A multi-target automatic tracking method and system based on unmanned intelligent turntable

Abstract

The present application provides a multi-target automatic tracking method and system based on an unmanned intelligent turntable. Among them, the method uses a detector to synchronously collect electromagnetic echoes and environmental depth data to construct an observation data set; separate thermal radiation, surface texture and motion features, establish a four-dimensional feature space with time and space coordinates and identify occluded areas; dynamically adjust signal parameters to enhance the characteristics of the occluded area and feedback optimization records; calculate the optimal scanning path based on the parameter optimization value, drive the sensor array to scan at multiple angles to generate a weighted data set; combine historical trajectory data, optimize the trajectory prediction model through distributed computing and parallel optimization, coordinate turntable parameter synchronization and decision-making, and finally achieve efficient reconstruction of the continuous trajectory of multiple targets in an occluded environment. The technical solution provided by the present application realizes the reconstruction of the trajectory of multiple targets in an occluded environment through multi-dimensional feature fusion and dynamic parameter optimization, thereby improving the environmental adaptability and tracking efficiency of the unmanned system.

Description

Technical Field

[0001] The present application relates to the technical field of multi-target intelligent tracking in an occluded environment, and in particular to a multi-target automatic tracking method and system based on an unmanned intelligent turntable. Background Art

[0002] In civilian scenarios such as intelligent monitoring of urban transportation hubs, crowd control in large venues, and wildlife observation in nature reserves, unmanned intelligent turntable systems must achieve uninterrupted tracking of multiple types of dynamic targets in densely obscured environments. The core technical requirements for such applications are that the system must have multimodal data synchronization capabilities, effectively integrating electromagnetic sensing and environmental depth data to establish a unified spatiotemporal benchmark; the system must have adaptive feature extraction capabilities, dynamically enhancing identifiable features when the target is partially obscured; and the system must have collaborative perception and trajectory continuity assurance capabilities, achieving trajectory prediction and reconstruction during occlusion periods through multi-angle scanning optimization and distributed computing, meeting the needs of highly robust tracking in complex scenarios.

[0003] The current mainstream approach to multi-target tracking in densely occluded environments is based on single-sensor sequential scanning and static feature matching. This approach uses a single high-precision sensor to periodically scan the monitoring area along a pre-set fixed path. After identifying the target using background subtraction or motion detection algorithms, the system extracts a universal template stored in a feature library for matching and positioning. When occlusion is detected, the system uses a linear motion model based on the target's historical trajectory to predict the current position, while maintaining the original scanning path and parameters. All data processing and trajectory estimation are centralized in the processor, with scanning resources allocated according to pre-set priority rules.

[0004] The defects of the existing solutions are mainly manifested in that, although its single-sensor sequential scanning mechanism can maintain basic monitoring coverage, it cannot effectively fuse multimodal sensor data in dynamic occlusion scenarios; at the same time, although the static feature library matching strategy can handle conventional target recognition, it lacks the ability to enhance dynamic features when the target is partially occluded, and relying solely on the linear motion model to extrapolate the position is prone to trajectory interruption and information loss; more prominently, although the centralized processor processing architecture simplifies data flow management, it has a large computing power delay in the face of the computing requirements of concurrent tracking of multiple targets, resulting in an increase in the target association error rate during the occlusion period. Summary of the Invention

[0005] The present application provides a multi-target automatic tracking method and system based on an unmanned intelligent turntable, which is used to solve the problems of trajectory breakage, feature loss and parameter synchronization lag in the existing technology when tracking multiple targets in dynamic occlusion scenarios.

[0006] In a first aspect, the present application provides a multi-target automatic tracking method based on an unmanned intelligent turntable, comprising:

[0007] The electromagnetic echo signal and the environmental depth data of the target area are collected by the detector of the unmanned intelligent turntable, and the electromagnetic echo signal and the environmental depth data are synchronously calibrated in the time dimension to form an observation data set;

[0008] Based on the physical property differences of different sensors in the observation data set, the target's thermal radiation characteristics, surface texture characteristics, and motion state characteristics are separated, and a four-dimensional feature space containing time and space coordinates is constructed. At the same time, abnormal areas on the target surface are identified as occlusion areas;

[0009] Adjusting the signal parameters of the detector to enhance the target features of the occluded area according to the target distribution density in the four-dimensional feature space and the feature conflict strength of the occluded area, and feeding the adjusted signal parameters back to the four-dimensional feature space to form a feature enhancement record;

[0010] calculating an optimal scanning path based on the optimized signal parameter values ​​in the feature enhancement record and the collected electromagnetic echo signal, and driving the sensor array to perform multi-angle scanning according to the optimal scanning path to generate a weighted scanning data set;

[0011] Based on the scanning data set and historical trajectory data, a target trajectory prediction model is established, and the target trajectory prediction model is optimized in parallel using distributed computing nodes to coordinate the parameter synchronization and decision optimization of the unmanned intelligent turntable, thereby realizing the reconstruction of continuous trajectories of multiple targets in an occluded environment.

[0012] Optionally, the target trajectory prediction model is established based on the scanning data set and historical trajectory data, and the target trajectory prediction model is optimized in parallel using distributed computing nodes to coordinate parameter synchronization and decision optimization of the unmanned intelligent turntable to achieve reconstruction of continuous trajectories of multiple targets in an occluded environment, including:

[0013] extracting position coordinates, time stamps, and path weight values ​​from the scan data set, forming a spatiotemporal sequence of scan data based on the position coordinates and time stamps, and converting the path weight values ​​into trajectory confidence weights;

[0014] Aligning the position sequence in the historical trajectory data with the spatiotemporal sequence of the scan data, establishing a dynamic displacement equation based on the trajectory confidence weight, and completing the construction of the target trajectory prediction model;

[0015] After deploying the target trajectory prediction model to distributed computing nodes, each node processes a subset of target data and optimizes the target trajectory prediction model by exchanging parameters of the dynamic displacement equation to minimize the spatial deviation between the predicted position and the scan data;

[0016] Based on the optimization result, the unmanned intelligent turntable is coordinated to feed back the predicted position to the detector unit to adjust the scanning angle to focus on the predicted area, and at the same time, the trajectory confidence weight is compared with the set threshold to adjust the signal transmission power of the detector;

[0017] The prediction process of the target trajectory prediction model and the parameter updating process of the dynamic displacement equation are iteratively performed in an occlusion environment, and the trajectory interruption points are connected by extrapolation calculation of the dynamic displacement equation to generate a continuous trajectory sequence.

[0018] Optionally, the generating a continuous trajectory sequence by connecting trajectory interruption points through extrapolation calculation of the dynamic displacement equation includes:

[0019] When a trajectory interruption is detected, the last valid position coordinate is recorded as the starting interruption point, and the first recaptured position coordinate is recorded as the ending interruption point;

[0020] Using the starting interruption point as the initial position, applying the dynamic displacement equation to predict the position at the next time point, inserting the predicted position as a virtual scanning point into the scanning data set and assigning a decaying trajectory confidence weight to form an extended data set, and updating the parameters of the dynamic displacement equation based on the extended data set;

[0021] Iteratively predict from the starting interruption point in chronological order to an intermediate time point, using the position of the intermediate time point at this time as the forward prediction position, and iteratively predict from the ending interruption point in chronological reverse order to the same intermediate time point, using the position of the intermediate time point at this time as the reverse prediction position;

[0022] When the spatial deviation between the forward prediction position and the backward prediction position is less than a tolerance threshold, the starting interruption point, the intermediate connection point and the ending interruption point are connected to form a continuous trajectory segment.

[0023] Optionally, the calculating an optimal scanning path based on the signal parameter optimization values ​​in the feature enhancement record in combination with the collected electromagnetic echo signal, and driving the sensor array to perform multi-angle scanning according to the optimal scanning path to generate a weighted scanning data set includes:

[0024] generating a regional scanning priority based on the feature enhancement value in the feature enhancement record, wherein a region with a higher feature enhancement value corresponds to a higher scanning priority;

[0025] identifying low signal strength areas and high priority areas whose scanning priorities are greater than a preset threshold value based on the intensity distribution of the electromagnetic echo signal, taking the highest priority area as the starting point of the path, forcibly covering all the high priority areas and low signal strength areas while minimizing the total steering angle, and generating an optimal scanning path;

[0026] generating a steering angle sequence of the detector unit based on the optimal scanning path, driving the sensor array to scan according to the steering angle sequence, and obtaining the position coordinates, time stamp, electromagnetic echo signal value and environmental depth value of each scanning point;

[0027] A path weight value equal to the feature enhancement value of the area in which each scanning point is located is assigned, and the information of all the scanning points is integrated to form a scanning data set including the position coordinates, time stamp, electromagnetic echo signal value, environmental depth value and path weight value.

[0028] Optionally, based on the physical property differences of different sensors in the observation data set, separating the thermal radiation characteristics, surface texture characteristics, and motion state characteristics of the target, constructing a four-dimensional feature space containing time and space coordinates, and identifying abnormal areas on the target surface as occlusion areas, includes:

[0029] Extracting electromagnetic echo signal values ​​and environmental depth values ​​as well as corresponding time stamps and position coordinates based on differences in physical characteristics of different sensors in the observation data set;

[0030] Based on the change pattern of the electromagnetic echo signal value, the thermal radiation characteristics are separated by comparing the signal strength threshold; based on the local fluctuation pattern of the environmental depth value, the surface texture characteristics are separated by calculating the difference between adjacent depth values; based on the change trend of the position coordinates and time stamps, the motion state characteristics are separated by calculating the position offset of consecutive time points;

[0031] Constructing a four-dimensional feature space by associating corresponding time tags and position coordinates according to the thermal radiation feature, the surface texture feature, and the motion state feature;

[0032] In the four-dimensional feature space, abnormal areas on the target surface are identified by detecting mutations in the electromagnetic echo signal values ​​and the ambient depth values. When the surface texture features or the thermal radiation features of the abnormal area show an abnormal decrease in intensity value in a local area, it is marked as an occluded area.

