Dynamic target rapid perception method and system based on deep learning

Abstract

The present invention discloses a method and system for fast perception of dynamic targets based on deep learning. The method comprises: based on the joint calibration of the perception system, performing environmental perception and collecting data from different data sources; based on the collected data, performing scene acquisition, obtaining map data and posture information; based on the acquired map data and posture information, performing normalized data preprocessing of multivariate features, and performing feature matching on the data after normalized data preprocessing through superpixel segmentation and semantic analysis; based on the data obtained by feature matching, using sparse convolutional network and motion estimation fusion to perceive the dynamic target. The present invention solves the technical problem of low accuracy in dynamic target recognition in related technologies.

Description

Technical Field

[0001] The present invention relates to the field of AI intelligence, and in particular to a method and system for rapid perception of dynamic targets based on deep learning. Background Art

[0002] With the development of power technology and the improvement of power quality requirements of power users, there is an urgent need to build dynamic target rapid perception technology for substations. The current technical status is as follows:

[0003] (1) The current reduction in construction land has led to a decrease in the distance between equipment. Due to the lack of accurate sensing and positioning technology and methods, the construction of live-line operations near substations has become more difficult and the safe working space has been reduced;

[0004] (2) The working space of urban line channels has become crowded, lacking non-contact active dynamic identification and judgment of safety distance;

[0005] (3) The sensitivity and distance accuracy of the distance sensing devices based on field strength theory currently available on the market are greatly affected by voltage, electromagnetic field, and transmission environment medium. At present, there is no mature product for the high-precision dynamic target rapid perception monitoring platform of power plants. It is urgent to study new technologies and methods to ensure the safety of personnel and equipment at various live-line construction sites such as power plants, substations, and distribution stations.

[0006] In summary, although various detection algorithms have been developed for dynamic target perception and detection, most deep learning-based target perception methods rely on a large number of annotation labels, require a lot of work, are not applicable to uncooperative spatial targets, and the versatility of these algorithms needs to be further improved. The accuracy and robustness of the algorithms are not strong enough to quickly and accurately lock targets and calculate the state and category parameters of dynamic targets in complex environments. In addition, the extraction of different feature values ​​often only focuses on one aspect of the image information and cannot express the target more comprehensively.

[0007] To address the above-mentioned problems, no effective solution has been proposed yet. Summary of the invention

[0008] The embodiments of the present invention provide a method and system for rapid perception of dynamic targets based on deep learning, so as to at least solve the technical problem of low accuracy in dynamic target recognition in related technologies.

[0009] According to one aspect of an embodiment of the present invention, a method for fast perception of dynamic targets based on deep learning is provided, comprising: based on the joint calibration of the perception system, performing environmental perception, collecting data from different data sources; based on the collected data, performing scene acquisition, obtaining map data and posture information; based on the acquired map data and posture information, performing normalized data preprocessing of multivariate features, and performing feature matching on the normalized data preprocessed by superpixel segmentation and semantic analysis; based on the data obtained by feature matching, using a sparse convolutional network and motion estimation fusion to perceive the dynamic target.

[0010] According to another aspect of an embodiment of the present invention, a deep learning-based rapid perception system for dynamic targets is also provided, including: a perception module, configured to perform environmental perception based on joint calibration of the perception system, and collect data from different data sources; a scene acquisition module, configured to perform scene acquisition based on the collected data, and obtain map data and posture information; a data processing module, configured to perform normalized data preprocessing of multi-features based on the acquired map data and posture information, and perform feature matching on the preprocessed normalized data through superpixel segmentation and semantic analysis; and a recognition module, configured to perceive the dynamic target based on the data obtained by feature matching, using a sparse convolutional network and motion estimation fusion.

[0011] According to another aspect of an embodiment of the present invention, a non-contact safety distance active warning method is also provided, comprising: executing the deep learning-based dynamic target rapid perception method as described above; and actively issuing a warning when the perceived dynamic target exceeds the safety distance threshold.

[0012] In an embodiment of the present invention, based on the joint calibration of the perception system, environmental perception is performed and data is collected from different data sources; based on the collected data, scene acquisition is performed to obtain map data and posture information; based on the acquired map data and posture information, normalized data preprocessing of multivariate features is performed, and feature matching is performed on the preprocessed normalized data through superpixel segmentation and semantic analysis; based on the data obtained by feature matching, a sparse convolutional network and motion estimation fusion are used to perceive the dynamic target, thereby solving the technical problem of low accuracy in dynamic target recognition in related technologies, and thus having the technical effect of quickly and accurately identifying dynamic targets. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0014] Figure 1is a flow chart of a non-contact safety distance active warning method according to a first embodiment of the present invention;

[0015] Figure 2 is a flow chart of a non-contact safety distance active warning method according to a second embodiment of the present invention;

[0016] Figure 3 is a flowchart of the workflow of the precise positioning subsystem according to an embodiment of the present invention;

[0017] Figure 4 is a flowchart of the working process of the intelligent error compensation subsystem according to an embodiment of the present invention;

[0018] Figure 5 is a flowchart of a workflow of an environment monitoring subsystem according to an embodiment of the present invention;

[0019] Figure 6 is a flowchart of another workflow of the environment monitoring subsystem according to an embodiment of the present invention;

[0020] Figure 7 is a structural schematic diagram of a non-contact safety distance active warning system according to an embodiment of the present invention;

[0021] Figure 8 is a flow chart of a method for rapid perception of dynamic targets according to a first embodiment of the present invention;

[0022] Fig. 9 is a flow chart of a method for rapid perception of dynamic targets according to a second embodiment of the present invention;

[0023] Fig.10 is a schematic diagram of feature extraction according to an embodiment of the present invention;

[0024] Fig.11 is a flow chart of a method for rapid perception of dynamic targets according to a third embodiment of the present invention;

[0025] Fig.12 is a flow chart of a method for rapid perception of dynamic targets according to a fourth embodiment of the present invention;

[0026] Fig.13 is a schematic diagram of the structure of a dynamic target rapid perception system according to an embodiment of the present invention;

[0027] Fig.14 4 is a schematic diagram of the structure of a non-contact safety distance active warning system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0028] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0029] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0030] Overview

[0031] For dynamic target perception, dynamic target detection is the first step to achieve recognition. Dynamic target detection is to accurately extract the changing area from the background image in the sequence image. The main purpose of dynamic recognition perception is to determine the appropriate control amount based on the angle deviation and distance deviation obtained by path recognition to minimize the two deviations.

