material of clothing commonly used by workers can be used as the identification benchmark, and there is no specific limitation.

[0047] As an example, when the material identification result is clothing, it indicates that the hook is back on the body (since people usually wear clothing, and safety suspension points are mostly made of metal or other materials). That is, it is determined that the hook is not attached to the safety suspension point. Therefore, the processor in the controller controls the alarm module in the intelligent safety belt detection device to issue an alarm through sound, light, and / or vibration, prompting the worker to adjust the hook position. When the material identification result is metal, it indicates that the hook is not back on the body (since safety suspension points are usually made of metal components). That is, it is determined that the hook is attached to the safety suspension point, meaning the hook is correctly worn, and no alarm needs to be triggered. In this way, real-time detection and risk warning of the hook status are achieved, effectively determining whether the hook is correctly back on the body, effectively avoiding safety hazards caused by the hook being mistakenly attached to an unsafe suspension point, and improving detection accuracy. It should be understood that the above is only an example and does not constitute a limitation of the present invention.

[0048] In one embodiment, the object category recognition parameters include the following parameters: The rate of environmental signals acquired by the ranging sensor; The number of reflected photons collected by the ranging sensor; The number of single-photon avalanche diodes enabled in the ranging sensor; The distance measurement noise estimate collected by the distance sensor.

[0049] In this embodiment, the object category recognition parameters may include any two, three, or all of the above parameters, and are not limited here.

[0050] Preferably, the object category recognition parameters can include all of the above parameters. That is, the object category recognition parameters include all of the above parameters, which helps to improve the accuracy of material recognition. Moreover, since the data of each parameter in the object category recognition parameters is small, even if all the parameters included in the object category recognition parameters are input, the total data size is still small, which has high engineering practicality.

[0051] As an example, such as Figure 4As shown, when the distance parameter is less than a preset distance threshold, the distance parameter and object category recognition parameter are input into a pre-trained object category recognition model. Specifically, the distance parameter, ambient rate, number of reflected photons, number of SPADs enabled, and distance measurement noise estimate (sigma) can be input into the pre-trained object category recognition model so that the model outputs the material recognition result of the tested object (e.g., skin, metal, or clothing). By inputting these five parameters into the pre-trained model, the material of the tested object can be reflected from multiple dimensions, further improving the accuracy and robustness of the material recognition, ensuring the reliable recognition effect of the intelligent safety belt detection device. Even with all parameters input, the amount of data is very small, greatly reducing the requirements for processor speed and power, as well as data storage, further improving detection real-time performance and efficiency. Real-time alarms can be issued for violations, further ensuring operational safety.

[0052] In one embodiment, such as Figure 4 and Figure 5 As shown, the object category recognition model includes an input layer, a dual-branch attention module, a feature fusion module, and a classification and output module. The dual-branch attention module includes an original feature branch module and an extended feature branch module. The input layer is connected to the original feature branch module, and the original feature branch module is connected to the feature fusion module. The input layer is connected to the extended feature branch module, and the extended feature branch module is connected to the feature fusion module. The feature fusion module is connected to the classification and output module.

[0053] As an example, the input layer is connected to the original feature branch module, which in turn is connected to the feature fusion module; the input layer is connected to the extended feature branch module, which in turn is connected to the feature fusion module; and the feature fusion module is connected to the classification and output module. Through this structural configuration, the input layer can simultaneously drive the original and extended feature branch modules to work in parallel, each extracting feature information from different levels. Subsequently, the feature fusion module performs weighted fusion of the features output from the two branches and inputs the fused features into the classification and output module. This achieves an organic unity between shallow response speed and deep representation capability, ensuring not only the real-time performance of material recognition but also significantly improving the model's adaptability and robustness in complex scenes through a multi-level feature complementarity mechanism, effectively enhancing the accuracy of object material recognition.

[0054] In one embodiment, such as Figure 4 As shown, in step S202, the distance parameters and object category recognition parameters are input into the pre-trained object category recognition model so that the object category recognition model outputs the material of the tested object. This includes the following steps: S220: Input the distance parameter and object category recognition parameter into the input layer so that the input layer outputs the original feature vector; S221: Input the original feature vector into the original feature branch module so that the original feature branch module outputs the weighted original feature vector; S222: Input the original feature vector into the extended feature branch module so that the extended feature branch module outputs a weighted extended feature vector; S223: Input the weighted original feature vector and the weighted extended feature vector into the feature fusion module so that the feature fusion module outputs the fused feature vector; S224: Input the fused feature vector into the classification and output module so that the classification and output module outputs the material recognition result of the tested object.