[0033] Optionally, adjusting the signal parameters of the detector to enhance the target features of the occluded area according to the target distribution density in the four-dimensional feature space and the feature conflict strength of the occluded area, and feeding the adjusted signal parameters back to the four-dimensional feature space to form a feature enhancement record, includes:

[0034] Counting the number of feature points in a unit space area according to the degree of aggregation of position coordinates in the four-dimensional feature space to calculate the target distribution density;

[0035] In the occlusion area, based on the intensity value deviations of the surface texture feature and the thermal radiation feature, respectively comparing the intensity values ​​of the surface texture feature and the thermal radiation feature in the normal area and the occlusion area to calculate corresponding feature conflict intensities;

[0036] According to the target distribution density and the feature conflict strength, the signal parameters of the detector are adjusted by a signal parameter optimization rule to enhance the target features of the occluded area;

[0037] The adjusted signal parameters are applied to the detector, the signal of the occluded area is recollected and the values ​​of the target distribution density and the feature conflict strength in the four-dimensional feature space are updated, and the adjustment process of the signal parameters, the target distribution density and the feature conflict strength is recorded to form a feature enhancement record.

[0038] Optionally, the detector of the unmanned intelligent turntable collects electromagnetic echo signals and environmental depth data of the target area, and synchronously calibrates the electromagnetic echo signals and the environmental depth data in the time dimension to form an observation data set, including:

[0039] A detector group including an electromagnetic sensor and a depth sensor is set on the unmanned intelligent turntable. The electromagnetic sensor transmits electromagnetic waves at fixed time intervals and receives reflected signals to generate electromagnetic echo signals. The depth sensor obtains depth information of the target area as environmental depth data based on optical principles.

[0040] When the electromagnetic sensor and the depth sensor generate data, a time stamp uniformly managed by the internal clock of the unmanned intelligent turntable is automatically added, and the electromagnetic echo signal and the environmental depth data are decomposed into discrete data points including a time stamp, a position coordinate, and a corresponding signal strength value;

[0041] Based on the time stamp, the electromagnetic echo signal is paired one-to-one with the discrete data points of the environmental depth data to form a data unit, and the data unit is integrated into a structured data set including the time stamp, location coordinates, and corresponding signal strength values ​​as an observation data set.

[0042] In a second aspect, the present application provides a multi-target automatic tracking system based on an unmanned intelligent turntable, comprising:

[0043] An acquisition module is used to collect electromagnetic echo signals and environmental depth data of the target area through a detector of an unmanned intelligent turntable, and synchronously calibrate the electromagnetic echo signals and the environmental depth data in the time dimension to form an observation data set;

[0044] A construction module is used to separate the thermal radiation characteristics, surface texture characteristics, and motion state characteristics of the target based on the physical property differences of different sensors in the observation data set, construct a four-dimensional feature space containing time and space coordinates, and identify abnormal areas on the target surface as occlusion areas;

[0045] an adjustment module, configured to adjust the signal parameters of the detector to enhance the target features of the occluded area according to the target distribution density in the four-dimensional feature space and the feature conflict strength of the occluded area, and feed the adjusted signal parameters back to the four-dimensional feature space to form a feature enhancement record;

[0046] a scanning module, configured to drive the sensor array to perform multi-angle scanning, calculate an optimal scanning path based on the signal parameter optimization values ​​in the feature enhancement record and the collected electromagnetic echo signal, so as to generate a weighted scanning data set;

[0047] The prediction module is used to establish a target trajectory prediction model based on the scanning data set and historical trajectory data, and to use distributed computing nodes to parallel optimize the target trajectory prediction model to coordinate the parameter synchronization and decision optimization of the unmanned intelligent turntable, thereby realizing the reconstruction of continuous trajectories of multiple targets in an occluded environment.

[0048] In a third aspect, the present application provides a computing device comprising a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement a multi-target automatic tracking method based on an unmanned intelligent turntable as described in the first aspect above.

[0049] In a fourth aspect, the present application provides a computer storage medium storing a computer program. When the computer program is executed by a computer, it implements a multi-target automatic tracking method based on an unmanned intelligent turntable as described in the first aspect.

[0050] In an example of the present application, the electromagnetic echo signal and environmental depth data of the target area are collected by the detector of the unmanned intelligent turntable, and the electromagnetic echo signal and the environmental depth data are synchronously calibrated in the time dimension to form an observation data set; based on the physical property differences of different sensors in the observation data set, the thermal radiation characteristics, surface texture characteristics and motion state characteristics of the target are separated, and a four-dimensional feature space containing time and space coordinates is constructed, and abnormal areas on the target surface are identified as occlusion areas; according to the target distribution density in the four-dimensional feature space and the feature conflict intensity of the occlusion area, the signal parameters of the corresponding detector are adjusted to enhance the target features of the occlusion area, and the adjusted signal parameters are fed back to the four-dimensional feature space to form a feature enhancement record; based on the signal parameter optimization value in the feature enhancement record, the optimal scanning path is calculated in combination with the collected electromagnetic echo signal, and the sensor array is driven to perform multi-angle scanning according to the optimal scanning path to generate a weighted scanning data set; based on the scanning data set and historical trajectory data, a target trajectory prediction model is established, and the target trajectory prediction model is parallel optimized using distributed computing nodes to coordinate the parameter synchronization and decision optimization of the unmanned intelligent turntable to achieve reconstruction of continuous trajectories of multiple targets in an occlusion environment.

[0051] The technical solution of this application has the following beneficial effects:

[0052] This application effectively solves the technical problems of target feature degradation and trajectory breakage in complex occlusion environments through the time synchronization calibration of multimodal sensor data and the dynamic construction of a four-dimensional feature space. By separating the characteristics of thermal radiation, surface texture, and motion state, and combining a feedback mechanism for the intensity of feature conflict in the occlusion area, it significantly improves the robust recognition capability of target features in dynamic occlusion scenarios. Based on a multi-angle scanning strategy driven by signal parameter optimization and a distributed parallel optimization model, it realizes the coordinated processing of dynamic adjustment of the scanning path and trajectory prediction, overcoming the performance bottleneck of the traditional single-sensor scanning framework in dense occlusion scenarios and ensuring the integrity and stability of multi-target continuous trajectory reconstruction.

[0053] Furthermore, by fusing the position coordinates, time stamps, and path weight values ​​of the scan data set, a spatiotemporal sequence is constructed and aligned with the historical trajectory data. A dynamic displacement equation is established based on the trajectory confidence weight to form a target trajectory prediction model. After the model is deployed to distributed computing nodes, each node processes a subset of the data in parallel and exchanges parameters to optimize the model, minimizing the spatial deviation between the predicted position and the scan. The turntable scanning angle and detector signal power are dynamically adjusted based on the optimization results. At the same time, prediction and parameter updates are iteratively performed in an occluded environment. The trajectory interruption points are connected through extrapolation of the dynamic displacement equation to generate a continuous trajectory sequence. This method solves the technical problems of multi-target trajectory breakage and association errors in occluded environments through distributed parallel optimization and iterative updates of the dynamic displacement equation, improves the accuracy and continuity of trajectory prediction, and enhances the completeness of trajectory reconstruction. At the same time, the dynamic signal power adjustment mechanism based on the scan confidence weight effectively reduces the target missed detection rate in densely occluded scenes and ensures the robustness of multi-target tracking in complex environments.

[0054] These and other aspects of the present application will become more readily apparent from the description of the following embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0056] Figure 1 A flowchart of a multi-target automatic tracking method based on an unmanned intelligent turntable provided by the present application is shown;

[0057] Figure 2 A scene diagram showing a multi-target automatic tracking method based on an unmanned intelligent turntable provided by the present application is shown;

[0058] Figure 3 The present invention provides a schematic structural diagram of a multi-target automatic tracking system based on an unmanned intelligent turntable;

[0059] Figure 4 A schematic structural diagram of a computing device provided by the present application is shown. DETAILED DESCRIPTION

[0060] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0061] In some of the processes described in the specification and claims of this application and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this document or may be executed in parallel. The serial numbers of the operations, such as 101, 102, etc., are only used to distinguish between different operations, and the serial numbers themselves do not represent any order of execution. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to being different types.

[0062] Research shows that existing multi-target tracking schemes in densely occluded environments rely on single-sensor temporal scanning and static feature matching mechanisms. Their fixed-path scanning mode makes it difficult to integrate the multimodal spatiotemporal information of electromagnetic echoes and environmental depth data, resulting in feature misalignment in complex occlusion scenarios such as gaps between buildings. Static feature library matching strategies cannot dynamically enhance the local identifiable features of occluded targets, and the extrapolated trajectory of linear motion models is prone to breakage. The centralized processor architecture lacks computing power when tracking multiple targets concurrently, which can easily cause delays in scanning parameter optimization and target association errors during the occlusion period, restricting the reliability of trajectory continuity reconstruction.

[0063] To address the above issues, the present invention proposes a multi-target automatic tracking method based on an unmanned intelligent turntable. The core of this method is to construct a dynamically optimizable four-dimensional feature space through the spatiotemporal synchronization calibration of electromagnetic echoes and environmental depth data, and to achieve parallel optimization of scanning parameter adaptive adjustment and trajectory prediction models based on distributed computing. This method solves the information limitations of a single sensor through the spatiotemporal fusion of multimodal sensing data. The dynamic feature enhancement mechanism overcomes the feature degradation defects of static template matching in occluded scenarios. The distributed parallel optimization architecture reduces the trajectory prediction delay to the millisecond level, improving computing efficiency compared to traditional centralized processing architectures. The adaptive adjustment of signal parameters improves the accuracy of target feature extraction in occluded areas, thereby improving the reliability and integrity of continuous trajectory reconstruction in complex scenarios.

[0064] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without making creative efforts are within the scope of protection of this application.