[0032] This application constructs a unified data benchmark framework in multiple time and space through a normalized expression method of multi-source on-site information scale, time, relationship, etc. based on inertial, 5G, GNSS and three-dimensional map database, realizes feature segmentation of map data, and uses a sparse convolutional neural network architecture to extract typical features of real-life construction site maps and intelligent online recognition and perception. At the same time, a learning strategy is used to optimize the parameters, realizing rapid perception and recognition of dynamic targets under quantitative control of safe distances.

[0033] In addition, through non-contact active dynamic identification of safe distances, safety management and control of construction sites can be achieved, effectively ensuring the overall management of project construction, greatly improving the safety, timeliness and technology of power transmission and transformation project construction, and effectively ensuring the safety of power transmission and transformation project construction; achieving the goal of safe construction of all power transmission and transformation project machinery and personnel.

[0034] Example 1

[0035] According to an embodiment of the present invention, a non-contact safety distance active warning method is provided, such as Figure 1 As shown, the method includes:

[0036] Step S102, based on the inertial device, the positioning data of the engineering vehicle and the operator are obtained.

[0037] Based on inertial devices, such as inertial sensors, the posture information of the engineering vehicle and the operator at the construction site is obtained, and a posture motion constraint model is established by analyzing the motion characteristics of the engineering vehicle and the operator; the posture error of the operator and the engineering vehicle with motion constraints is estimated using the optimal estimation method; and the posture information of the operator and the engineering vehicle is obtained using a weak strapdown navigation failure error model.

[0038] Afterwards, multi-frequency channel estimation is used to obtain signal strengths and multipath delay clusters at different frequencies. A multipath space-time compensation model is used to achieve precise positioning in strong electric field interference to obtain auxiliary positioning information. The auxiliary positioning information is used to correct the posture information of the operating personnel and the engineering vehicle to obtain corrected positioning data.

[0039] Step S104, using Kalman filtering and inertial posture error estimation, error compensation is performed on the inertial device to obtain positioning data after error compensation.

[0040] Using Kalman filtering and inertial posture error estimation, error compensation for the inertial device may include at least one of the following: compensating for the gyro output error of the inertial device based on at least one of the following: gyro original output value vector, coarse calibration zero position output, gyro angular velocity zero position output value vector, temperature-related zero position output, acceleration-related zero position output, random noise signal, coarse calibration proportional coefficient of proportional coefficient matrix, proportional coefficient of input angular velocity, cross-coupling coefficient matrix; filtering and denoising the gyro noise under dynamic conditions to compensate for the random shaking error of the vehicle carrier.

[0041] Among them, filtering and denoising the dynamic gyro noise can include: performing fast Fourier transform on the gyro noise in a moving state; performing a first genetic operation on the gyro noise after the fast Fourier transform, randomly generating a binary sequence C, and extracting a noise signal from the gyro noise after the fast Fourier transform based on the binary sequence; performing an inverse fast Fourier transform, transforming the extracted noise signal into a continuous signal, and comparing the continuous signal with the gyro noise signal in a static state to obtain an inverted sequence; and multiplying the inverted sequence with the gyro noise after the fast Fourier transform to perform filtering and denoising.

[0042] Step S106, obtaining the work site environment data of the engineering vehicle and personnel;

[0043] For example, based on sensor self-calibration and system joint calibration, the work site environment is perceived and field data is collected; the collected data is preprocessed to filter out interference factors and extract features; the effective feature point extraction method of three-dimensional map data is used to perform multiple feature matching on the extracted features to reconstruct the three-dimensional scene of the engineering site in real time.

[0044] Then, based on the map data visualization technology and data automatic association analysis technology, the superpixels of the typical environmental feature data of the map data are segmented and the semantic relationship is analyzed; using the sparse convolutional neural network architecture, based on the semantic relationship and the segmented superpixels of the typical environmental feature data of the map data, the target recognition and active monitoring of the workers and construction vehicles in the three-dimensional scene of the construction site are carried out.

[0045] Step S108, determining whether the engineering vehicle or the personnel is within the non-contact safety distance of the substation based on the error-compensated positioning data and the work site environment data, and actively issuing an early warning if they are outside the non-contact safety distance of the substation.

[0046] For example, based on the positioning data after error compensation and the work site environment data, the risk status of the engineering vehicle and the operator is classified using a multi-dimensional scene linkage visualization engine; based on the pre-configured warning parameters and the classified risk status, it is determined whether the engineering vehicle or the operator is within the non-contact safety distance of the substation.

[0047] This embodiment uses inertial positioning to detect live construction environments and vehicle interference, and compensates for the shaking errors of engineering vehicles, thereby solving the problem of spatial positioning in weak / no satellite environments. In addition, satellite enhanced positioning technology is used to assist inertial positioning, providing high-precision spatial positioning and achieving safe management and control of the construction site.

[0048] The non-contact active warning method in this embodiment can provide rapid, effective and safe warning before an accident occurs, effectively avoiding loss of personnel and equipment. At the same time, it can improve the technology and safety of existing power plants and promote the digitalization, informatization and technological upgrading of power plants.

[0049] Example 2

[0050] According to an embodiment of the present invention, a non-contact safety distance active warning method is provided. The method can be applied in the following scenarios: the project application environment includes five conditions: 500 kV AC site, 220 kV AC site, 35 kV AC site, 500 kV flexible DC commutation site, and large deep foundation pit operation scene. Figure 2 As shown, the method comprises the following steps:

[0051] Step S201, calling the precise positioning subsystem.

[0052] The precise positioning subsystem is based on inertial autonomous positioning and orientation as its core, supplemented by satellite navigation positioning enhancement technology. It establishes a multi-source fusion autonomous precise positioning system based on factors such as signal strength, environmental conditions, and movement form.

[0053] Step S202, calling the error intelligent compensation subsystem.

[0054] The intelligent error compensation subsystem divides the vehicle's working inertia data into periods and extracts features, builds a model using a deep neural network, and solves the compensation problem using Kalman filtering and inertial posture error estimation.

[0055] Step S203, calling the environment monitoring subsystem.

[0056] The environmental monitoring subsystem establishes a unified reference framework for data in multiple time and space based on inertial, 5G, GNSS and three-dimensional map database, extracts effective feature points, combines deep learning algorithms, and uses visualization technology to complete fusion display.

