Unique identification method and system for global target based on multi-source sensing data fusion

Abstract

The invention discloses a global target unique identification method and system based on multi-source sensing data fusion, and the method comprises the steps: collecting the state information data, namely a target observation state value, of a target through a multi-source sensor device, and removing abnormal data; calculating a detection overlapping region between the multi-source sensor devices; carrying out filtering processing on target state information data detected by each sensor device by using a Kalman filtering algorithm based on the detection overlapping region; matching the target prediction state value obtained by filtering with the target observation state value by using a genetic algorithm; according to a matching result, parameters in the Kalman filtering algorithm are corrected, and model parameters of the Kalman filtering algorithm are updated in multiple iterations; and based on the updated Kalman filtering algorithm model, realizing prediction of a target state. According to the method, sensor information is fused, the problems that a single sensor target is inaccurate, static target detection is poor, a large target is mistakenly segmented, video detection is prone to being interfered by the environment and the like are solved, and high precision, high robustness and track continuity of target detection in the holographic traffic scene are achieved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent transportation technology, and in particular, it is a method for unique identification of targets across the entire domain based on multi-source sensor data fusion. Background Technology

[0002] With the rapid development of Intelligent Transportation Systems (ITS), traffic perception technology has gradually become a core component of smart city construction. Among these, radar and video (radar-visual) fusion detection equipment, due to its all-weather, high-precision target detection capabilities, is widely used in urban roads, intersections, highways, and other scenarios to achieve real-time monitoring and management of traffic targets such as vehicles and pedestrians. However, in actual deployment, due to the spatial distribution and overlapping detection ranges of multiple sensor devices, the same target may be repeatedly perceived from different device perspectives. This leads to problems such as target ghosting (i.e., the same target being rendered multiple times) or target splitting (i.e., a single target being misidentified as multiple fragments) during holographic scene reconstruction in the central system. This not only reduces the accuracy of traffic data but may also interfere with subsequent trajectory prediction, behavior analysis, and traffic control decisions. For example, at intersections, multiple devices may be installed in different directions, and their detection areas inevitably overlap. When a target (such as a vehicle) enters an overlapping area, each device can independently detect and report the target information. However, due to the spatiotemporal calibration errors between devices, differences in detection accuracy, and dynamic changes in the target's motion state, the central system has difficulty directly determining whether these detection results belong to the same physical entity.

[0003] 1) The same target may be assigned different temporary identifiers (IDs) in the detection data of different sensors, causing the system to mistakenly regard it as multiple independent targets, thus generating duplicate virtual objects in the holographic scene. Furthermore, the calibration between different sensors produces detection errors, which further exacerbates the complexity of target data processing.

[0004] 2) For high-frequency target data of full-domain device detection, under the traditional architecture, (1) the performance of the single-node data processing mode is insufficient, resulting in poor real-time performance during the target fusion process, invalid targets participate in fusion matching, and it is difficult to make efficient use of resources in effective computing resources; (2) the system access capability is limited, and the system cannot be horizontally expanded when facing more device detection data access. Summary of the Invention

[0005] The purpose of this invention is to address the problems existing in the prior art by providing a global target uniqueness identification method based on multi-source sensor data fusion. To address issues such as target overlap, inter-sensor errors, and low computational performance in multi-target detection, this invention proposes a method that, during the device fusion process, cleans the detection data and splits it into tasks across various computing nodes, thereby locking the target to be fused and processing it in parallel on different computing nodes. It iterates through the Euclidean distance generated by target matching to form the target matching cost, improves the genetic algorithm by introducing a multi-mutation operator and a time-termination iteration method, comprehensively considering uncertainties such as target mutation in sensor target detection, inter-sensor errors, and target motion, and calculates the sample matching fitness to achieve globally optimal target matching. Kalman filtering is then used to achieve target tracking. This invention maximizes the utilization of computational resources while ensuring the accuracy and robustness of actual target tracking, effectively solving the target fusion problem in holographic scenes.