[0065] Figure 1 The present invention provides a flowchart of a multi-target automatic tracking method based on an unmanned intelligent turntable. Figure 1 As shown, the method includes:

[0066] 101. Collect electromagnetic echo signals and environmental depth data of the target area through a detector of an unmanned intelligent turntable, and synchronously calibrate the electromagnetic echo signals and the environmental depth data in a time dimension to form an observation data set;

[0067] Optionally, step 101 may specifically include the following steps:

[0068] 1011. A detector group including an electromagnetic sensor and a depth sensor is arranged on the unmanned intelligent turntable. The electromagnetic sensor transmits electromagnetic waves at fixed time intervals and receives reflected signals to generate electromagnetic echo signals. The depth sensor obtains depth information of the target area as environmental depth data based on optical principles.

[0069] 1012. When the electromagnetic sensor and the depth sensor generate data, a time stamp uniformly managed by the internal clock of the unmanned intelligent turntable is automatically added, and the electromagnetic echo signal and the environmental depth data are decomposed into discrete data points including a time stamp, a position coordinate, and a corresponding signal strength value;

[0070] 1013. Based on the time stamp, the electromagnetic echo signal is paired one-to-one with the discrete data points of the environmental depth data to form a data unit, and the data unit is integrated into a structured data set including the time stamp, location coordinates, and corresponding signal strength values ​​as an observation data set.

[0071] In the above scheme, the electromagnetic echo signal refers to the time-varying signal generated by the electromagnetic sensor emitting electromagnetic waves and receiving the target's reflected signal. It contains the target's position coordinates, motion state, and material properties, and can be used for target detection, classification, and tracking. Environmental depth data refers to discrete point cloud data obtained based on optical principles. It reflects the spatial distance between the target and the sensor and the three-dimensional structure of the scene, and can be used to establish a unified spatiotemporal reference and obstacle modeling. Discrete data points are atomic data units composed of timestamps, spatial coordinates, and signal strength parameters, providing structured input for feature extraction and multi-sensor correlation analysis. Data units are combinations of electromagnetic echo signals and environmental depth data points linked by timestamps. They reflect the physical state and spatial distribution characteristics of the target at a specific moment and support cross-modal data fusion. Structured datasets are multidimensional observation data sets stored in a unified format, containing time-series-aligned electromagnetic features, geometric features, and environmental context information, supporting efficient querying and distributed computing.

[0072] In an embodiment of the present application, first, through step 1011, a detector group consisting of an electromagnetic sensor and a depth sensor is installed on the robotic arm or movable base of the unmanned intelligent turntable. The electromagnetic sensor generates an electromagnetic echo signal by periodically emitting electromagnetic pulses and receiving the reflected signal of the target area. Its operating frequency and transmission power can be dynamically adjusted according to the complexity of the environment. For example, a high-frequency narrow beam is used in the metal detection scene. The depth sensor obtains the depth distribution map of the target area based on optical principles and generates environmental depth data reflecting the geometric characteristics of the space, such as structured light projection or ToF time of flight measurement. The electromagnetic sensor and the depth sensor are physically connected through the hardware interface of the turntable control module, and the timing consistency of the electromagnetic emission and depth sampling is ensured by a synchronization signal trigger. For example, after each pulse is emitted by the electromagnetic sensor, the depth sensor immediately starts single-frame acquisition to avoid data misalignment due to time difference.

[0073] Then, in step 1012, when the electromagnetic sensor and depth sensor generate raw data, they automatically attach a timestamp generated by the turntable's internal high-precision clock, which has microsecond accuracy, to form a raw data stream with time stamps. For example, the electromagnetic echo signal is recorded as "timestamp + reception intensity + azimuth / elevation angle," while the environmental depth data is marked as "timestamp + pixel coordinates + depth value." The system then uses a data parsing module to split the continuous signal stream into discrete data points based on the timestamp. Each data point contains time information, spatial coordinates, and the corresponding signal's electromagnetic echo amplitude or grayscale value of the depth value. This process is implemented using a preset parsing protocol, such as using JSON format to encapsulate data fields to ensure compatibility between data structures of different sensors.

[0074] Finally, in step 1013, the electromagnetic echo data and the environmental depth data are matched based on the unified timestamp. For example, the electromagnetic echo intensity value of 50dB at the timestamp t=0.5s and the depth data at the same time (coordinates x=1.2m, y=3.4m, depth 2.1m) form a data unit. During the matching process, the system uses a sliding window algorithm or a time tolerance threshold to process short misalignments caused by sensor delays and eliminates unmatched items. Finally, all data units are integrated into a structured data set as an observation data set in time series, using standardized fields. Storage supports direct input into machine learning models or 3D reconstruction algorithms. For example, the observation dataset generated by a single scan can contain thousands of records, each of which accurately describes the target's position, electromagnetic characteristics, and spatial depth at a specific moment, providing complete input for subsequent environmental modeling and target recognition.

[0075] In practical applications, in industrial warehouse inspection scenarios, an unmanned intelligent turntable is equipped with a detector system consisting of a 24GHz millimeter-wave radar and a structured light depth camera, synchronously collecting data from the target area at a rate of 20 frames per second. As the turntable scans the shelf area, the millimeter-wave radar emits a conical beam at 50ms intervals to capture electromagnetic echo signals from the metal casing of the moving AGV. The depth camera uses infrared structured light projection to generate a 1280×720 resolution depth point cloud. The system uses microsecond synchronization triggers to ensure strict alignment between radar pulse emission and depth camera exposure. All data is accurately timestamped by the turntable's main control chip. The data analysis module breaks down the radar echo into discrete points containing timestamps, polar coordinates, and signal strength, and simultaneously parses the depth data into timestamps, pixel coordinates, and depth values. Using a sliding window matching algorithm, the radar data point with an intensity of 61dB and coordinates of θ = 13.1° at time t = 1622543950.124000 is dynamically associated with the depth point in the same frame at a depth of 3.02m and world coordinates of x = 2.31m, y = 1.87m, and z = 3.02m, forming a structured data unit. A 10-second scan generates a total of 200 matching frames. Each record contains the following standardized fields: The dataset successfully integrates the electromagnetic reflection characteristics and three-dimensional spatial position of the AGV, providing cross-modal input for subsequent multi-target motion trajectory reconstruction, with a measured positioning error of less than 0.1m.

[0076] The above-mentioned 101 overall solution realizes deep correlation perception of electromagnetic characteristics and geometric forms through multi-modal sensor collaborative acquisition and high-precision spatiotemporal synchronization technology. Electromagnetic sensors and depth sensors form a detector group on an unmanned turntable, and achieve microsecond-level time alignment based on a unified clock source to eliminate the time misalignment problem of cross-modal data; the data parsing engine breaks down the continuous signal stream into discrete data points with timestamps, processes short delays between sensors through a dynamic time warping algorithm, and uses a spatial interpolation algorithm to compensate for equipment installation differences, and finally integrates the matched multi-dimensional data units into a standardized observation data set in time series. Through spatiotemporal consistency calibration and structured fusion, this solution significantly improves target recognition accuracy and system decision reliability in complex industrial scenarios, providing high-fidelity input for dynamic obstacle trajectory prediction and three-dimensional environment modeling.

[0077] 102. Based on the physical property differences of different sensors in the observation data set, separate the target's thermal radiation characteristics, surface texture characteristics, and motion state characteristics, construct a four-dimensional feature space containing time and space coordinates, and simultaneously identify abnormal areas on the target surface as occlusion areas;

[0078] Optionally, step 102 may specifically include the following steps:

[0079] 1021. Extracting electromagnetic echo signal values ​​and environmental depth values ​​as well as corresponding time stamps and position coordinates based on differences in physical characteristics of different sensors in the observation data set;

[0080] 1022. Based on the change pattern of the electromagnetic echo signal value, separate the thermal radiation feature by comparing the signal strength threshold; based on the local fluctuation pattern of the environmental depth value, separate the surface texture feature by calculating the difference between adjacent depth values; based on the change trend of the position coordinates and time stamps, separate the motion state feature by calculating the position offset of consecutive time points;

[0081] 1023. Construct a four-dimensional feature space by associating corresponding time tags and position coordinates according to the thermal radiation feature, the surface texture feature, and the motion state feature;

[0082] 1024. In the four-dimensional feature space, abnormal areas on the target surface are identified by detecting mutations in the electromagnetic echo signal value and the ambient depth value. When the surface texture feature or the thermal radiation feature of the abnormal area shows an abnormal decrease in intensity value in a local area, it is marked as an occluded area.

[0083] In the above scheme, thermal radiation characteristics refer to dynamic signals that reflect the temperature distribution characteristics of the target surface, including temperature intensity and spatial distribution parameters, which can be used to identify local overheating or low-temperature areas. Surface texture characteristics refer to dynamic signals that reflect the geometric deformation characteristics of the target surface, including local gradient amplitude and texture complexity parameters, which can be used to detect surface scratches or concave-convex deformations. Motion state characteristics refer to dynamic signals that reflect the spatial displacement characteristics of the target, including motion vectors and trajectory curvature parameters, which can be used to track the target's motion path and speed changes. Occlusion areas refer to abnormal signal areas that reflect the obscured characteristics of the target surface, including electromagnetic intensity drops and depth gradient mutation parameters, which can be used to mark obstacles or self-occluded parts in the sensor's field of view.

[0084] In the embodiment of the present application, first, in step 1021, by analyzing the physical characteristics of the electromagnetic sensor and the environmental depth sensor, such as the electromagnetic wave reflection characteristics and the optical imaging principle, the electromagnetic echo signal value, the environmental depth value and the corresponding time stamp and position coordinates are extracted from the observation data set. For example, the following image is extracted from the observation data set: structured data.