[0057] The environmental monitoring subsystem mainly performs the following three steps: risk superposition assessment, early warning assessment and multi-dimensional risk assessment, graded early warning and multi-dimensional statistical analysis.

[0058] Risk overlay assessment: Utilize the multi-dimensional scene linkage visualization engine to realize visual simulation display, achieve accurate positioning of equipment and operations on real-life maps, classification of risk points, and display of risk status, and realize hierarchical display of risk overlay assessment results in the region on the map.

[0059] Early warning assessment: Establish a multi-objective resource scheduling model for emergency rescue under a multi-level grid, realize the visualization and nodeization of early warning parameter configuration, visualize the process display and operation, and realize the customized and flexible configuration of early warning assessment indicators.

[0060] Multi-dimensional risk assessment, graded warning and multi-dimensional statistical analysis: Establish a visual intelligent identification and warning platform that supports edge computing microservice framework and elastic dynamic aggregation, support multi-dimensional risk assessment, graded warning and multi-dimensional statistical analysis, and fully support monitoring on-site command decision-making.

[0061] Step S204, determining whether the operator has made an erroneous operation.

[0062] Step S205, calling the control and decision subsystem.

[0063] Based on the above subsystems, the control decision subsystem provides safety protection and active warning when the on-site operator makes an error operation or exceeds the safety distance.

[0064] Step S206, sound an alarm and make a response.

[0065] Step S207: do not issue an alarm and continue monitoring.

[0066] This embodiment effectively solves the problems of quantitative control of construction safety distance, single monitoring means, large amount of environmental monitoring data, and gaps in construction environment monitoring. Through non-contact active dynamic identification of safety distance, it eliminates errors of construction personnel and vehicles, makes safety warnings before accidents occur, realizes safety management of construction sites, avoids serious production accidents, effectively ensures the safety management of engineering project construction, greatly improves the safety, timeliness and technology of power transmission and transformation project construction, effectively ensures the construction of power transmission and transformation projects, and the solution is universal and can be applied to most work scenarios after appropriate modification according to the needs of different work scenarios.

[0067] In addition, by unifying multi-source data under a unified time and space benchmark, the efficiency of data management, statistics, and organization is significantly improved, solving the problem of complex data processing in the above systems. At the same time, it fills the domestic technical gap of lack of wearable portable high-precision positioning equipment, promotes the standardization, productization, and industrialization of key technologies, information services, and instrumentation, and provides scientific and technological support for the continuous, stable, and safe operation and development of power plants.

[0068] Example 3

[0069] According to an embodiment of the present invention, a non-contact safety distance active warning method is provided. The steps in this embodiment are similar to those in Embodiments 1 and 2, except for the method performed by the precise positioning subsystem.

[0070] The precise positioning subsystem includes an inertial autonomous positioning module and an elastic fusion precise positioning technology module. The precise positioning subsystem has the following working procedures: Figure 3 As shown, the following steps S302 to S304 are included.

[0071] Step S302, obtaining position and posture information based on an inertial device.

[0072] The inertial autonomous positioning module of the precise positioning subsystem solves the problem of obtaining the posture information of construction vehicles and operators at the construction site. By analyzing the motion characteristics of personnel and construction vehicles, a vehicle posture motion constraint model is established. The optimal estimation method is used to obtain the posture error estimation method of motion-constrained personnel and vehicles, and the posture error that best meets the actual situation is calculated. Combined with the weak strapdown navigation failure error model, the posture information of operators and construction vehicles can be accurately obtained.

[0073] Step S304: perform precise positioning based on 5G regional integration.

[0074] The elastic fusion precise positioning technology module of the precise positioning subsystem uses a 5G-based regional elastic fusion precise collaborative positioning method to perform channel estimation at multiple frequency points, obtain information such as the strength of signals at different frequency points, multipath time delay cluster distribution, and establish a multipath space-time compensation model to achieve precise positioning in a strong electric field interference environment.

[0075] Example 4

[0076] According to an embodiment of the present invention, a non-contact safety distance active warning method is provided. The steps in this embodiment are similar to the steps in Embodiments 2 and 3, except that the method performed by the error intelligent compensation subsystem. The workflow of the error intelligent compensation subsystem is as follows: Figure 4 shown.

[0077] First, the gyro output error characteristics are analyzed.

[0078] Error sources include deterministic error of inertial devices, random error, vehicle random shaking error, etc. The errors of inertial devices are divided into: zero position error, proportional coefficient error, cross-coupling error, and random error.

[0079] First, the gyro output error characteristics are analyzed, and its error model is:

[0080]

[0081] is the angular velocity output vector after error compensation; A q is the original output value vector of the three-axis gyroscope; N1 is the coarse calibration zero-position output, N represents the gyroscope angular velocity zero-position output value vector, N(T) is the temperature-related zero-position output, and N(a) is the acceleration-related zero-position output; v(t) is the random noise signal; the proportional coefficient matrix is ​​represented by SF, SF1 is the coarse calibration proportional coefficient, and SF2(Ω) is the proportional coefficient corresponding to the input angular velocity; the cross-coupling coefficient matrix is ​​represented by CR.

[0082] Secondly, the random shaking error of the vehicle carrier is corrected.

[0083] 1) Use deep neural network method to establish inertial sensor model.

[0084] 2) Based on the principle of adaptive moment estimation, the sensor parameter optimization method of engineering vehicles is used to improve the perception accuracy.

[0085] Finally, the dynamic gyro noise is filtered and denoised.

[0086] The gyro noise signal in a static state is selected as a reference. In order to improve the quality of the gyro signal under normal walking, the dynamic gyro noise is the main processing object of the filtering noise reduction algorithm. The steps are as follows:

[0087] 1) Perform fast Fourier transform on the noisy gyro data in motion.

[0088] 2) Perform the first genetic operation to randomly generate a binary sequence C.

[0089] 3) Perform inverse fast Fourier transform to transform the extracted noise signal into a continuous signal and compare it with the gyro noise signal in a static state.

[0090] 4) Multiply the inverted sequence with the gyro noise after the fast Fourier transform of the motion state to complete the entire filtering and noise reduction process.

[0091] In this embodiment, the vehicle working inertia data is divided into periods and features are extracted, a deep neural network is used to build a model, and Kalman filtering and inertial posture error estimation are used to solve compensation, so that engineering vehicles and operators can be positioned more accurately.