[0006] The technical solution to achieve the purpose of this invention is: a method for unique global target identification based on multi-source sensor data fusion, wherein the method realizes periodic updates of overlapping region calculation and node calculation tasks, including the following steps:

[0007] Step 1, Data Acquisition and Processing: Collect target status information data, i.e., target observation status values, through multi-source sensor devices, and remove abnormal data; the status information data includes the target's latitude and longitude location information, velocity, and acceleration;

[0008] Step 2: Calculate the detection overlap area between multi-source sensor devices;

[0009] Step 3: Based on the detected overlapping area, use the Kalman filter algorithm to filter the target state information data detected by each sensor device;

[0010] Step 4: Use a genetic algorithm to match the target predicted state value obtained from the filtering process with the target observed state value;

[0011] Step 5: Based on the matching results, the parameters in the Kalman filter algorithm are corrected. Through multiple iterations, the Kalman filter algorithm model parameters are updated.

[0012] Step 6: Based on the updated Kalman filter algorithm model, predict the target state.

[0013] Furthermore, step 1 specifically includes:

[0014] Step 1-1: Build an application system including computing nodes and distribution nodes, and deploy sensor devices in a traffic scenario, while establishing a network cascading relationship between the sensor devices; if there are overlapping detection areas between sensor devices, these sensor devices must be in the same network (to eliminate the influence of the network on different sensor devices); the computing nodes deploy various computing algorithm models for data processing; the distribution nodes deploy data routing rules for initializing data and various computing algorithm models, and also for storing and sharing external inputs;

[0015] Steps 1-2: Collect target status information data through sensor devices and transmit it to the distribution node;

[0016] Steps 1-3: The distribution node performs dirty data processing on the received data to remove abnormal data.

[0017] Furthermore, step 2, calculating the detection overlap region between the multi-source sensor devices, specifically includes:

[0018] Step 2-1: The distribution node constructs a cascade relationship of sensor devices across the entire domain, forming a device group relationship; each device group has a structured relationship from upstream to downstream of the detection area.

[0019] Step 2-2: The distribution node completes the construction of the device group detection data task, establishes the corresponding relationship between each device group and each computing node, and forms a distribution strategy (automatically generated or manually configured). This distribution strategy must ensure that there is no overlap between overlapping areas of different device groups. At the same time, the distribution node initializes the parameters of the Kalman filter algorithm and the genetic algorithm.

[0020] Steps 2-3: For each device group, perform frame merging processing on the data collected by all sensor devices within it.

[0021] Steps 2-4 involve calculating the overlapping area for each device group, specifically including:

[0022] The convex hull algorithm is used to traverse the target position information collected by all sensor devices in the device group, and the overlapping region formed by the device group is calculated (the overlapping region is stored in the form of a set of feature points, and the point sets are connected in order to form the final effective overlapping polygon region).

[0023] Step 2-5: Determine whether there is overlap between the overlapping areas corresponding to each equipment group. If so, return to step 2-2; otherwise, end the entire process.

[0024] Furthermore, steps 2-4 specifically include:

[0025] The target location information collected by all sensor devices in the device group is traversed using the convex hull algorithm, and the overlapping area formed by the device group is calculated.

[0026] Furthermore, the overlapping region is stored in the form of a set of feature points, and the point sets are connected in sequence to form the final effective overlapping polygonal region.

[0027] Further, in step 3, based on the detected overlapping region, the target state information data detected by each sensor device is filtered using a Kalman filter algorithm, specifically including:

[0028] Step 3-1: Determine whether the target is within a certain detection overlap area. If so, proceed to step 3-2.

[0029] Step 3-2: Determine if the current input is the first input. If it is the first input, proceed to step 3-3; otherwise, proceed to step 3-4.