[0085] Next, in step 1022, based on the differences in the physical characteristics of the sensors, a differentiation algorithm is used to separate the target features. The electromagnetic echo signal is subjected to a sliding average filter to eliminate instantaneous noise, and then based on the dynamic threshold, the high-intensity signal in the temperature abnormality area is separated, such as the 55dB thermal radiation feature on the surface of the metal workpiece; secondly, the environmental depth value is Gaussian smoothed to suppress high-frequency noise, and the Sobel operator is used to calculate the local gradient amplitude. Pixels with gradient values ​​exceeding 5cm / m are retained as surface texture mutation areas, such as the depth gradient mutation caused by metal scratches; finally, the displacement vector between consecutive frames is calculated by the Lucas-Kanade optical flow method, and the target motion trajectory is predicted in combination with the Kalman filter to output the velocity vector. and trajectory curvature parameters, such as the 0.1m / s lateral displacement and 0.05m / s longitudinal offset generated by the workpiece's rotational motion; ultimately achieving efficient separation of thermal radiation characteristics, surface texture characteristics, and motion state characteristics.

[0086] Then, in step 1023, the separated features are associated according to the time and space coordinates to construct a structure including time t, space coordinates (x, y, z), electromagnetic intensity , depth gradient depth and motion vector The four-dimensional feature space of . The fields t,x,y,z, , , , , achieving unified representation of multi-source heterogeneous data. For example, a frame of data is stored as: , providing high-dimensional feature input for subsequent target recognition and behavior prediction in complex scenarios.

[0087] Finally, in step 1024, an anomaly detection algorithm is used to identify occluded areas on the target surface in the four-dimensional feature space. A first-order difference operation is performed on the electromagnetic intensity and environmental depth gradient sequences, and a dynamic threshold is used to mark mutation points. A morphological closing operation is then used to aggregate adjacent mutation points to form a continuous occluded area outline. A strength verification mechanism based on surface texture gradients or thermal radiation intensity is used to distinguish between self-occlusion and external occlusion. For example, if the electromagnetic intensity at the edge of a metal workpiece drops sharply from 60dB to 30dB due to obstruction by the sensor bracket, and the depth gradient suddenly increases from 5cm / m to 20cm / m, the system marks it as an occluded area through multi-dimensional feature verification. This information can be used by the path planning module for dynamic obstacle avoidance.

[0088] In actual applications, in industrial quality inspection scenarios, the deployed unmanned turntable is equipped with high-frequency electromagnetic sensors and ToF depth cameras, which achieve microsecond time alignment through a synchronous controller. The electromagnetic sensor scans the surface of the metal casting at a frequency of 20Hz, generating a time-varying electromagnetic echo signal containing thermal radiation characteristics. The depth camera simultaneously obtains 3D point cloud data with a precision of 0.05mm. The turntable control module adds a 200ns precision timestamp to each frame of data. For example, the electromagnetic echo is marked as , the depth data is marked as . The data parsing module splits the continuous signal into discrete points, and matches the time-stamp aligned electromagnetic data points (1625000000.123ns, 62dB) with the depth data points (1625000000.123ns, 3.2mm) through the sliding window algorithm to form a data unit containing time-space coordinates and multi-dimensional features. A dynamic threshold is used to separate the thermal radiation features. After a 5-point sliding average filter is performed on the electromagnetic signal, a 58dB threshold is set to separate the local overheating area 62dB±2dB on the casting surface; the Sobel operator is used to calculate the depth gradient, and the surface texture mutation area of ​​the casting porosity defect with a gradient >8mm / m is identified; the displacement vector between consecutive frames is calculated by the Lucas-Kanade optical flow method, and the trajectory is predicted by combining the Kalman filter to output the velocity vector (0.15m / s, 0.08m / s) and curvature parameters. The constructed four-dimensional feature space includes time, space coordinates, electromagnetic intensity, depth gradient and motion vector. For example, a defect point is stored as In four-dimensional space, the electromagnetic intensity drop from 62dB to 32dB and the depth gradient sudden increase from 8mm / m to 25mm / m were detected by first-order difference detection. The abnormal area was aggregated by morphological closing operation and marked as a φ15cm circular occlusion area formed by the sensor bracket occlusion.

[0089] The overall solution described above, 102, uses multimodal sensor data fusion and feature decoupling technology to efficiently separate thermal radiation, surface texture, and motion state features based on the physical property differences between electromagnetic echoes and environmental depth values, thereby constructing a spatiotemporally aligned four-dimensional feature space. Dynamic threshold detection and morphological analysis are used to accurately identify abnormal areas on the target surface in the four-dimensional feature space, and a multi-dimensional feature strength verification mechanism is used to distinguish between self-occlusion and external occlusion. This method effectively integrates multi-source heterogeneous data, breaking through the perception limitations of a single sensor in complex industrial scenarios, significantly improving target recognition accuracy and obstacle avoidance response efficiency, and providing robust perception support for high-precision motion control.

[0090] 103. Adjusting the signal parameters of the detector to enhance the target features of the occluded area based on the target distribution density in the four-dimensional feature space and the feature conflict strength of the occluded area, and feeding the adjusted signal parameters back to the four-dimensional feature space to form a feature enhancement record;

[0091] Optionally, step 103 may specifically include the following steps:

[0092] 1031. Counting the number of feature points in a unit space area according to the degree of aggregation of position coordinates in the four-dimensional feature space to calculate target distribution density;

[0093] 1032. In the occlusion area, based on the intensity value deviations of the surface texture feature and the thermal radiation feature, respectively compare the intensity values ​​of the surface texture feature and the thermal radiation feature in the normal area and the occlusion area to calculate corresponding feature conflict intensities;

[0094] 1033. Adjust the signal parameters of the detector according to the target distribution density and the feature conflict intensity using a signal parameter optimization rule to enhance the target features of the occluded area.

[0095] 1034. Apply the adjusted signal parameters to the detector, re-collect the signal of the occluded area and update the values ​​of the target distribution density and the feature conflict intensity in the four-dimensional feature space, record the adjustment process of the signal parameters, the target distribution density and the feature conflict intensity, and form a feature enhancement record.

[0096] In the above scheme, target distribution density refers to the degree of aggregation of target feature points within a unit volume in the four-dimensional feature space. The spatial distribution characteristics of the target are quantified by counting the number distribution of feature points such as electromagnetic intensity and depth gradient in a specific area. Feature conflict intensity refers to the degree of difference between the occluded area and the normal area in surface texture, thermal radiation and other features. By comparing the mean, variance and gradient changes of multi-dimensional feature values, the feature anomaly caused by occlusion is quantified. Target features refer to the set of multi-dimensional physical properties of the target in the four-dimensional feature space, including electromagnetic echo intensity, environmental depth gradient, motion state parameters, etc., which reflect the geometric structure and dynamic behavior of the target. Feature enhancement records refer to the closed-loop feedback log formed by combining the parameter configuration sequence generated by adjusting the detector signal parameters and the changes in the target distribution density and feature conflict intensity in the re-collected data.

[0097] In this embodiment, step 1031 first involves dividing the space into fixed-size cells (e.g., 0.1 m × 0.1 m × 0.1 m) using a gridding method based on the degree of clustering of position coordinates (x, y, z) in the four-dimensional feature space. This gridding employs equidistant region partitioning to ensure consistent spatial resolution within each cell. The number of electromagnetic intensity and depth gradient feature points within each grid cell is then counted, and the statistical results are smoothed using Gaussian kernel density estimation to eliminate discrete noise and generate the target distribution density. Specifically, a kernel function is used to weight the contribution of each feature point, generating a continuous density distribution map. For example, in industrial inspection, surface defects on metal workpieces form high-density clusters due to dense electromagnetic echoes. A density heat map can visually display these defect concentrations. This process quantifies spatial distribution characteristics and provides a spatial benchmark for subsequent parameter optimization.

[0098] Next, in step 1032, the surface texture gradient and thermal radiation intensity are extracted from the occluded area, and the occluded area is separated from the normal area through morphological operations. For example, the occluded range of the stent is marked based on connected domain analysis. The feature values ​​of the occluded area and the normal state of the same target are compared to calculate the difference in surface texture gradient amplitude. The formula is as follows: , and calculate the thermal radiation intensity difference and variance change at the same time. The thermal radiation intensity difference formula is as follows: Weighted Euclidean distance is used to fuse multi-dimensional differences and generate a conflict intensity matrix. For example, sensor bracket occlusion causes the texture gradient to drop from 50 to 15, and the thermal radiation intensity to drop from 60dB to 28dB, resulting in a conflict intensity of 0.7. This process accurately distinguishes between self-occlusion and external occlusion through multi-feature comparison.

[0099] Then, in step 1033, a dynamic weight allocation algorithm is used to adjust detector parameters based on the target distribution density and feature conflict intensity. The electromagnetic sensor gain is reduced in high-density areas to prevent signal overexposure, while the gain is increased in low-density areas to enhance weak signals. The filter threshold is relaxed in high-conflict areas to preserve abnormal features. For example, the gain is reduced in areas with dense defects to prevent signal overexposure, while the gain is increased in areas with background noise to enhance weak signals. In high-conflict areas, the filter threshold is relaxed, adjusting the mean filter window from 5×5 to 3×3 to preserve abnormal features. The target trajectory is predicted using a Kalman filter, and the sampling frequency is adjusted to match the target's dynamic characteristics. A genetic algorithm is used to optimize the parameter combination, with the objective function being to maximize feature discernibility. The adjusted gain value and sampling interval are then output. This process dynamically adapts to environmental changes through a closed-loop feedback mechanism.

[0100] Finally, in step 1034, the adjusted signal parameters are applied to the detector to re-collect the electromagnetic echo and depth data of the occluded area, and the incremental learning algorithm is used to fuse the new data with the original data. The density heat map in the four-dimensional feature space is dynamically updated, and the change in conflict intensity is quantified through differential operations. For example, after adjustment, the electromagnetic intensity is restored from 45dB to 55dB, and the conflict intensity is reduced from 0.9 to 0.5; then the parameter adjustment sequence, density trend chart and conflict comparison table are stored in the log in Protobuf format to form an enhanced archive containing multi-dimensional feature change records. This archive supports subsequent algorithms to achieve adaptive parameter optimization through retrospective analysis. For example, in industrial detection, historical records can be used to quickly locate the sensor sensitivity attenuation law and improve the robustness of target recognition in complex occlusion scenarios.