[0092] Example 5

[0093] According to an embodiment of the present invention, a non-contact safety distance active warning method is provided. The steps in this embodiment are similar to the steps in embodiments 2 to 4, except for the method performed by the environment monitoring subsystem.

[0094] This embodiment uses positioning data in a spatial rectangular coordinate system or a spatial geodetic coordinate system, and the point cloud data and positioning data under different references must be unified under the same time and space reference. The workflow of the environmental monitoring subsystem implemented in this embodiment is as follows: Figure 5 As shown, the process includes the following steps S502 to S506.

[0095] Step S502: environmental perception.

[0096] Based on calibration methods such as sensor self-calibration and system joint calibration, environmental perception of the work site is carried out.

[0097] Step S504: feature extraction and real-time reconstruction of the three-dimensional scene.

[0098] The collected data is preprocessed to filter out interference factors and extract features. Using map data image representation technology and data automatic association analysis technology, map data segmentation is achieved through superpixel segmentation and semantic relationship analysis of typical environmental feature data. Using the effective feature point extraction method of three-dimensional map data, multiple feature matching is performed to achieve real-time reconstruction of the three-dimensional scene of the construction site.

[0099] Step S506, active warning.

[0100] Combined with deep learning, the sparse convolutional neural network architecture is used to realize the recognition and active monitoring of typical targets on the real-life map of the construction site. The high-precision instantaneous attitude solution of IMU is integrated with 5G and GNSS to obtain the complete scene.

[0101] Example 6

[0102] According to an embodiment of the present invention, a non-contact safety distance active warning method is provided. The steps in this embodiment are similar to the steps in embodiments 2 to 4, except for the method performed by the environment monitoring subsystem.

[0103] This embodiment uses positioning data in a spatial rectangular coordinate system or a spatial geodetic coordinate system, and the point cloud data and positioning data under different references must be unified under the same time and space reference. The workflow of the environmental monitoring subsystem implemented in this embodiment is as follows: Figure 6 As shown, the process includes the following steps S602 to S606.

[0104] Step S602: environmental perception.

[0105] First, data collection is performed to obtain, for example, map data, inertial data, and 5G+GNSS data.

[0106] Then the system is calibrated. Based on calibration methods such as sensor self-calibration and system joint calibration, the work site environment perception is carried out.

[0107] Step S604, information processing.

[0108] Information processing mainly includes two parts: data processing and data fusion.

[0109] Data processing includes data preprocessing, multivariate segmentation, feature extraction, etc. Data fusion includes unified benchmark, feature level, etc. Through information processing, data preprocessing is performed on the collected data to complete interference factor filtering and feature extraction.

[0110] Data fusion is mainly about unifying the benchmark and determining the feature level.

[0111] Step S606: environment reconstruction and target positioning.

[0112] Environmental reconstruction and target positioning mainly include: environmental modeling, target recognition and active monitoring.

[0113] Environmental modeling is mainly based on environmental features, multiple feature matching, and 3D reconstruction. Target recognition is mainly template matching and feature extraction. Active monitoring is mainly target positioning and target monitoring.

[0114] In this embodiment, the map data is segmented by combining the map data visualization technology and the data automatic association analysis technology, using the superpixel segmentation and semantic relationship analysis methods of typical environmental feature data. In combination with deep learning, a sparse convolutional neural network architecture is used to realize the recognition and active monitoring of typical targets in the real-life construction site map.

[0115] Example 7

[0116] According to an embodiment of the present invention, a non-contact safety distance active warning device is provided. The non-contact safety distance active warning device in this embodiment is as follows: Figure 7 As shown, including:

[0117] The precise positioning subsystem 72 is configured to obtain the positioning data of the engineering vehicle and the operating personnel based on the inertial device; the intelligent error compensation subsystem 74 is configured to use Kalman filtering and inertial posture error estimation to perform error compensation on the inertial device and obtain the positioning data after error compensation; the environmental monitoring subsystem 76 is configured to obtain the working site environmental data of the engineering vehicle and the personnel; the control and decision subsystem 78 is configured to determine whether the engineering vehicle or the personnel is within the non-contact safety distance of the substation based on the positioning data after error compensation and the working site environmental data, and actively issue an early warning if they are outside the non-contact safety distance of the substation.

[0118] The non-contact safety distance active warning device in this embodiment can implement the methods in the above embodiments 1 to 6, which will not be repeated here.

[0119] This embodiment consists of four subsystems: a precise positioning subsystem, an intelligent error compensation subsystem, an environmental monitoring subsystem, and a control and decision-making subsystem, which are respectively responsible for personnel and equipment positioning, disturbance and noise elimination, power plant working environment monitoring, and safety measures response functions, and have the beneficial effect of accurate early warning.

[0120] The following embodiments will focus on the method for rapid perception of dynamic targets. On the basis of the method for rapid perception of dynamic targets, an active warning can be issued when the dynamic target exceeds the safety distance threshold. In other words, the method for rapid perception of dynamic targets in the following embodiments is part of the non-contact safety distance active warning method. The non-contact safety distance active warning method includes: a method for rapid perception of dynamic targets and an early warning method, wherein the early warning method is to actively issue an early warning when the dynamic target identified by the method for rapid perception of dynamic targets exceeds the safety distance threshold.

[0121] Example 8

[0122] According to an embodiment of the present invention, a method for fast perception of dynamic targets based on deep learning is provided. Figure 8 As shown, the method includes:

[0123] Step S802, based on the joint calibration of the perception system, perform environmental perception and collect data from different data sources;

[0124] Step S804, based on the collected data, performing scene acquisition to obtain map data and position information;

[0125] Step S806, based on the acquired map data and position information, normalized data preprocessing of multi-features is performed, and feature matching is performed on the normalized data preprocessed by superpixel segmentation and semantic analysis.

[0126] In an exemplary embodiment, performing normalized data preprocessing of multivariate features based on the acquired map data and posture information includes at least one of the following: performing information scale normalized data preprocessing on the acquired map data and posture information; performing time series normalized data preprocessing on the acquired map data and posture information; performing global normalized data preprocessing on the acquired map data and posture information.