[0055] As an example, distance parameters and object category recognition parameters are input to the input layer, which maps the input distance parameters and object category recognition parameters into a raw feature vector, for example, a 5-dimensional raw feature vector [d, a, p, s, σ], where d is the distance parameter, a is the ambient signal rate, p is the number of reflected photons, s is the number of single-photon avalanche diodes, and σ is the distance measurement noise estimate; subsequently, the raw feature vector is input to the raw feature branch module (...). Figure 5 Right-side branch) and extended feature branch module ( Figure 5 (Left branch); where the original feature vector, after being processed by the original feature branch module, outputs a weighted original feature vector; simultaneously, the original feature vector, after being processed by the extended feature branch module, outputs a weighted extended feature vector; then, the weighted original feature vector and the weighted extended feature vector are synchronously input into the feature fusion module (e.g., Figure 5The model employs a fusion process (Concatenate) to concatenate and linearly transform the weighted original feature vector and the weighted extended feature vector, forming a fused feature vector. The dimension of the fused feature vector is the sum of the dimensions of the weighted original and extended feature vectors. This preserves the overall consistency of low-dimensional features while also considering the detailed representation of high-dimensional features. Finally, the fused feature vector is input to the classification and output module. After classification processing, the module outputs the material recognition result of the tested object, such as the material recognition result for skin, metal, or clothing. Thus, based on the collaborative processing between these modules, accurate material differentiation of the tested object can be achieved with only a small number of input parameters, ensuring recognition accuracy while also maintaining a lightweight model, low computational cost, high real-time performance, and strong environmental robustness.

[0056] In one embodiment, such as Figure 5 As shown, in step S221, which involves inputting the original feature vector into the original feature branch module so that the original feature branch module outputs a weighted original feature vector, the following steps are included: S2211: Perform a fully connected transformation on the original feature vector to obtain the first feature vector of the first target dimension, which is equal to the dimension of the original feature vector; S2212: Perform the Softmax function operation on the first feature vector to obtain the attention weight vector of the first target dimension; S2213: Based on the attention weight vector and the original feature vector, obtain the weighted original feature vector of the first target dimension.

[0057] In this embodiment, the fully connected transformation refers to the dimensionality-invariant feature space transformation of the original feature vector through a linear mapping.

[0058] As an example, firstly, a linear transformation can be performed on the original feature vector through the first fully connected layer (Dense layer) in the original feature branch module, outputting a first feature vector of the first target dimension. Here, the first target dimension is equal to the dimension of the original feature vector. For example, a 5-dimensional original feature vector can be converted into a 5-dimensional first feature vector, achieving pre-adjustment of feature importance. Next, the first activation layer of the original feature branch module applies the Softmax function to the first feature vector, normalizing its dimensional values ​​to obtain a 5-dimensional attention weight vector (the sum of its elements is 1). Finally, the attention weight vector and the original feature vector are combined in the first computational layer (Multiply layer) of the original feature branch module to obtain a weighted original feature vector. For example, the attention weight vector and the original feature vector are multiplied element-wise (e.g., corresponding element-wise multiplication) to obtain the weighted original feature vector. Alternatively, the attention weight vector and the original feature vector are input into two independent fully connected layers for nonlinear transformation to generate intermediate vectors, which are then multiplied element-wise to obtain the weighted original feature vector. This highlights key features that are more representative in the material recognition task (e.g., the priority of distance parameters), without limitation here. Thus, by applying attention weighting to the original feature branch module, the object category recognition model's ability to distinguish the importance of different feature dimensions is enhanced, contributing to improved overall recognition performance and robustness.