[0101] In actual applications, in the unmanned turntable bearing defect detection scenario, the system found a high target distribution density in the bearing raceway area based on four-dimensional feature space analysis. There were 120 feature points within a grid of 0.05m³, and the average texture gradient in the normal area of ​​the bracket rotation occlusion area was 45cm / m and the thermal radiation was 58dB, while the occlusion area dropped sharply to 12cm / m and 32dB, and the feature conflict intensity reached 0.8. Based on this, the laser radar and millimeter wave radar parameters were dynamically adjusted, and the laser sampling rate was increased from 10Hz to 30Hz to capture surface microcracks under the high-speed rotation of the turntable 2πrad / s. At the same time, the millimeter wave transmission power was increased from 20dBm to 28dBm to penetrate the oil and dirt occlusion. After re-collection, the electromagnetic intensity in the occlusion area recovered to 53dB, the texture gradient rebounded to 38cm / m, and the conflict intensity dropped to 0.3. The feature space was updated to generate enhanced records. .

[0102] The above-mentioned 103 overall solution realizes the intelligent closed-loop optimization of the detector signal parameters by dynamically sensing the target distribution density in the four-dimensional feature space and accurately diagnosing the conflict intensity between the surface texture and thermal radiation characteristics of the occluded area. It can effectively identify the target cluster area, locate the root cause of the feature distortion caused by occlusion, and adaptively adjust the signal parameters accordingly, significantly enhancing the feature expression of the target in the occluded area, and improving its detectability and recognition in complex environments. The optimized parameters are immediately fed back to the detector for signal re-collection, updating the feature space data and verifying the enhancement effect. At the same time, the entire process of parameter adjustment, distribution density change and conflict intensity evolution is fully recorded to form a traceable feature enhancement history. Ultimately, this mechanism continuously improves the system's target detection robustness in occlusion scenarios, and provides reliable data support and decision-making basis for subsequent analysis and optimization.

[0103] 104. Calculate an optimal scanning path based on the signal parameter optimization values ​​in the feature enhancement record and the collected electromagnetic echo signal, and drive the sensor array to perform multi-angle scanning according to the optimal scanning path to generate a weighted scanning data set;

[0104] Optionally, step 104 may specifically include the following steps:

[0105] 1041. Generate a regional scanning priority based on the feature enhancement value in the feature enhancement record, where a region with a higher feature enhancement value corresponds to a higher scanning priority;

[0106] 1042. Identify low signal strength areas and high priority areas whose scanning priorities are greater than a preset threshold value based on the intensity distribution of the electromagnetic echo signal, use the highest priority area as the starting point of the path, forcibly cover all high priority areas and low signal strength areas while minimizing the total steering angle, and generate an optimal scanning path;

[0107] 1043. Generate a steering angle sequence of the detector unit based on the optimal scanning path, drive the sensor array to scan according to the steering angle sequence, and obtain the position coordinates, time stamp, electromagnetic echo signal value, and environmental depth value of each scanning point;

[0108] 1044. And assign a path weight value equal to the feature enhancement value of the area to each of the scanning points, integrate the information of all the scanning points, and form a scanning data set including the position coordinates, time stamp, electromagnetic echo signal value, environmental depth value and path weight value.

[0109] In the above scheme, regional scanning priority refers to a weighting factor generated based on the feature enhancement values ​​in the feature enhancement record. It quantifies the urgency of sensor scanning in different areas, prioritizing coverage of high-priority areas to improve detection efficiency. The optimal scanning path refers to the sensor motion trajectory generated by the path planning algorithm. It must cover all high-priority areas and low-signal-strength areas while minimizing the total steering angle to ensure scanning efficiency and data integrity. The steering angle sequence refers to the sequence of steering instructions for sensor units derived from the optimal scanning path. It contains parameters for angle change direction and amplitude, driving the sensor to perform multi-angle scanning along a preset path. The path weight value is a weighting factor dynamically assigned based on the feature enhancement value and environmental depth value of the scanning point. It is used to indicate the importance of the data at that point in the overall analysis. The scanning dataset refers to the set of raw data collected by the sensor array along the optimal path. It contains structured data such as location coordinates, time stamps, electromagnetic echo signal values, environmental depth values, and path weight values, supporting subsequent feature analysis and algorithm optimization.

[0110] In an embodiment of the present application, first, step 1041 is performed based on the historical scanning parameter optimization values ​​stored in the feature enhancement record and the target distribution density, conflict intensity and other feature parameters of the corresponding area, and a regional scanning priority matrix is ​​generated through a dynamic weight allocation algorithm. The matrix weights and fuses the feature enhancement value with the perception data of the electromagnetic echo intensity and thermal radiation deviation. For example, a decision model is constructed using a hierarchical analysis method, and the feature enhancement value is used as the criterion layer weight, and a dynamic priority is generated in combination with the degree of electromagnetic signal attenuation. High-priority areas will be marked as forced scanning targets to ensure that resources are allocated to key detection areas first. For example, an area with a feature enhancement value ≥ 0.8 and an electromagnetic intensity lower than 50% of the mean is marked as a high-priority area, and a forced scan is performed to ensure that resources are allocated to this area first.

[0111] Next, step 1042 geometrically associates the high priority area with the low signal strength area through spatial overlay analysis to construct a coverage constraint graph. Using the improved traveling salesman problem solving algorithm, starting from the highest priority area, dynamic programming is used to generate the shortest path covering all constraint nodes. For example, using A The algorithm optimizes local paths and incorporates a steering angle penalty function to enforce coverage of all high-priority and low-signal areas while minimizing the total steering angle. For example, after associating high-priority areas with signal attenuation areas, the system plans a spiral scanning path, reducing the total steering angle from 120° to 65°, ensuring full coverage of critical areas.

[0112] Then, in step 1043, a steering angle sequence for the detector unit is generated based on the optimal scanning path decomposition, such as a stepwise adjustment from 0° to 15° to 30°. The multi-degree-of-freedom robotic arm and the sensor array are driven to move synchronously according to the steering angle sequence. During the scanning process, four-dimensional spatiotemporal data is collected for each scanning point, including position coordinates, time stamps, electromagnetic echo signal values, and ambient depth values. For example, when the scanning path is decomposed into a steering instruction sequence from 0° to 15° to 30°, the lidar and infrared sensors are driven to synchronously collect data. The acquired scanning point position is (X=120.5mm, Y=30.2mm, Z=5.1mm), the timestamp is 2025-06-10T14:30:00.000Z, the electromagnetic intensity is -85dB, and the depth value is 5.12m.

[0113] Finally, step 1044 is used to assign a path weight value to each scanning point. This value is calculated by weighting the feature enhancement value and the environmental depth value of the area in a preset ratio. Subsequently, the system integrates the position coordinates, time stamp, electromagnetic echo signal value, environmental depth value and path weight value of the scanning point to construct a structured data set containing spatiotemporal metadata, signal features and weight identifiers to support subsequent machine learning model training and 3D reconstruction algorithm processing. For example, a scanning point is located in a high-priority corrosion area, has a feature enhancement value of 0.85, and an environmental depth of 1.1m. Its weight is calculated as The point data is marked as a high-confidence sample and used for subsequent defect classification model training.

[0114] In actual application, a defect point is stored as In four-dimensional space, the electromagnetic intensity drop from 58dB to 35dB and the depth gradient mutation from 10mm / m to 28mm / m were detected by first-order difference detection. The abnormal areas were aggregated by morphological closing operation and marked as the φ8cm fan-shaped occlusion area formed by the occlusion of the robotic arm support. The space was divided into 0.05m×0.05m×0.05m units according to the grid division method. Statistics showed that the density of feature points in the 0.3m×0.2m defect-dense area reached 1200 points / m³. The dynamic weight allocation algorithm was triggered to increase the electromagnetic sensor gain in the occlusion area from 1.2 times to 1.8 times to penetrate metal interference. At the same time, the background area filter window was adjusted from 7×7 to 5×5 to suppress noise. The sampling frequency was optimized by genetic algorithm, and the sampling interval of the occlusion area was shortened from 20ms to 5ms to capture dynamic features. After adjusting the parameters and incorporating new data through an incremental learning algorithm, the electromagnetic intensity in the four-dimensional feature space was restored to 52dB, and the conflict intensity was reduced from 0.85 to 0.35. Feature enhancement records show that this parameter combination increased the defect detection rate to 99.5%. This log includes the parameter adjustment sequence, density change curve, and conflict comparison table, providing a basis for adaptive optimization of subsequent algorithms.

[0115] The overall solution described above, 104, utilizes the optimized parameters in feature enhancement records, combined with the characteristics of electromagnetic echo signals, to achieve intelligent scanning control of the sensor array. By mapping feature enhancement values ​​to regional scanning priorities, key target areas and weak signal areas are dynamically targeted. A path optimization algorithm enforces coverage of all high-value areas while minimizing steering losses, generating an efficient scanning path. The sensor array is driven along this path to precisely collect multi-dimensional data synchronized in time and space, while assigning each scanning point a path weight derived from the feature enhancement value. Ultimately, a multi-dimensional scanning dataset is formed that integrates target value weights with physical signals, significantly improving the pertinence and completeness of data collection in complex scenarios and providing a structured information foundation with a value gradient for subsequent analysis.

[0116] 105. Based on the scanning data set and historical trajectory data, a target trajectory prediction model is established, and the target trajectory prediction model is optimized in parallel using distributed computing nodes to coordinate parameter synchronization and decision optimization of the unmanned intelligent turntable to achieve reconstruction of continuous trajectories of multiple targets in an occluded environment.