[0030] Step 3-3: Construct the Kalman filter model and initialize the effective state vector and Kalman filter model parameters;

[0031] Among them, the effective state vector for:

[0032]

[0033] In the formula, For longitude, For latitude, for Directional speed, for Directional speed, for acceleration in direction for Acceleration in the direction of; where the target is set to due north. The direction is set to due east. direction;

[0034] Return to step 3-1;

[0035] Steps 3-4 involve using Kalman filtering to process the target's state information data, thereby enabling the prediction of the next frame's data based on the previous frame's data.

[0036] Furthermore, step 4, which involves matching the target predicted state value obtained through filtering with the target observed state value using a genetic algorithm, specifically includes:

[0037] Step 4-1: Initial population size, generating the initial population;

[0038] Step 4-2: Determine whether the termination iteration condition is met. If yes, proceed to step 4-5; otherwise, execute step 4-3.

[0039] Step 4-3, calculate the fitness function:

[0040]

[0041]

[0042] In the formula, It is the fitness of the sample to match the target. The cost incurred when detecting a match between the target and the tracker's state prediction is given, where M is the number of elements in the detector's input set of observations, and N is the number of elements in the tracker's output set generated by the filtering algorithm. Let be the Euclidean distance between the observed and predicted values, and let i and j be the tracker index and the observed target index, respectively. It is the Euclidean distance between the observed value and the tracker's predicted value. Let j be the observation vector of the j-th target. Let be the predicted state vector of the i-th tracker, i.e., the predicted value generated by the tracker in the Kalman filter; parameters This represents the tolerance for missed detections in the sample. When the Euclidean distance exceeds this range, it is considered an unreasonable match. The samples formed by this matching scheme will be mutated in subsequent iterations based on the mutation operator.

[0043] Then proceed to step 4-4;

[0044] Step 4-4: Select, crossover, and mutate the population samples to generate a new population, then proceed to step 4-2.

[0045] Steps 4-5: End the process and output the globally optimal individual as the optimal matching result.

[0046] Furthermore, in step 4-4, a missed detection operator, a target switching operator, and a target mutation operator are introduced to mutate the sample. The confidence level and confidence interval are defined as follows: As shown:

[0047]

[0048] In the formula, For the missed detection operator, for Confidence level, Switch operators for the target. for Confidence level, For the target mutation operator, for The confidence level.

[0049] Furthermore, in step 5, the parameters in the Kalman filter algorithm are corrected based on the matching results, specifically including:

[0050] Step 5-1: Determine whether the cumulative number of predictions recorded by the tracker has reached the preset cleanup threshold. If yes, proceed to step 5-5; otherwise, proceed to step 5-2.

[0051] Step 5-2: Calculate the Euclidean distance between the predicted value and the observed value, and determine whether the preset threshold condition is met. If so, proceed to step 5-3; otherwise, proceed to step 5-4.

[0052] Step 5-3: Correct the filter parameters, then proceed to step 5-6;

[0053] Step 5-4: Record the observation information without updating the filter parameters, and then proceed to step 5-6.

[0054] Step 5-5, clean the tracker;

[0055] Steps 5-6: Update and output the tracker information.

[0056] On the other hand, a global target unique identification system is provided, the system comprising:

[0057] The first module is used to: collect target status information data, i.e., target observation status values, through multi-source sensor devices, and remove abnormal data; the status information data includes the target's latitude and longitude location information, velocity, and acceleration;

[0058] The second module is used to calculate the detection overlap area between multi-source sensor devices;

[0059] The third module is used to: based on the detection overlap area, use the Kalman filter algorithm to filter the target state information data detected by each sensor device;

[0060] The fourth module is used to match the target predicted state value obtained by filtering with the target observed state value using a genetic algorithm.

[0061] The fifth module is used to: correct the parameters in the Kalman filter algorithm based on the matching results, and update the Kalman filter algorithm model parameters in multiple iterations;

[0062] The sixth module is used to implement the prediction of the target state based on the updated Kalman filter algorithm model.