[0059] In one embodiment, such as Figure 4 and Figure 5 As shown, in step S222, which involves inputting the original feature vector into the extended feature branch module so that the extended feature branch module outputs a weighted extended feature vector, the following steps are included: S2221: Perform a fully connected transformation on the original feature vector to obtain a second feature vector with a second target dimension, which is higher than the dimension of the original feature vector; S2222: Perform a fully connected transformation on the second feature vector to obtain the third feature vector of the second target dimension; S2223: Perform the Softmax function operation on the third feature vector to obtain the extended attention weight vector of the second target dimension; S2224: Based on the extended attention weight vector and the second feature vector, obtain the weighted extended feature vector of the second target dimension.

[0060] In this embodiment, the fully connected transformation refers to the feature space transformation that increases the dimension of the original feature vector through linear mapping. For example, the original 5-dimensional feature vector is expanded into a second 20-dimensional feature vector.

[0061] As an example, firstly, the original feature vector can be fully transformed by the second fully connected layer (Dense layer) of the extended feature branch module to obtain a second feature vector of the second target dimension. That is, it is mapped to a second feature vector with a higher dimension than the original feature vector. For example, the original feature vector [d, a, p, s, σ] of dimension 5 can be mapped to a second feature vector [f1, f2, ..., f] of dimension 20 through the second fully connected layer. 20 To enhance feature representation, the second feature vector is then input into the third fully connected layer (Dense layer) of the extended feature branch module for further linear transformation or weight reconstruction, resulting in a third feature vector whose dimension remains the same as the second target dimension. For example, the second feature vector [f1, f2, ..., f...] is transformed into a third feature vector. 20 After performing linear or weighted transformations through the third fully connected layer, a third feature vector [g1, g2, ..., g] with a dimension of 20 is obtained. 20 Next, the third feature vector is input to the second activation layer of the extended feature branch module. The values ​​of each dimension are normalized using the Softmax function to obtain the extended attention weight vector. For example, for the third feature vector [g1, g2, ..., g...], ... 20 The Softmax function is executed to normalize the weights into an extended attention weight vector [w1, w2, ..., w]. 20 Finally, the extended attention weight vector and the third feature vector are combined through the second computation layer (Multiply layer) of the extended feature branch module to obtain the weighted extended feature vector. For example, the extended attention weight vector and the third feature vector are multiplied element-wise to output the weighted extended feature vector, or another method as described above is used, which is not limited here. For example, the extended attention weight vector [w1, w2, ..., w 20 Each weight w in ] ᵢ With the third eigenvector [g1, g2, ..., g 20 The corresponding component g in ] ᵢ Multiplying them one by one, we get the weighted extended feature vector [w1] g1,w2 g2, …, w 20 g 20 In this way, the original feature information can be further expanded in high-dimensional space and saliency weights can be integrated, effectively improving the model's discriminative ability and sensitivity to fine-grained differences, and enhancing the adaptability and robustness of the overall material recognition results in diverse scenarios. It should be understood that the above is merely an example and does not constitute a limitation of the present invention.

[0062] In one embodiment, such as Figure 5 As shown, in step S224, the object category recognition model also includes a residual network module; wherein, the feature fusion module is connected to the residual network module, and the residual network module is connected to the classification and output module; The fused feature vector is input into the classification and output module, so that the classification and output module outputs the material recognition result of the tested object, including: S2241: Input the fused feature vector into the residual network module so that the residual network module outputs an enhanced feature vector; S2242: Input the enhanced feature vector into the classification and output module so that the classification and output module outputs the material recognition result of the tested object.

[0063] In this embodiment, the dimension of the fused feature vector is the sum of the dimensions of the weighted original feature vector and the weighted extended feature vector. For example, if the original feature vector is 5-dimensional and the weighted extended feature vector is 20-dimensional, then the fused feature vector is 25-dimensional.

[0064] As an example, by inputting the fused feature vector into the residual network module, the residual network module outputs an enhanced feature vector. This enhanced feature vector is then input into the classification and output module, which outputs the material identification result of the tested object. Thus, by introducing the residual network module to deeply enhance the fused features, not only is the model's ability to capture complex relationships between features strengthened, but the accuracy and robustness of the final material classification result are also improved.

[0065] In one embodiment, such as Figure 5 As shown, in step S2241, which involves inputting the fused feature vector into the residual network module so that the residual network module outputs an enhanced feature vector, the following steps are included: S2241A: The fused feature vector is sequentially input into a multi-layer fully connected layer so that the multi-layer fully connected layer outputs an enhanced fused feature vector of the third target dimension, which is equal to the dimension of the fused feature vector; S2241B: Based on the enhanced fusion feature vector and the fusion feature vector, obtain the enhanced feature vector of the third target dimension.