[0117] Optionally, step 105 may specifically include the following steps:

[0118] 1051. Extracting position coordinates, time stamps, and path weight values ​​from the scan data set, forming a spatiotemporal sequence of scan data based on the position coordinates and time stamps, and converting the path weight values ​​into trajectory confidence weights;

[0119] 1052. Align the position sequence in the historical trajectory data with the spatiotemporal sequence of the scan data, establish a dynamic displacement equation based on the trajectory confidence weight, and complete the construction of the target trajectory prediction model;

[0120] 1053. After deploying the target trajectory prediction model to the distributed computing nodes, each node processes a subset of the target data and optimizes the target trajectory prediction model by exchanging parameters of the dynamic displacement equation to minimize the spatial deviation between the predicted position and the scan data.

[0121] 1054. Based on the optimization result, the unmanned intelligent turntable is coordinated to feed back the predicted position to the detector unit to adjust the scanning angle to focus on the predicted area, and the trajectory confidence weight is compared with the set threshold to adjust the signal transmission power of the detector;

[0122] 1055. Iteratively execute the prediction process of the target trajectory prediction model and the parameter update process of the dynamic displacement equation in an occlusion environment, and connect the trajectory interruption points through extrapolation calculation of the dynamic displacement equation to generate a continuous trajectory sequence.

[0123] Among them, step 1055 specifically includes the following processes: when a trajectory interruption is detected, the last valid position coordinate is recorded as the starting interruption point, and the first recaptured position coordinate is recorded as the ending interruption point; the dynamic displacement equation is applied to predict the position of the next time point with the starting interruption point as the initial position, the predicted position is inserted into the scanning data set as a virtual scanning point and assigned an attenuated trajectory confidence weight to form an extended data set, and the parameters of the dynamic displacement equation are updated based on the extended data set; iteratively predicting from the starting interruption point to an intermediate time point in chronological order, and using the position of the intermediate time point at this time as the forward prediction position, and simultaneously iteratively predicting from the ending interruption point to the same intermediate time point in reverse chronological order, and using the position of the intermediate time point at this time as the reverse prediction position; when the spatial deviation between the forward prediction position and the reverse prediction position is less than the tolerance threshold, connecting the starting interruption point, the intermediate connection point and the ending interruption point to form a continuous trajectory segment.

[0124] In the above solution, historical trajectory data refers to the chronological position sequence of a moving object in space and time, consisting of coordinate points and millisecond-precision timestamps, which are used to analyze motion patterns and behavioral patterns. Trajectory confidence weights are quantitative metrics calculated based on the reliability of trajectory points. These weights are dynamically adjusted based on characteristics such as rate, repeatability, and round-trip time, to filter out noisy or anomalous trajectories. The target trajectory prediction model is a prediction framework that integrates physical dynamics with machine learning algorithms. It takes historical trajectories and environmental semantics as input and outputs a sequence of future positions. A continuous trajectory sequence is a complete spatiotemporal trajectory after eliminating interruptions, connecting virtual points in occluded intervals through an extrapolation algorithm.

[0125] In an embodiment of the present application, first, step 1051 extracts key features from the original scan data, including the three-dimensional coordinates, timestamp and regional weight value of the target. The position coordinates reflect the geometric position of the target, the timestamp records the time of data collection, and the regional weight value is calculated by signal strength or target density to identify the importance of the region. The system sorts the discrete data by time to form a spatiotemporal sequence, and suppresses noise through the Kalman filter algorithm to improve the accuracy of position estimation. For example, the dense point cloud collected by the lidar is filtered to retain stable feature points and remove outliers. Subsequently, the path weight value is mapped to a trajectory confidence weight in the range of 0-1 through the Sigmoid function. The weight of the high confidence area can reach 0.9, and the low confidence area is reduced to 0.3. This process provides high-quality input for subsequent prediction models to ensure that the algorithm focuses on key areas.

[0126] Then, in step 1052, the historical trajectory data is time-synchronized with the current spatiotemporal sequence, and cubic spline interpolation is used to fill time gaps and ensure consistent resolution between the two. For example, if the historical trajectory is 10 frames per second and the scanned data is 5 frames per second, linear interpolation is used to generate intermediate frame data. Based on the aligned data, the system establishes a dynamic displacement equation that combines the uniform velocity model with the predictive capabilities of the LSTM network, giving higher priority to areas with higher weights. For example, due to the high confidence weight of the elbow area of ​​the industrial pipeline, the displacement equation prioritizes matching the historical curvature parameters, generating a predicted trajectory that conforms to physical laws. At the same time, an adaptive learning rate mechanism is introduced to optimize the equation parameters, and the coefficients are adjusted using the gradient descent method to reduce prediction error.

[0127] Then, in step 1053, the trajectory prediction model is deployed to a distributed computing cluster, with each node independently processing a subset of the target data. Nodes use the Adam optimizer to adjust the parameters of the dynamic displacement equation, aggregate the gradients through the parameter server, and update the global model. For example, node A calculates the gradient of the displacement equation for a certain region and uploads it to the server. The server then integrates the gradients of all nodes and sends the updated value. The communication protocol uses an asynchronous messaging mechanism to reduce latency and improve throughput. In a smart warehousing scenario, when 10 AGVs are operating simultaneously, distributed optimization reduces path prediction latency by 40%, and the global consistency error is kept within 5 centimeters.

[0128] Then, based on the optimized predicted position, step 1054 coordinates the unmanned turntable to adjust the detector's scanning angle. For example, a PID controller is used to increase the turntable's speed to 200° / second to align with the predicted area. The system compares the trajectory confidence weight with a threshold. If the weight is below the threshold, the signal transmission power is reduced; otherwise, the power is increased to the peak value. For example, when a high-confidence target is detected in a security scenario, the turntable quickly turns and activates the millimeter-wave radar's high-power mode, enabling identification of small targets within 300 meters. Power regulation is combined with low-precision computing technology to reduce energy consumption while maintaining detection accuracy.

[0129] Finally, through step 1055, a repair mechanism based on dynamic displacement equations and bidirectional prediction is proposed to address the trajectory interruption problem caused by signal obstruction or equipment interference in complex environments. When the system detects that the trajectory signal is lost or the position jump exceeds the preset threshold, it first records the last valid position before the interruption as the starting interruption point, and captures the first valid position after the signal is restored as the ending interruption point. Taking the starting interruption point as the benchmark, the position of the next time step is predicted by extrapolation through the dynamic displacement equation, and a virtual scanning point is generated and assigned a confidence weight that decays with time, which is inserted into the original data set to form an extended data set. Subsequently, the system re-optimizes the parameters of the dynamic displacement equation based on the extended data set, such as adjusting the kinematic model coefficients through the gradient descent method to improve the prediction accuracy. After the parameters are updated, the system starts from the starting interruption point. and terminate breakpoints Iterate the forecast forward and backward according to the time series until the same intermediate time point is reached Based on the optimized dynamic equation, Start calculating the position frame by frame and generate a forward trajectory sequence .from Reverse calculation to generate backward trajectory sequence Calculate the two series at the same intermediate time point Position deviation at a given moment: like If the error is less than the tolerance threshold, the trajectory is judged to be repairable. This process introduces an adaptive tolerance mechanism, which dynamically adjusts the threshold according to the complexity of the environment and uses cubic spline interpolation to and exist The interpolation function is smoothly connected to generate continuous trajectory segments. For example, , where the coefficients a, b, c, d are given by the two end trajectories at The position, velocity, and acceleration at each point are calculated. Finally, the system merges the starting interruption point, intermediate connection points, and final interruption point into a complete trajectory and updates the global trajectory database. This step verifies the trajectory smoothness to ensure that the repaired trajectory conforms to the laws of physical motion.

[0130] In actual applications, in the unmanned turntable bearing ball tracking scenario, the system extracts the data point with coordinates x=0.325m, y=0.148m, z=0.082m at the timestamp of 12.800 seconds and its 0.7 path weight based on the scan data set, converts it into a trajectory confidence weight of 0.85, and then combines the historical trajectory to establish a dynamic displacement equation through distributed node parallel optimization. When the trajectory of the bearing ball is detected to be interrupted at the timestamp of 13.2 seconds, the coordinates of the last valid position point P1 are recorded as x=0.332m, y=0.153m, z=0.085m, and the coordinates of the first valid point P2 are recaptured at the timestamp of 13.6 seconds with coordinates of x=0.348m, y=0.162m, z=0.092m. Using P1 as the starting interruption point, the position of the timestamp of 13.4 seconds is forward predicted. x=0.340m, y=0.158m, z=0.088m; use P2 as the end point to reversely predict the position at the same time point x=0.341m, y=0.157m, z=0.089m; calculate the spatial deviation =0.0017 meters, less than the tolerance threshold of 0.005 meters. Continuous trajectory segments were generated using cubic spline interpolation, with interpolation coefficients a=0.008, b=-0.12, c=1.35, and d=0.085. The synchronized unmanned turntable increased the lidar scanning angular rate to 200 degrees per second, focusing on the predicted area. The millimeter-wave transmission power was increased to 30 decibel milliwatts to penetrate oil and dirt obstructions. After five iterations of optimization, the trajectory interruption repair rate reached 95.3%, and target tracking latency was reduced from 120 milliseconds to 28 milliseconds, achieving millimeter-level continuous reconstruction of the turntable's 12 ball bearing motion trajectories.

[0131] The overall solution of the aforementioned 105 constructs a dynamic displacement-driven target trajectory prediction model by fusing weighted scan data with historical trajectory information, and utilizes distributed computing to achieve parallel model optimization. The system first converts path weights into trajectory confidence weights to establish a spatiotemporal prediction model. Distributed nodes collaboratively optimize model parameters to continuously reduce the spatial deviation between the predicted position and the actual scan. Based on the optimization results, the unmanned turntable dynamically adjusts the scanning focus area and detector transmit power, forming a closed perception-decision loop. To address trajectory interruptions caused by occlusion, a bidirectional iterative prediction and virtual trajectory point insertion mechanism is employed, generating a continuous trajectory sequence through convergence verification of forward and reverse predictions. Ultimately, lossless reconstruction of multi-target trajectories is achieved in complex occlusion environments, significantly improving the robustness and continuity of trajectory prediction.