[0063] Compared with the prior art, the present invention has the following significant advantages:

[0064] (1) The present invention selects sensor detection information as the input information of the detector, each node loads the algorithm model, and the data is distributed to multiple computing nodes for parallel processing, which improves the data processing capability during the fusion process and meets the business requirements of supporting the fusion of multiple devices at the same time.

[0065] (2) This invention utilizes the convex hull algorithm to preheat the associated device to detect overlapping areas, sets the target location area detection before matching, filters targets outside the area into the fusion algorithm, reduces the target dataset to be matched, effectively improves the algorithm fusion efficiency, and achieves full target traversal matching.

[0066] (3) This invention takes into account the actual situation of sensor engineering applications. By dividing multiple independent overlapping regions and configuring a dedicated genetic algorithm matching node in each region, the mutation operator probability, iteration number and termination time can be adaptively adjusted based on the difference in target passage time, so as to achieve parallel matching of multiple regions without interference, thereby improving matching accuracy and real-time performance.

[0067] (4) The present invention sets multiple state threshold constraints on the target in the tracker, realizes accurate positioning and timely deletion of the tracked target, greatly reduces the frequent ID switching and target loss phenomenon in target tracking. Compared with the prior art, the method is highly flexible and switches the tracking mode according to the tracking state and environmental changes to realize accurate tracking of the detected target, enhances the comprehensiveness of multi-target recognition and tracking, and provides a new solution for holographic scene visualization.

[0068] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0069] Figure 1 This is a core flowchart of a global target unique identification method based on multi-source sensor data fusion in one embodiment.

[0070] Figure 2 This is a network architecture diagram of a device deployment method for a global target unique identification method based on multi-source sensor data fusion in one embodiment.

[0071] Figure 3 This is a schematic diagram of the road equipment detection area in one embodiment of a global target unique identification method based on multi-source sensor data fusion.

[0072] Figure 4 This is a schematic diagram of a typical intersection device detection method based on multi-source sensor data fusion for unique target identification across the entire domain, as shown in one embodiment.

[0073] Figure 5 This is a schematic diagram of a service node for a global target unique identification method based on multi-source sensor data fusion in one embodiment.

[0074] Figure 6This is a flowchart illustrating the data relationship construction process of a global target unique identification method based on multi-source sensor data fusion in one embodiment.

[0075] Figure 7 This is a flowchart of the sensor filtering process in a global target unique identification method based on multi-source sensor data fusion, as shown in one embodiment.

[0076] Figure 8 This is a flowchart of the genetic algorithm fusion process for a global target unique identification method based on multi-source sensor data fusion in one embodiment.

[0077] Figure 9 This is a fusion result image of a road scene based on a global target unique identification method using multi-source sensor data fusion in one embodiment.

[0078] Figure 10 This is a fusion result diagram of a global target unique identification method based on multi-source sensor data fusion in an intersection scenario, as shown in one embodiment. Detailed Implementation

[0079] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0080] It should be noted that if the embodiments of the present invention involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of the components in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.

[0081] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0082] In one embodiment, combined Figure 1 This paper provides a method for unique global target identification based on multi-source sensor data fusion. The method realizes the periodic updating of overlapping region calculation and node calculation tasks, and includes the following steps:

[0083] Step 1, Data Acquisition and Processing: Collect target status information data, i.e., target observation status values, through multi-source sensor devices, and remove abnormal data; the status information data includes the target's latitude and longitude location information, velocity, and acceleration;

[0084] Step 2: Calculate the detection overlap area between multi-source sensor devices;

[0085] Step 3: Based on the detected overlapping area, use the Kalman filter algorithm to filter the target state information data detected by each sensor device;

[0086] Step 4: Use a genetic algorithm to match the target predicted state value obtained from the filtering process with the target observed state value;

[0087] Step 5: Based on the matching results, the parameters in the Kalman filter algorithm are corrected. Through multiple iterations, the Kalman filter algorithm model parameters are updated.

[0088] Step 6: Based on the updated Kalman filter algorithm model, predict the target state.