[0066] As an example, the fused feature vector can be sequentially input into three fully connected layers of the residual network module. For instance, the fused feature vector can first be input into the fourth fully connected layer (dense-1 layer) of the residual network module, so that the fourth fully connected layer outputs a first residual vector of the fifth target dimension, where the fifth target dimension is higher than the dimension of the fused feature vector. For example, a fused feature vector of dimension 25 can be mapped to a first residual vector of dimension 32 through the fourth fully connected layer. Next, the first residual vector is input into the fifth fully connected layer (dense-2 layer) of the residual network module, so that the fifth fully connected layer outputs a second residual vector of the sixth target dimension, where the sixth target dimension is lower than the dimension of the fused feature vector. For example, a fused feature vector of dimension 25 can be mapped to a first residual vector of dimension 32 through the fourth fully connected layer. The first residual vector with a dimension of 32 is mapped to a second residual vector with a dimension of 16 through the fifth fully connected layer. Then, the second residual vector is input to the sixth fully connected layer (dense-3 layer) so that the sixth fully connected layer outputs an enhanced fusion feature vector of the third target dimension. For example, the second residual vector with a dimension of 16 is mapped to an enhanced fusion feature vector of 25 dimensions through the sixth fully connected layer. Subsequently, the residual connected layer (Add layer) of the residual network module obtains the enhanced feature vector of the third target dimension based on the enhanced fusion feature vector and the fusion feature vector. For example, the enhanced fusion feature vector is added to or multiplied element-wise with the original fusion feature vector (e.g., corresponding element-wise addition) to obtain the fusion enhanced intermediate vector. Thus, the dimension-up (25 to 32), dimension-down (32 to 16), and then dimension-up (16 to 25) processing paths in the residual network module can fully activate the nonlinear expressive power of the fusion features in different feature spaces while preserving the original fusion information, achieving higher-dimensional semantic modeling and information fusion, and providing a stronger discriminative and more robust input foundation for subsequent classification.

[0067] In one embodiment, such as Figure 5 As shown, in step S224, which involves inputting the enhanced feature vector into the classification and output module so that the classification and output module outputs the material of the tested object, the following steps are included: S2243: Perform a fully connected transformation on the enhanced feature vector to obtain the fourth feature vector of the fourth target dimension. The fourth target dimension is the same as the number of types of materials of the object being measured. S2244: Perform the Softmax function operation on the fourth feature vector to obtain the material probability distribution vector. Each dimension of the material probability distribution vector corresponds to the predicted probability of a material in the tested object. S2245: Determine the index corresponding to the maximum value in the material probability distribution vector, and map the index to the corresponding material label; S2246: Determine the material identification result of the object being tested based on the material label.

[0068] As an example, the enhanced feature vector can be fully transformed using the seventh fully connected layer (Dense layer) of the classification and output module to obtain a fourth feature vector of the fourth target dimension. This fourth target dimension is the same as the number of material types of the object being tested. For example, if there are three material types (e.g., skin, clothing, and metal), the corresponding fourth target dimension is 3-dimensional. For instance, an enhanced feature vector with dimension 25 can be mapped to a 3-dimensional fourth feature vector through the seventh fully connected layer. Next, the Softmax function is applied to the fourth feature vector to normalize it into a probability distribution, resulting in a material probability distribution vector. For example, the obtained material probability distribution vector might be [0.90, 0.95, ...]. [0.75], where each dimension of the material probability distribution vector corresponds to the predicted probability of a material in the tested object. For example, the first dimension corresponds to a predicted probability of 90% for skin, the second dimension corresponds to a predicted probability of 95% for clothing, and the third dimension corresponds to a predicted probability of 33% for 75 metal. Then, by comparing these three probability values, the maximum value is found to be in the second dimension, so this index is mapped to the material label of clothing. Finally, based on this material label, the material of the tested object is determined to be clothing. In this way, the deeply refined features can be directly converted into confidence outputs for each category, and the final material can be determined in one step through the maximum value index, improving the overall recognition accuracy and execution efficiency.