[0127] In an exemplary embodiment, the acquired map data and posture information are subjected to normalized data preprocessing of information scale, including: performing standardization processing on the acquired map data and posture information, wherein the standardization processing includes at least one of the following: unifying the data encoding method of data from different data sources into the same method; unifying the data format of data from different data sources into the same format; performing matching processing on the standardized map data and posture information, wherein the matching processing includes at least one of the following: matching together different attributes of different data sources that identify the same thing; identifying the same data from different data sources as duplicates; and performing data screening on the matched map data and posture information to select consistent dimension data.

[0128] In an exemplary embodiment, the acquired map data and posture information are preprocessed for time series normalization, including: based on the pattern correlation between adjacent time series and the characteristics of continuous data generation in the time dimension, the acquired map data and posture information from different data sources are temporally aligned to facilitate subsequent algorithms.

[0129] In an exemplary embodiment, the acquired map data and posture information are subjected to global normalized data preprocessing, including: windowing and segmenting the time series to obtain multiple time subsequences of fixed window lengths; based on the relationship factor of adjacent time series, the time subsequence in the current window is subjected to normalized data preprocessing using the maximum and minimum values ​​of the time subsequence in the current window.

[0130] In an exemplary embodiment, before performing feature matching on the data preprocessed by normalized data through superpixel segmentation and semantic analysis, the method further includes: defining the error state vector of the perception system as a column vector; determining the relationship between the angular velocity and acceleration of the inertial device of the perception system and the measurement value based on the column vector; determining the update equation of the inertial device based on the determined relationship between the angular velocity and acceleration of the inertial device and the measurement value; and determining the attitude error of the inertial device using the determined update equation.

[0131] For example, using the determined update equation to determine the attitude error of the inertial device includes: obtaining a transfer matrix between frames of the map data based on map feature point extraction and matching relationships; using the transfer matrix to solve corresponding attitude angle information, using the angular increment change value of the inertial device as a weight adjustment factor and solving the attitude error based on the attitude angle information.

[0132] Step S808, based on the data obtained by feature matching, a sparse convolutional network and motion estimation fusion are used to perceive the dynamic target.

[0133] This embodiment unifies the benchmark of data in multiple time and space, and improves the multi-source information processing capability of data; it uses a sparse convolutional neural network architecture to extract typical features of the construction site real-life map and realize intelligent online recognition, which is suitable for multiple types of targets, has strong adaptability, and high recognition accuracy.

[0134] This embodiment can be applied to live construction sites to quickly identify workers and engineering vehicles. On the one hand, this embodiment can effectively reduce the probability of power outages, personal accidents and equipment accidents caused by subjective experience and traditional sensor signal judgment, and effectively solve the main safety hazards in live construction operations. On the other hand, it is also conducive to the supervision of management during the construction process, further enhances the image of the power grid industry, effectively improves power supply reliability and user satisfaction with electricity, and has obvious social benefits. Therefore, the method proposed in this embodiment has practical value.

[0135] In addition, the contactless safe distance active dynamic target recognition algorithm proposed in this embodiment is a normalized expression method of multi-source on-site information scale, time, relationship and other multi-factor features based on inertia, 5G, GNSS and three-dimensional map database, constructs a unified data benchmark framework in multiple time and space, realizes feature segmentation of map data, and combines deep learning algorithms to use sparse convolutional neural network architecture to realize typical feature extraction and intelligent online recognition and perception of real-life construction site maps. At the same time, learning strategies are used to optimize parameters. The unified spatial benchmark is to establish the target and environmental posture information that needs to be perceived in a unified coordinate system or obtain it through matrix transformation, thereby improving the recognition and monitoring of different spatial targets.

[0136] Furthermore, the perception system in this embodiment can quickly and accurately calculate the state and category parameters of dynamic targets in complex environments, significantly improving the accuracy and speed of dynamic target recognition.

[0137] In addition to being applied to live construction sites, the dynamic target rapid perception method provided by this embodiment can be widely used in various fields, providing a new technical route for realizing universal, proactive, high-precision, and cost-effective dynamic recognition rapid perception technology. In particular, the dynamic target rapid perception strategy based on deep learning proposed in this embodiment can be applied to target recognition monitoring based on deep learning in other fields, so the method proposed in this embodiment is more universal.

[0138] Example 9

[0139] According to an embodiment of the present invention, a method for rapid perception of dynamic targets based on deep learning is provided.

[0140] The premise for the realization of the fast perception method of dynamic targets based on deep learning is the completion of the perception system, scene acquisition, normalized expression, and feature extraction. The whole process starts with the perception system, which uses a three-dimensional map library, 5G+GNSS, and inertial devices to collect multi-source data, complete the acquisition of information scale, time and space relationship scene information, and then perform normalized expression of multi-features to construct a three-dimensional operation scene; based on nonlinear least squares, relaxation method, Bayes estimation theory and other methods, superpixel segmentation and semantic analysis are performed to perform feature matching, and the information obtained from feature matching is used to establish a dynamic target perception autonomous learning model using a sparse convolutional neural network and motion estimation fusion model to complete the rapid recognition of dynamic targets and real-life map display.

[0141] In this application, multivariate feature normalization refers to the normalized expression of multivariate features such as scale, time, and relationship of multi-source field information. Target recognition refers to screening out the target part from spatial or temporal distribution data information and performing feature judgment.

[0142] Fig. 9 is a flow chart of a method for rapid dynamic target perception based on deep learning provided in this embodiment. Fig. 9 As shown, the method includes:

[0143] Step S902: the perception system perceives environmental data.

[0144] The perception system for live working sites mainly includes 5G, GNSS, inertial devices and high-precision three-dimensional map libraries.

[0145] Step S904: scene acquisition.

[0146] The information obtained by the perception system is used to obtain map data and the position information of operators and engineering vehicles.

[0147] Step S906: normalized expression and feature matching.

[0148] In view of the on-site working environment, based on multi-source data from inertial, 5G, GNSS and 3D map database, the normalized expression of multivariate features of information scale, time and space relationship is completed, and a 3D working scene is constructed. Superpixel segmentation and semantic analysis are performed based on nonlinear least squares method, relaxation method and Bayes estimation theory to perform feature matching. In addition, the map data is filtered and segmented, feature data is extracted, and real-time reconstruction of the 3D scene of the construction site is achieved through multiple feature matching. Among them, the feature extraction process is as follows: Fig.10 shown.

[0149] In this embodiment, the normalized expression mainly includes the normalized expression of information scale, the normalized expression of time series, and the global normalized expression.