[0089] Furthermore, in one embodiment, step 1 specifically includes:

[0090] Step 1-1: Build an application system including computing nodes and distribution nodes, and deploy sensor devices in a traffic scenario, while establishing a network cascading relationship between the sensor devices; if there are overlapping detection areas between sensor devices, these sensor devices must be in the same network (the purpose is to eliminate the influence of the network on different sensor devices); the computing nodes deploy various computing algorithm models for data processing; the distribution nodes deploy data routing rules for initializing data and various computing algorithm models, and also for storing and sharing external inputs;

[0091] Steps 1-2: Collect target status information data through sensor devices and transmit it to the distribution node;

[0092] In steps 1-3, the distribution node performs dirty data processing on the received data to remove abnormal data (to prevent abnormal data caused by sensor point drift or sensor jitter from entering the subsequent algorithm model). For example, it cleans target type data, removes noise, and filters out all targets except vehicles, including pedestrians and bicycles, while retaining target types such as buses, trucks, vans, and cars, and extracts targets with intense and obvious movement for prediction and fusion.

[0093] Furthermore, in one embodiment, step 2, calculating the detection overlap region between the multi-source sensor devices, specifically includes:

[0094] Step 2-1: The distribution node constructs a cascade relationship of sensor devices across the entire domain, forming a device group relationship; each device group has a structured relationship from upstream to downstream of the detection area.

[0095] Step 2-2: The distribution node completes the construction of the device group detection data task, establishes the corresponding relationship between each device group and each computing node, and forms a distribution strategy (which can be automatically generated or manually configured without special restrictions). This distribution strategy must ensure that there is no overlap between overlapping areas of different device groups. At the same time, the distribution node initializes the parameters of the Kalman filter algorithm and the genetic algorithm.

[0096] Steps 2-3: For each device group, perform frame merging processing on the data collected by all sensor devices within it.

[0097] Steps 2-4 involve calculating the overlapping area for each device group, specifically including:

[0098] The convex hull algorithm is used to traverse the target position information collected by all sensor devices in the device group, and the overlapping region formed by the device group is calculated (the overlapping region is stored in the form of a set of feature points, and the point sets are connected in order to form the final effective overlapping polygon region).

[0099] Step 2-5: Determine whether there is overlap between the overlapping areas corresponding to each equipment group. If so, return to step 2-2; otherwise, end the entire process.

[0100] Preferably, in some embodiments, steps 2-4 specifically include:

[0101] The target location information collected by all sensor devices in the device group is traversed using the convex hull algorithm, and the overlapping area formed by the device group is calculated.

[0102] Preferably, in some embodiments, the overlapping region is stored in the form of a set of feature points, and the point sets are connected sequentially to form a final effective overlapping polygonal region.

[0103] Furthermore, in one embodiment, in step 3, based on the detected overlapping region, the target state information data detected by each sensor device is filtered using a Kalman filter algorithm, specifically including:

[0104] Step 3-1: Determine whether the target is within a certain detection overlap area. If so, proceed to step 3-2.

[0105] Step 3-2: Determine if the current input is the first input. If it is the first input, proceed to step 3-3; otherwise, proceed to step 3-4.

[0106] Step 3-3: Construct the Kalman filter model and initialize the effective state vector and Kalman filter model parameters;

[0107] Among them, the effective state vector for:

[0108]

[0109] In the formula, For longitude, For latitude, for Directional speed, for Directional speed, for acceleration in direction for Acceleration in the direction of; where the target is set to due north. The direction is set to due east. direction;

[0110] Return to step 3-1;

[0111] Steps 3-4 involve using Kalman filtering to process the target's state information data, thereby enabling the prediction of the next frame's data based on the previous frame's data.

[0112] Furthermore, in one embodiment, step 4, which involves matching the target predicted state value obtained through filtering with the target observed state value using a genetic algorithm, specifically includes:

[0113] Step 4-1: Initial population size, generating the initial population;

[0114] Step 4-2: Determine whether the termination iteration condition is met. If yes, proceed to step 4-5; otherwise, execute step 4-3.