[0132] The following is a complete example for steps 101 to 105. Figure 2 As shown in the figure, in the auto parts warehouse, the unmanned intelligent turntable is equipped with a 77GHz millimeter-wave radar and a binocular stereo depth camera, which synchronously collects AGV channel data at a frequency of 30Hz. When the turntable monitors 6 cargo-carrying AGVs, the millimeter-wave radar transmits frequency-modulated continuous waves at 33ms intervals to capture the strong electromagnetic echoes of the metal shelves and the medium-intensity reflections of the AGV body; the depth camera generates a 1920×1080 depth map with an accuracy of ±2mm through parallax calculation. The system uses hardware synchronization triggers to ensure that the radar emission pulse is strictly aligned with the camera exposure, and all data is attached with a microsecond-level timestamp generated by the FPGA. The parsing module splits the radar signal into discrete points of timestamp, azimuth, and signal strength, and the depth data is parsed into timestamp, world coordinates, and depth value. Through time sliding window matching, the radar point at time t=1689152000.124000 is associated with the depth point in the same frame, and finally a structured observation dataset containing 2 million records is constructed, each in the format of , the time and space synchronization error is less than 0.1ms.

[0133] Then, based on the above data set, the system separated multi-dimensional features: first, after applying 76-78GHz bandpass filtering to the electromagnetic signal, the thermal radiation features were extracted through dynamic threshold 40dB segmentation, and the AGV motor was identified as a continuously overheated area of ​​>60dB; Canny edge detection was used to calculate the local gradient of the depth data, and the gradient threshold was set to >8cm / m to separate the surface texture features, successfully capturing the concave deformation gradient peak of 22cm / m on the surface of the AGV cargo box; at the same time, based on 10 consecutive frames of position data, the motion vector was calculated by the improved Lucas-Kanade algorithm, and the lateral speed of AGV03 of 0.8m / s±0.05m / s and the rotational angular velocity of 15° / s were separated. Then a four-dimensional feature space was constructed, and the expanded data structure was as follows: For example, the eigenvector of AGV02 at t=1689152000.456 is Finally, through joint anomaly detection, when shelf occlusion caused the electromagnetic intensity of an AGV to drop sharply from 52dB to 28dB and the depth gradient to suddenly change from 12cm / m to 35cm / m, the system marked an occluded area of ​​1.2m×0.8m. The thermal radiation difference between this area and the normal area was greater than 24dB, and the texture gradient difference was greater than 18cm / m.

[0134] The system then quantified spatial features of the obstructed area, using a 0.25 m³ voxel grid to calculate target density. This revealed a density of only 12 points / m³ in the obstructed AGV04's tail, while the density in the normal area was greater than 35 points / m³. The system also calculated feature conflict intensity by integrating thermal radiation intensity deviations and texture gradient deviations, achieving a normalized weighted conflict value of 0.85. The detector was dynamically adjusted based on optimization rules: the millimeter-wave radar gain was increased from 25dB to 40dB, and the beamwidth was narrowed from 45° to 20° to focus on the obstructed area. The depth camera exposure time was increased from 20ms to 40ms, and HDR mode was enabled. After the parameters took effect, a rescan was performed, revealing significant improvements in the new data: the electromagnetic intensity at the AGV04's tail returned to 45dB, the depth gradient rebounded to 12cm / m, and the conflict intensity dropped to 0.32. The entire process was recorded as a structured feature enhancement log.

[0135] Based on the optimal parameters in the feature enhancement record, radar gain 40dB / beam 20°, camera exposure 40ms, the system calculates the optimal scanning path. Construct an electromagnetic intensity weight map, set the weight of the occlusion recovery area to 2.0, and the normal area to 1.0; superimpose the conflict heat map, and add an exploration factor of 1.5 in the high historical conflict area; use the improved A The algorithm plans the path, and the cost function integrates distance weight, feature weight and conflict factor. The total length of the generated path is 35.6m, including 128 scanning points, of which 82 points are densely distributed in the occlusion area, accounting for 64%. The gimbal is driven to perform multi-angle scanning at an angular velocity of 15° / s. In the original occlusion area of ​​AGV04, a spiral progressive trajectory radius is used from 0.5m to 0.2m, and each point stays for 100ms to collect multimodal data. Finally, a weighted scanning data set is generated, and the path weight coefficient of each data is between 0.5-2.0. For example, the key point P45 data is marked as ,The total amount of the data set is 150,000, covering the complete surface features of 6 AGVs.

[0136] Finally, the current scan dataset was integrated with 100,000 historical trajectory data to construct an LSTM-CNN hybrid prediction model. The CNN branch processes spatial features as input, using depth gradient maps and heatmaps. Three convolution layers extract local textures. The LSTM branch processes temporal features as input, using velocity vectors and angular velocity sequences. The model has a 128-unit memory state memory. Training is performed in parallel on a Spark distributed cluster, with one driver node coordinating tasks and three executor nodes handling model partitions. Node1 optimizes the CNN convolution kernel, Node2 fine-tunes the LSTM hidden layer, and Node3 integrates spatiotemporal features. After 200 iterations, the model's trajectory prediction error on the test set was reduced to 0.15 meters. After deployment, the model coordinates turntable decisions. When AGV05 enters a historically high-conflict zone, the model triggers parameter adjustments 300 milliseconds in advance and generates an avoidance pathpoint sequence [x=8.12, y=4.56]→[x=8.35, y=4.78]. Finally, in a continuous occlusion environment, the trajectory reconstruction completeness of the six AGVs reached 99.2%, with an average positioning error of 0.18m, which is 4.3 times more accurate than traditional methods.

[0137] Figure 3 The present invention provides a schematic diagram of a multi-target automatic tracking system based on an unmanned intelligent turntable. Figure 3 As shown, the system includes:

[0138] An acquisition module 31 is configured to collect electromagnetic echo signals and environmental depth data of a target area through a detector of an unmanned intelligent turntable, and synchronously calibrate the electromagnetic echo signals and the environmental depth data in a time dimension to form an observation data set;

[0139] A construction module 32 is configured to separate the target's thermal radiation characteristics, surface texture characteristics, and motion state characteristics based on the differences in physical properties of different sensors in the observation data set, construct a four-dimensional feature space containing time and space coordinates, and identify abnormal areas on the target surface as occlusion areas;

[0140] An adjustment module 33 is configured to adjust the signal parameters of the detector to enhance the target features of the occluded area according to the target distribution density in the four-dimensional feature space and the feature conflict strength of the occluded area, and feed the adjusted signal parameters back to the four-dimensional feature space to form a feature enhancement record;

[0141] a scanning module 34 for driving the sensor array to perform multi-angle scanning, calculating an optimal scanning path based on the signal parameter optimization values ​​in the feature enhancement record and the collected electromagnetic echo signals to generate a weighted scanning data set;

[0142] The prediction module 35 is used to establish a target trajectory prediction model based on the scanning data set and historical trajectory data, and use distributed computing nodes to parallel optimize the target trajectory prediction model to coordinate the parameter synchronization and decision optimization of the unmanned intelligent turntable to achieve reconstruction of continuous trajectories of multiple targets in an occluded environment.

[0143] Figure 3 The multi-target automatic tracking system based on unmanned intelligent turntable can perform Figure 1 The implementation principle and technical effects of the multi-target automatic tracking method based on an unmanned intelligent turntable described in the illustrated embodiment will not be elaborated on here. The specific manner in which each module and unit performs operations in the multi-target automatic tracking system based on an unmanned intelligent turntable in the above embodiment has been described in detail in the embodiments of the method and will not be elaborated on here.

[0144] In one possible design, Figure 3 The multi-target automatic tracking system based on an unmanned intelligent turntable of the embodiment shown can be implemented as a computing device, such as Figure 4 As shown, the computing device may include a storage component 41 and a processing component 42;

[0145] The storage component 41 stores one or more computer instructions, wherein the one or more computer instructions are called and executed by the processing component 42 .

[0146] The processing component 42 is used for the above Figure 1 The embodiment provides a multi-target automatic tracking method based on an unmanned intelligent turntable.

[0147] The processing component 42 may include one or more processors to execute computer instructions to complete all or part of the steps in the above method. Of course, the processing component may also be implemented as one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above method.

[0148] The storage component 41 is configured to store various types of data to support operations at the terminal. The storage component can be implemented by any type of volatile or non-volatile memory device, or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk, or optical disk.

[0149] Of course, a computing device may also include other components, such as input / output interfaces, display components, communication components, etc.

[0150] The input / output interface provides an interface between the processing component and the peripheral interface module, which can be an output device, an input device, etc.

[0151] The communication component is configured to facilitate, among other things, wired or wireless communications between the computing device and other devices.

[0152] Among them, the computing device can be a physical device or an elastic computing host provided by a cloud computing platform, etc. In this case, the computing device can refer to a cloud server, and the above-mentioned processing components, storage components, etc. can be basic server resources rented or purchased from the cloud computing platform.

[0153] The present application also provides a computer storage medium storing a computer program, wherein the computer program can achieve the above-mentioned Figure 1 The embodiment shown is a multi-target automatic tracking method based on an unmanned intelligent turntable.