[0069] In one embodiment, the ranging sensor includes a laser ranging sensor. The laser used by the laser ranging sensor to acquire measurement parameters is non-visible light, which is laser light whose wavelength is not within the visible spectrum range of the human eye.

[0070] In this embodiment, non-visible light refers to electromagnetic waves that cannot be directly perceived by the human eye, such as a laser with a wavelength of 940nm. This invention does not limit this; as long as the laser is within the range that is invisible to the human eye, the application requirements of this invention can be met.

[0071] As an example, since the laser rangefinder uses non-visible light as a light source (i.e., the laser used by the laser rangefinder is non-visible light), the present invention can still operate stably even in environments with complex or sudden changes in lighting conditions, such as high-altitude operations (e.g., suddenly entering a dark area, nighttime operations, tunnels, etc.), and is not affected by the lack or change of visible light in the environment, thus having the ability to work continuously around the clock.

[0072] Experiments have shown that in environments with no visible light (e.g., a completely dark room with no active lighting), the laser rangefinder can still accurately collect measurement data of the object being measured. Based on the collected measurement data, the accuracy rate of identifying the material type of the object can reach over 92%, which is basically consistent with the recognition accuracy under normal lighting conditions. This effectively verifies the robustness and reliability of the non-visible light-based laser rangefinder in complex lighting environments.

[0073] Example 2 Secondly, such as Figure 1 As shown, an intelligent seat belt detection device 100 is provided. The intelligent seat belt detection device includes a distance sensor 101 and a controller 102. The controller 102 is connected to the distance sensor 101 and is used to implement the steps of the artificial intelligence-based intelligent seat belt hook detection method of the first aspect described above. To avoid repetition, it will not be described again here.

[0074] Thirdly, such as Figure 6 As shown, a smart seat belt 300 is provided. The smart seat belt 300 includes a smart seat belt detection device 100 as described in the second aspect and a seat belt hook 301. The smart seat belt detection device 100 is disposed on the seat belt hook 301.

[0075] Fourthly, a computer-readable storage medium is provided, which stores a computer program. When the computer program is executed by a processor, it implements the steps of the artificial intelligence-based intelligent seat belt hook detection method described in the first aspect above. To avoid repetition, these steps will not be repeated here.

[0076] In this invention, "multiple" refers to two or more.

[0077] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0078] In this invention, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, in this invention, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0079] Unless otherwise specified, all steps of this invention may be performed sequentially or randomly. For example, if the method includes steps A and B, it means that the method may include steps A and B performed sequentially, or it may include steps B and A performed sequentially. For example, if the method may also include step C, it means that step C may be added to the method in any order. For example, the method may include steps A, B, and C, or it may include steps A, C, and B, or it may include steps C, A, and B, etc.

[0080] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting intelligent seatbelt hooks based on artificial intelligence, characterized in that, include: The measurement parameters returned by the ranging sensor are obtained. The measurement parameters include distance parameters and object category identification parameters. The distance parameters are the distance values ​​of the ranging sensor relative to the measured object. The object category identification parameters include parameters for identifying the material of the measured object. When the distance parameter is less than a preset distance threshold, the distance parameter and the object category recognition parameter are input into a pre-trained object category recognition model so that the object category recognition model outputs the material recognition result of the tested object. Based on the material identification results and the preset material category, determine whether the hook of the smart safety belt should be hooked back onto the human body.

2. The intelligent seatbelt hook detection method based on artificial intelligence according to claim 1, characterized in that, The object category recognition parameters include the following parameters: The rate of environmental signal acquired by the ranging sensor; The number of reflected photons collected by the ranging sensor; The number of single-photon avalanche diodes enabled in the ranging sensor; The distance measurement noise estimate acquired by the ranging sensor.