[0150] 1. Normalization method of information scale

[0151] 1) Standardization: The purpose of standardization is to make the data encoding methods and data formats of different data sources the same, laying the foundation for the next step of data matching (code standardization process in data standardization).

[0152] 2) Matching. Data matching has two aspects. One is to match different attributes of the same thing from different data sources to make the data more complete. The other is to mark the same data from different data sources as duplicates, laying the foundation for the next step of screening. For example: duplicate customer names from different data sources.

[0153] 3) Screening: The main purpose of data screening is to select consistent dimensions as master data, that is, the consistent dimension data that is finally delivered.

[0154] 2. Time Series Normalization Method

[0155] Time series normalization mainly completes the time alignment of multi-source data. Due to the particularity of time series, for example, the pattern correlation between adjacent sequences and the fact that data is continuously generated in the time dimension, the selection of normalization methods should also be as conducive to subsequent algorithm / model work as possible.

[0156] 3. Global Normalization Method

[0157] On the basis of global normalization, combined with the idea of ​​windowing in wavelet / Fourier transform, windowed normalization is produced.

[0158] The maximum and minimum values ​​of the sequence in the window are used to normalize the time series in the window. This method solves the problem of sequence amplitude differences under different measurement conditions when the time span is large to a certain extent. Given a time series S, after S is windowed and segmented, S has a total of n sequences with a window length of L: S = {s1, s2, ..., s n}. Adaptively normalize S, and the normalized sequence is: R = {r1, r2, ..., r n},in

[0159]

[0160] Among them, α and β are the relationship factors of adjacent time series, i = {1, 2...n}, d i is the iteration process variable, d i-1 is the iteration process variable, r i is the normalized sequence, f i is the iteration process variable, f i-1 is the iteration process variable.

[0161] Step S908, identifying dynamic targets.

[0162] The target recognition process is to filter out the target part from the spatial or temporal distribution data information and make feature judgment. Machine learning is a useful tool to solve this kind of problem. In the process of machine learning, appropriate algorithms and classifiers are used and appropriate prior features are selected to identify dynamic targets.

[0163] Example 10

[0164] According to an embodiment of the present invention, a method for rapid perception of dynamic targets based on deep learning is provided.

[0165] In this embodiment, based on calibration methods such as sensor self-calibration and system joint calibration, the work site environment perception is carried out. The collected data is preprocessed to complete the interference factor filtering and feature extraction. Real-life map fusion includes feature-level and spatial unified benchmark fusion. By studying the effective feature point extraction method of three-dimensional map data, the real-time reconstruction of the three-dimensional scene of the construction site is achieved through multiple feature matching; combining the map data graphical representation technology and data automatic association analysis technology, the superpixel segmentation and semantic relationship analysis methods of typical environmental feature data are studied to achieve the segmentation of map data, and combined with deep learning, the sparse convolutional neural network architecture is used to realize the rapid recognition of typical dynamic targets in the real-life map of the construction site.

[0166] First, perform environmental perception. After completing data collection, obtain map data, inertial data, and 5G+GNSS data, etc. Then perform system calibration. Based on calibration methods such as sensor self-calibration and system joint calibration, perform environmental perception at the work site.

[0167] Secondly, information processing is performed. Information processing mainly includes two parts: data processing and data fusion. Data processing includes data preprocessing, multivariate segmentation, feature extraction, etc. Data fusion includes unified benchmark, feature level, etc. Through information processing, data preprocessing is performed on the collected data to complete interference factor filtering and feature extraction. Data fusion mainly unifies the benchmark and determines the feature level.

[0168] Finally, environmental reconstruction and target positioning are performed. Environmental reconstruction and target positioning mainly include: environmental modeling, target recognition and active monitoring. Environmental modeling is mainly based on environmental features, multiple feature matching, and three-dimensional reconstruction. Target recognition is mainly template matching and feature extraction. Active monitoring is mainly target positioning and target monitoring.

[0169] In this embodiment, the map data is segmented by combining the map data visualization technology and the data automatic association analysis technology, using the superpixel segmentation and semantic relationship analysis methods of typical environmental feature data. In combination with deep learning, a sparse convolutional neural network architecture is used to realize the recognition and active monitoring of typical targets in the real-life construction site map.

[0170] Embodiment 11

[0171] According to an embodiment of the present invention, a method for fast perception of dynamic targets based on deep learning is provided. The method is similar to the methods in Embodiments 8 to 10, except that the dynamic targets are identified.

[0172] Fig.11 is a flow chart of a method for dynamic target recognition according to an embodiment of the present invention. Fig.11 As shown, the method comprises the following steps:

[0173] Step S1102, obtaining information data.

[0174] The perception system senses the environment and obtains environmental data.

[0175] Step S1104, expand the column vector.

[0176] The inertial, 5G, and GNSS system error state vectors are defined as 15×1 column vectors:

[0177]

[0178] in, is the difference in attitude angle errors measured by IMU, 5G and GNSS at time k, defined as a three-dimensional column vector containing roll angle, pitch angle and yaw angle; is the zero bias value output by the gyroscope; δr k is the three-axis position error in the navigation system; δv k is the three-axis velocity error under the navigation system; is the zero bias value of the accelerometer output.

[0179] The relationship between angular velocity ω and acceleration a and the measured value can be expressed as:

[0180]

[0181] Among them, ω m To measure the angular velocity, n g is the angular velocity compensation in the navigation system, a m To measure acceleration, n a is the acceleration compensation in the navigation system.

[0182] The update equation described by the state space convolution model of the dynamic system can be expressed as:

[0183]

[0184] Where k is the discrete time, is the system noise; G k is the noise matrix coefficient; V k is the observation noise matrix; H k is the observation matrix; Z k is the observed quantity, X k+1 is the state of the system at time (k+1), X k+1 ∈R n , Z k+1 ∈R m is the observation matrix of the corresponding signal, f(x k ) is the state function of the system at time k, is the transfer matrix from IMU to navigation coordinate system, ω b is the angular velocity under IMU, a b is the acceleration under IMU. f(X k ) is defined as follows:

[0185]

[0186] in, is the difference in attitude angle error output by the gyroscope, is the zero bias value output by the accelerometer, is the zero bias value output by the gyroscope, Δt I The IMU sampling interval is Δt C The camera sampling interval is, is the attitude transfer matrix from the map pose to the IMU (from the installation structure of the IMU, we can know that is a fixed value), is the attitude angle error at time k.