[0154] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0155] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of the present embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0156] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0157] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A multi-target automatic tracking method based on an unmanned intelligent turntable, characterized in that: include: The electromagnetic echo signal and the environmental depth data of the target area are collected by the detector of the unmanned intelligent turntable, and the electromagnetic echo signal and the environmental depth data are synchronously calibrated in the time dimension to form an observation data set; Based on the physical property differences of different sensors in the observation data set, the target's thermal radiation characteristics, surface texture characteristics, and motion state characteristics are separated, and a four-dimensional feature space containing time and space coordinates is constructed. At the same time, abnormal areas on the target surface are identified as occlusion areas; Adjusting the signal parameters of the detector to enhance the target features of the occluded area according to the target distribution density in the four-dimensional feature space and the feature conflict strength of the occluded area, and feeding the adjusted signal parameters back to the four-dimensional feature space to form a feature enhancement record; calculating an optimal scanning path based on the optimized signal parameter values ​​in the feature enhancement record and the collected electromagnetic echo signal, and driving the sensor array to perform multi-angle scanning according to the optimal scanning path to generate a weighted scanning data set; Based on the scanning data set and historical trajectory data, a target trajectory prediction model is established, and the target trajectory prediction model is optimized in parallel using distributed computing nodes to coordinate the parameter synchronization and decision optimization of the unmanned intelligent turntable, thereby realizing the reconstruction of continuous trajectories of multiple targets in an occluded environment.

2. The method according to claim 1, characterized in that The target trajectory prediction model is established based on the scanning data set and historical trajectory data, and the target trajectory prediction model is optimized in parallel using distributed computing nodes to coordinate the parameter synchronization and decision optimization of the unmanned intelligent turntable to achieve reconstruction of continuous trajectories of multiple targets in an occluded environment, including: extracting position coordinates, time stamps, and path weight values ​​from the scan data set, forming a spatiotemporal sequence of scan data based on the position coordinates and time stamps, and converting the path weight values ​​into trajectory confidence weights; Aligning the position sequence in the historical trajectory data with the spatiotemporal sequence of the scan data, establishing a dynamic displacement equation based on the trajectory confidence weight, and completing the construction of the target trajectory prediction model; After deploying the target trajectory prediction model to distributed computing nodes, each node processes a subset of target data and optimizes the target trajectory prediction model by exchanging parameters of the dynamic displacement equation to minimize the spatial deviation between the predicted position and the scan data; Based on the optimization result, the unmanned intelligent turntable is coordinated to feed back the predicted position to the detector unit to adjust the scanning angle to focus on the predicted area, and at the same time, the trajectory confidence weight is compared with the set threshold to adjust the signal transmission power of the detector; The prediction process of the target trajectory prediction model and the parameter updating process of the dynamic displacement equation are iteratively performed in an occlusion environment, and the trajectory interruption points are connected by extrapolation calculation of the dynamic displacement equation to generate a continuous trajectory sequence.

3. The method according to claim 2, characterized in that The step of connecting the trajectory interruption points to generate a continuous trajectory sequence by extrapolating the dynamic displacement equation includes: When a trajectory interruption is detected, the last valid position coordinate is recorded as the starting interruption point, and the first recaptured position coordinate is recorded as the ending interruption point; Using the starting interruption point as the initial position, applying the dynamic displacement equation to predict the position at the next time point, inserting the predicted position as a virtual scanning point into the scanning data set and assigning a decaying trajectory confidence weight to form an extended data set, and updating the parameters of the dynamic displacement equation based on the extended data set; Based on the dynamic displacement equation after parameter update, iteratively predict from the starting interruption point to the intermediate time point in chronological order, use the position of the intermediate time point at this time as the forward prediction position, and record the forward trajectory sequence generated by the trajectory coordinates during the forward iterative prediction process; at the same time, iteratively predict from the ending interruption point in reverse chronological order to the same intermediate time point, use the position of the intermediate time point at this time as the reverse prediction position, and record the backward trajectory sequence generated by the trajectory coordinates during the reverse iterative prediction process; When the spatial deviation between the forward prediction position and the backward prediction position is less than a tolerance threshold, the forward trajectory sequence and the backward trajectory sequence are smoothly connected at the intermediate time point by cubic spline interpolation to generate a continuous trajectory segment.

4. The method according to claim 1, wherein The method of calculating an optimal scanning path based on the signal parameter optimization value in the feature enhancement record and combining the collected electromagnetic echo signal, and driving the sensor array to perform multi-angle scanning according to the optimal scanning path to generate a weighted scanning data set includes: generating a regional scanning priority based on the feature enhancement value in the feature enhancement record, wherein a region with a higher feature enhancement value corresponds to a higher scanning priority; identifying low signal strength areas and high priority areas whose scanning priorities are greater than a preset threshold value based on the intensity distribution of the electromagnetic echo signal, taking the highest priority area as the starting point of the path, forcibly covering all the high priority areas and low signal strength areas while minimizing the total steering angle, and generating an optimal scanning path; generating a steering angle sequence of the detector unit based on the optimal scanning path, driving the sensor array to scan according to the steering angle sequence, and obtaining the position coordinates, time stamp, electromagnetic echo signal value and environmental depth value of each scanning point; A path weight value equal to the feature enhancement value of the area in which each scanning point is located is assigned, and the information of all the scanning points is integrated to form a scanning data set including the position coordinates, time stamp, electromagnetic echo signal value, environmental depth value and path weight value.

5. The method according to claim 1, wherein The method separates the target's thermal radiation characteristics, surface texture characteristics, and motion state characteristics based on the physical property differences of different sensors in the observation data set, constructs a four-dimensional feature space containing time and space coordinates, and simultaneously identifies abnormal areas on the target surface as occlusion areas, including: Extracting electromagnetic echo signal values ​​and environmental depth values ​​as well as corresponding time stamps and position coordinates based on differences in physical characteristics of different sensors in the observation data set; Based on the change pattern of the electromagnetic echo signal value, the thermal radiation characteristics are separated by comparing the signal strength threshold; based on the local fluctuation pattern of the environmental depth value, the surface texture characteristics are separated by calculating the difference between adjacent depth values; based on the change trend of the position coordinates and time stamps, the motion state characteristics are separated by calculating the position offset of consecutive time points; Constructing a four-dimensional feature space by associating corresponding time tags and position coordinates according to the thermal radiation feature, the surface texture feature, and the motion state feature; In the four-dimensional feature space, abnormal areas on the target surface are identified by detecting mutations in the electromagnetic echo signal values ​​and the ambient depth values. When the surface texture features or the thermal radiation features of the abnormal area show an abnormal decrease in intensity value in a local area, it is marked as an occluded area.

6. The method according to claim 1, characterized in that The method further comprises: adjusting the signal parameters of the detector to enhance the target features of the occluded area according to the target distribution density in the four-dimensional feature space and the feature conflict strength of the occluded area, and feeding the adjusted signal parameters back to the four-dimensional feature space to form a feature enhancement record, including: Counting the number of feature points in a unit space area according to the degree of aggregation of position coordinates in the four-dimensional feature space to calculate the target distribution density; In the occlusion area, based on the intensity value deviations of the surface texture feature and the thermal radiation feature, respectively comparing the intensity values ​​of the surface texture feature and the thermal radiation feature in the normal area and the occlusion area to calculate corresponding feature conflict intensities; According to the target distribution density and the feature conflict strength, the signal parameters of the detector are adjusted by a signal parameter optimization rule to enhance the target features of the occluded area; The adjusted signal parameters are applied to the detector, the signal of the occluded area is recollected and the values ​​of the target distribution density and the feature conflict strength in the four-dimensional feature space are updated, and the adjustment process of the signal parameters, the target distribution density and the feature conflict strength is recorded to form a feature enhancement record.

7. The method according to claim 1, characterized in that The detector of the unmanned intelligent turntable collects electromagnetic echo signals and environmental depth data of the target area, and synchronously calibrates the electromagnetic echo signals and the environmental depth data in the time dimension to form an observation data set, including: A detector group including an electromagnetic sensor and a depth sensor is set on the unmanned intelligent turntable. The electromagnetic sensor transmits electromagnetic waves at fixed time intervals and receives reflected signals to generate electromagnetic echo signals. The depth sensor obtains depth information of the target area as environmental depth data based on optical principles. When the electromagnetic sensor and the depth sensor generate data, a time stamp uniformly managed by the internal clock of the unmanned intelligent turntable is automatically added, and the electromagnetic echo signal and the environmental depth data are decomposed into discrete data points including a time stamp, a position coordinate, and a corresponding signal strength value; Based on the time stamp, the electromagnetic echo signal is paired one-to-one with the discrete data points of the environmental depth data to form a data unit, and the data unit is integrated into a structured data set including the time stamp, location coordinates, and corresponding signal strength values ​​as an observation data set.

8. A multi-target automatic tracking system based on an unmanned intelligent turntable, characterized in that: include: An acquisition module is used to collect electromagnetic echo signals and environmental depth data of the target area through a detector of an unmanned intelligent turntable, and synchronously calibrate the electromagnetic echo signals and the environmental depth data in the time dimension to form an observation data set; A construction module is used to separate the thermal radiation characteristics, surface texture characteristics, and motion state characteristics of the target based on the physical property differences of different sensors in the observation data set, construct a four-dimensional feature space containing time and space coordinates, and identify abnormal areas on the target surface as occlusion areas; an adjustment module, configured to adjust the signal parameters of the detector to enhance the target features of the occluded area according to the target distribution density in the four-dimensional feature space and the feature conflict strength of the occluded area, and feed the adjusted signal parameters back to the four-dimensional feature space to form a feature enhancement record; a scanning module, configured to drive the sensor array to perform multi-angle scanning, calculate an optimal scanning path based on the signal parameter optimization values ​​in the feature enhancement record and the collected electromagnetic echo signal, so as to generate a weighted scanning data set; The prediction module is used to establish a target trajectory prediction model based on the scanning data set and historical trajectory data, and to use distributed computing nodes to parallel optimize the target trajectory prediction model to coordinate the parameter synchronization and decision optimization of the unmanned intelligent turntable, thereby realizing the reconstruction of continuous trajectories of multiple targets in an occluded environment.

9. A computing device, characterized in that It includes a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement a multi-target automatic tracking method based on an unmanned intelligent turntable as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that A computer program is stored, and when the computer program is executed by a computer, the multi-target automatic tracking method based on an unmanned intelligent turntable as described in any one of claims 1 to 7 is implemented.

Images