3. The intelligent seatbelt hook detection method based on artificial intelligence according to claim 1 or 2, characterized in that, The object category recognition model includes an input layer, a dual-branch attention module, a feature fusion module, and a classification and output module. The dual-branch attention module includes an original feature branch module and an extended feature branch module. The step of inputting the distance parameter and the object category recognition parameter into a pre-trained object category recognition model, so that the object category recognition model outputs the material recognition result of the tested object, includes: The distance parameter and the object category recognition parameter are input into the input layer so that the input layer outputs the original feature vector. The original feature vector is input into the original feature branch module, so that the original feature branch module outputs a weighted original feature vector; The original feature vector is input into the extended feature branch module, so that the extended feature branch module outputs a weighted extended feature vector; The weighted original feature vector and the weighted extended feature vector are input into the feature fusion module so that the feature fusion module outputs a fused feature vector; The fused feature vector is input into the classification and output module so that the classification and output module outputs the material recognition result of the tested object.

4. The intelligent seatbelt hook detection method based on artificial intelligence according to claim 3, characterized in that: The step of inputting the original feature vector into the original feature branch module, so that the original feature branch module outputs a weighted original feature vector, includes: A fully connected transformation is performed on the original feature vector to obtain a first feature vector of a first target dimension, wherein the first target dimension is equal to the dimension of the original feature vector; The first feature vector is subjected to the Softmax function to obtain the attention weight vector of the first target dimension; Based on the attention weight vector and the original feature vector, the weighted original feature vector of the first target dimension is obtained; and / or; The original feature vector is input to the extended feature branch module so that the extended feature branch module outputs a weighted extended feature vector, including: A fully connected transformation is performed on the original feature vector to obtain a second feature vector with a second target dimension, wherein the second target dimension is higher than the dimension of the original feature vector; A fully connected transformation is performed on the second feature vector to obtain the third feature vector of the second target dimension; The third feature vector is subjected to the Softmax function operation to obtain the extended attention weight vector of the second target dimension; Based on the extended attention weight vector and the second feature vector, the weighted extended feature vector of the second target dimension is obtained.

5. The intelligent seatbelt hook detection method based on artificial intelligence according to claim 4, characterized in that, The object category recognition model also includes a residual network module; The step of inputting the fused feature vector into the classification and output module, so that the classification and output module outputs the material recognition result of the tested object, includes: The fused feature vector is input into the residual network module so that the residual network module outputs an enhanced feature vector; The enhanced feature vector is input into the classification and output module so that the classification and output module outputs the material recognition result of the tested object.

6. The intelligent seatbelt hook detection method based on artificial intelligence according to claim 5, characterized in that, The step of inputting the fused feature vector into the residual network module so that the residual network module outputs an enhanced feature vector includes: The fused feature vector is sequentially input into a multi-layer fully connected layer so that the multi-layer fully connected layer outputs an enhanced fused feature vector of the third target dimension, which is equal to the dimension of the fused feature vector; The enhanced feature vector of the third target dimension is obtained based on the enhanced fusion feature vector and the fusion feature vector.

7. The intelligent seatbelt hook detection method based on artificial intelligence according to claim 5, characterized in that, The step of inputting the enhanced feature vector into the classification and output module, so that the classification and output module outputs the material recognition result of the tested object, includes: A fully connected transformation is performed on the enhanced feature vector to obtain a fourth feature vector of the fourth target dimension, wherein the fourth target dimension is the same as the number of types of materials of the object being tested. The fourth feature vector is subjected to the Softmax function to obtain the material probability distribution vector, and each dimension of the material probability distribution vector corresponds to the predicted probability of a material in the tested object. Determine the index corresponding to the maximum value in the material probability distribution vector, and map the index to the corresponding material label; Based on the material label, the material identification result of the tested object is determined.

8. The intelligent seatbelt hook detection method based on artificial intelligence according to claim 1, characterized in that, The ranging sensor includes a laser ranging sensor. The laser used by the laser ranging sensor to acquire the measurement parameters is non-visible light, which is laser light whose wavelength is outside the visible spectrum of the human eye.

9. An intelligent seatbelt detection device, characterized in that, The intelligent seat belt detection device includes a distance sensor and a controller, the controller being connected to the distance sensor and used to implement the steps of the artificial intelligence-based intelligent seat belt hook detection method as described in any one of claims 1 to 8.

10. A smart seat belt, characterized in that, The smart seat belt includes the smart seat belt detection device as described in claim 9 and the seat belt hook, wherein the smart seat belt detection device is disposed on the seat belt hook.

11. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the artificial intelligence-based intelligent seat belt hook detection method as described in any one of claims 1 to 8.

Images