[0187] According to the map feature point extraction and matching relationship, the transfer matrix R between frames can be obtained, and the attitude matrix in the navigation coordinate system can be expressed as:

[0188]

[0189] Among them, C 11 To C 33 Represents the acquired attitude data information, from which the corresponding attitude angle information is calculated:

[0190]

[0191] in, is the attitude angle at time k, ψkC is the yaw angle at time k, θ kC is the pitch angle at time k, γ kC is the rolling angle at time k. (7)

[0192] Among them, the attitude error as an observation can be expressed as:

[0193]

[0194] in, is the difference of the combined attitude angle of adaptive dynamic matching, is the attitude angle measured by IMU and 5G, The attitude angle measured by GNSS.

[0195] The posture error of the combined pose solution of adaptive dynamic matching can be expressed as:

[0196]

[0197] in, is the weight adjustment factor, is the average angular increment per unit time, Here, the IMU's angular increment change is used as the calculation amount of the weight adjustment factor γ, thus avoiding the problem of being unable to solve due to excessive positioning error or failure.

[0198] Step S1106: 3D reconstruction.

[0199] After obtaining the information scale, time and space relationship scene information, the multi-dimensional features are normalized and expressed to construct a three-dimensional operation scene.

[0200] Step S1108, input and output relationship.

[0201] During deep learning training, the optimizer will optimize the model parameters to find a set of optimal solutions. The input and output parameter relationship of each layer in the convolutional neural network can be closely related to the weight W of the neuron connection and the number of input and output channels.

[0202] Step S1110, analyzing the results through deep learning data features.

[0203] A dynamic target perception autonomous learning model is established using sparse convolutional neural network and motion estimation fusion model to complete dynamic target rapid recognition and real-life map display.

[0204] In this embodiment, the advantages of the IMU's instantaneous high-precision attitude solution and the absence of temporal drift in 5G and GNSS attitude calculations are integrated to solve the problem of random divergence in inertial attitude solution, give full play to the advantages of IMU attitude measurement, further improve the accuracy of attitude solution, and realize integrity map fusion recognition under dynamic motion.

[0205] Example 12

[0206] According to an embodiment of the present invention, a method for fast perception of dynamic targets based on deep learning is provided. Fig.12 As shown, the method includes:

[0207] Step S1201, obtaining posture information.

[0208] The three-axis accelerometer and three-axis gyroscope collect data, and the collected data is settled through strapdown inertial navigation to obtain position, speed and attitude.

[0209] Step S1202: motion control triggering.

[0210] Through the negative feedback of motion control, the zero velocity and zero angular velocity are adjusted and corrected, the advantages of IMU attitude measurement are brought into play, the accuracy of attitude solution is further improved, and the integrity map fusion under dynamic motion is realized.

[0211] Step S1203, obtaining location information.

[0212] Based on the three-dimensional map database, feature extraction and matching are performed to obtain the location information of engineering vehicles and operators.

[0213] Step S1204, performing error correction on the position information and posture information.

[0214] The obtained posture information, such as position, velocity and attitude data, and position information are input into the extended Kalman filter to perform error correction on the strapdown inertial navigation solution of the inertial device and the solution results of the position information of the map.

[0215] Step S1205, updating the posture.

[0216] Based on the position information, depth reconstruction is performed to achieve 3D reconstruction. Based on the position, velocity and attitude information calculated by the strapdown inertial navigation solution, loop monitoring and similarity calculation are performed. Based on the output results of 3D reconstruction and similarity calculation, the semi-dense map and posture are updated to obtain a complete scene.

[0217] Example 13

[0218] According to an embodiment of the present invention, a dynamic target rapid perception system based on deep learning is provided. Fig.13As shown, the system includes:

[0219] The perception module 132 is configured to perform environmental perception based on the joint calibration of the perception system and collect data from different data sources; the scene acquisition module 134 is configured to perform scene acquisition based on the collected data and obtain map data and posture information; the data processing module 136 is configured to perform normalized data preprocessing of multi-features based on the acquired map data and posture information, and perform feature matching on the preprocessed normalized data through superpixel segmentation and semantic analysis; the recognition module 138 is configured to perceive the dynamic target based on the data obtained by feature matching using a sparse convolutional network and motion estimation fusion.

[0220] The system in this embodiment can implement any one of the methods in Embodiments 8 to 12, and therefore, they will not be described in detail here.

[0221] Embodiment 14

[0222] According to an embodiment of the present invention, a non-contact safety distance active warning system is provided. Fig.14 As shown, it includes a visual identification and early warning platform 142, a positioning base station 143, and multiple mobile positioning terminals 141.

[0223] Visual identification warning platform 142 includes a non-contact safety distance active warning system based on 5G and deep learning, which runs on the Linux operating system. The main station is based on the Jiaweishi industrial computer TPCIPC-100, Intel Bay TrailJ1900 processor, and the processor has stable performance; the memory uses DDR3 memory technology and is 8G in size; the hard disk is a 512G SSD solid state drive and a 1T HDD mechanical hard disk. In terms of software, the wisdom tooth Windows and Linux platforms have the default RS232 protocol and can be changed to 422 / 485 protocol.

[0224] The positioning base station 143 uses the Qianxun Cube MC262M PRO GNSS positioning module, with a built-in dual-frequency RTK algorithm and the BDS B2a frequency band. The positioning accuracy is 0.02±1ppm and the speed measurement accuracy is <0.05m / s.

[0225] The mobile positioning terminal 141 uses PA-IMU-03D as the mobile positioning terminal, and its parameters are as follows: start-up time ≤ 5S, range ± 200° / s, zero bias repeatability ≤ 0.3° / h, zero bias stability ≤ 0.3° / h, scale factor nonlinearity ≤ 200ppm, scale factor repeatability ≤ 200ppm, bandwidth ≥ 300Hz, random walk coefficient ≤ 0.05° / h, type quartz accelerometer, range ± 10g (various ranges can be selected), bias ≤ 1mg, bias monthly comprehensive error < 5×10-5 g(1σ), bias temperature coefficient ≤50μg / ℃, scale factor monthly comprehensive error <80ppm(1σ), scale factor temperature coefficient <80ppm / ℃.

[0226] During the operation of the system, three mobile positioning terminals are installed at designated locations on the engineering vehicle (carrier), such as Fig.14 As shown, they are installed at A above the cab of the engineering vehicle, B at the highest point of the boom, and C above the operator's cab. The mobile positioning terminal at point A is responsible for measuring the actual position and posture of the vehicle, and the mobile positioning terminals at points B and C are used to measure the elongation length of the boom and the rotation angle of the boom relative to the carrier. With the position information of points A, B, and C, the real-time information of the position and angle of the engineering vehicle can be obtained.

[0227] The real-time information of the position and angle of the engineering vehicle is transmitted back to the visual identification and early warning platform. The platform compares and identifies the existing dangerous area coordinates with the engineering vehicle position information, and actively issues an early warning when the set early warning safety threshold is exceeded.

[0228] The visual identification and warning platform in this embodiment is equivalent to the non-contact safety distance active warning device in Example 7, and can implement the methods in the above-mentioned Examples 1 to 6, so they will not be repeated here.

[0229] Embodiment 15

[0230] An embodiment of the present invention further provides a storage medium having a program stored thereon, which enables a computer to execute the methods in embodiments 1 to 6 and 8 to 12 when the program is executed.

[0231] Optionally, in this embodiment, the above-mentioned storage medium may include but is not limited to: a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk, and other media that can store program codes.

[0232] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for rapid perception of dynamic targets based on deep learning, characterized in that: include: Based on the joint calibration of perception systems, environmental perception is performed to collect data from different data sources; Based on the collected data, the scene is acquired to obtain the map data and posture information; Based on the acquired map data and position information, normalized data preprocessing of multi-features is performed, and feature matching of the normalized data preprocessed is performed through superpixel segmentation and semantic analysis; Based on the data obtained by feature matching, a sparse convolutional network and motion estimation fusion are used to perceive the dynamic target; Among them, collecting data from different data sources includes: Based on inertial devices, the position and posture information of the engineering vehicles and workers at the construction site is obtained, and a position and posture motion constraint model is established by analyzing the motion characteristics of the engineering vehicles and workers; the position and posture errors of the workers and the engineering vehicles with motion constraints are estimated using the optimal estimation method; the position and posture information of the workers and the engineering vehicles is obtained using the weak strapdown navigation failure error model; By utilizing multi-frequency channel estimation, the signal strengths and multipath delay clusters of different frequencies are obtained. By utilizing a multipath spatiotemporal compensation model, accurate positioning in strong electric field interference is achieved to obtain auxiliary positioning information. By utilizing the auxiliary positioning information, the posture information of the operating personnel and the engineering vehicle is corrected to obtain corrected positioning data.

2. The method according to claim 1, characterized in that Based on the acquired map data and position information, performing normalized data preprocessing of multivariate features includes at least one of the following: Performing normalized data preprocessing of information scale on the acquired map data and position information; Perform time series normalization data preprocessing on the acquired map data and posture information; The acquired map data and posture information are subjected to global normalization data preprocessing.

3. The method according to claim 2, characterized in that The obtained map data and position information are preprocessed to normalize the information scale, including: The acquired map data and position information are standardized, wherein the standardization includes at least one of the following: unifying the data encoding methods of data from different data sources into the same method; unifying the data formats of data from different data sources into the same format; The standardized map data and the position information are matched, wherein the matching process includes at least one of the following: matching different attributes of the same object from different data sources together; marking the same data from different data sources as duplicates; The matched map data and position information are screened to select consistent dimension data.

4. The method according to claim 2, characterized in that: The acquired map data and posture information are preprocessed for time series normalization, including: based on the pattern correlation between adjacent time series and the characteristics of continuous data generation in the time dimension, the acquired map data and posture information from different data sources are temporally aligned for the benefit of subsequent algorithms.

5. The method according to claim 2, characterized in that: The global normalization data preprocessing of the acquired map data and posture information includes: After windowing and segmenting the time series, multiple time subsequences with fixed window lengths are obtained; Based on the relationship factor of adjacent time series, the time subsequence in the current window is subjected to normalized data preprocessing by utilizing the maximum value and the minimum value of the time subsequence in the current window.

6. The method according to claim 1, characterized in that Before performing feature matching on the data preprocessed by the normalized data through superpixel segmentation and semantic analysis, the method further includes: defining the error state vector of the perception system as a column vector; Determine the relationship between the angular velocity and acceleration of the inertial device of the perception system and the measurement value based on the column vector; Determining an update equation for the inertial device based on the determined relationship between the angular velocity and acceleration of the inertial device and the measurement value; The attitude error of the inertial device is determined using the determined update equation.

7. The method according to claim 6, characterized in that Determining the attitude error of the inertial device using the determined update equation includes: According to the map feature point extraction and matching relationship, a transfer matrix between frames of the map data is obtained; The transfer matrix is ​​used to calculate the corresponding attitude angle information, and the angle increment change value of the inertial device is used as a weight adjustment factor to calculate the attitude error based on the attitude angle information.

8. A non-contact safety distance active warning method, characterized in that: include: Execute the method for rapid perception of dynamic targets based on deep learning as described in any one of claims 1 to 7; Active warning will be issued when the perceived dynamic target exceeds the safety distance threshold.

9. A dynamic target rapid perception system based on deep learning, characterized in that: include: The perception module is configured to perform environmental perception based on the joint calibration of the perception system and collect data from different data sources; A scene acquisition module is configured to perform scene acquisition based on the collected data, and acquire map data and posture information; A data processing module is configured to perform normalized data preprocessing of multivariate features based on the acquired map data and position information, and perform feature matching on the normalized data preprocessed by superpixel segmentation and semantic analysis; A recognition module is configured to perceive the dynamic target based on the data obtained by feature matching by using a sparse convolutional network and motion estimation fusion; The perception module is further configured as follows: Based on inertial devices, the position and posture information of the engineering vehicles and workers at the construction site is obtained, and a position and posture motion constraint model is established by analyzing the motion characteristics of the engineering vehicles and workers; the position and posture errors of the workers and the engineering vehicles with motion constraints are estimated using the optimal estimation method; the position and posture information of the workers and the engineering vehicles is obtained using the weak strapdown navigation failure error model; Using multi-frequency channel estimation, we can obtain the signal strength and multipath delay clusters at different frequencies. Using the multipath time-space compensation model, we can achieve accurate positioning in strong electric field interference to obtain auxiliary positioning information. The auxiliary positioning information is used to correct the position information of the operator and the engineering vehicle to obtain corrected positioning data.

10. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed, the computer is caused to execute the method according to any one of claims 1 to 8.

Images