[0115] Step 4-3, calculate the fitness function:

[0116]

[0117]

[0118] In the formula, It is the fitness of the sample to match the target. The cost incurred when detecting a match between the target and the tracker's state prediction is given, where M is the number of elements in the detector's input set of observations, and N is the number of elements in the tracker's output set generated by the filtering algorithm. Let be the Euclidean distance between the observed and predicted values, and let i and j be the tracker index and the observed target index, respectively. It is the Euclidean distance between the observed value and the tracker's predicted value. Let j be the observation vector of the j-th target. Let be the predicted state vector of the i-th tracker, i.e., the predicted value generated by the tracker in the Kalman filter; parameters This represents the tolerance for missed detections in the sample. When the Euclidean distance exceeds this range, it is considered an unreasonable match. The samples formed by this matching scheme will be mutated in subsequent iterations based on the mutation operator.

[0119] Then proceed to step 4-4;

[0120] Here is a detailed explanation:

[0121] In the above formula, the cost of a sample is calculated. The calculation process involves multiple parameters, and the model input parameter is the number of valid observed targets. The set is The filter model is input A tracker, a collection When initializing samples, a heuristic initialization method based on detection-observation matching degree is adopted to generate multiple samples. The cost of the samples is calculated, which is converted into the calculation of the Euclidean distance between the target and the sample scheme. The cost matrix between the tracker and the detected target is calculated. Finally, the optimal sample is obtained in the genetic algorithm iteration. The cost matrix is ​​as follows:

[0122]

[0123] In the cost matrix The cost of matching the tracker with the observed target, The predicted state is the input through filtering. The target observation state transmitted by the sensor. The Euclidean distance between the tracker and the observed target;

[0124] in, This involves calculating the Euclidean distance after normalizing the target attributes, and calculating the elements of the 6-dimensional attribute cost matrix. Due to the small error in latitude and longitude data, the cost calculation method for each pairwise target is as follows:

[0125]

[0126] In the formula, This represents the tracker's predicted state of the target. The observation status of the observed target. It is the Euclidean distance to the target, where This represents the difference in latitude, in meters. This represents the difference in longitude, in meters. , The tracker predicts the longitude and latitude of the target. , For the tracker in direction and Predicted velocity values ​​in the direction of travel. , For the tracker in direction and Acceleration prediction values ​​in the direction, , For the target longitude and latitude observation values, , For the tracker in direction and Observations of velocity in the direction, , For the tracker in direction and Acceleration observations in the direction.

[0127] in, , It cannot be obtained directly and needs to be converted. After conversion, the attributes are normalized to calculate the Euclidean distance. The conversion formula is as follows:

[0128]

[0129] In the formula As a cosine function, after the above relationship transformation, the latitude and longitude changes are converted to... direction and The offset in direction, with the unit of offset being meters.

[0130] When normalizing 6-dimensional attributes, the attribute range is set for normalization. The normalization strategy formula is shown below:

[0131]

[0132] in, It is the actual value of the corresponding dimension attribute. It is the minimum value of this dimension attribute in the associated region. It is the maximum value in this dimension attribute, and will be the current The value is projected onto this interval to obtain the normalized result. .

[0133] Step 4-4: Select, crossover, and mutate the population samples to generate a new population, then proceed to step 4-2.

[0134] During the above process, if the tracker fails to match the observed target, a status marker is set, and the number of tracker matching failures is counted and accumulated.

[0135] Steps 4-5: End the process and output the globally optimal individual as the optimal matching result.

[0136] Preferably, in some embodiments, in step 4-4, a missed detection operator, a target switching operator, and a target mutation operator are introduced to mutate the sample, and the confidence level and confidence interval are defined as follows. As shown:

[0137]

[0138] In the formula, For the missed detection operator, for Confidence level, Switch operators for the target. for Confidence level, For the target mutation operator, for The confidence level.

[0139] Here, the algorithm introduces a missed detection mutation operator, which effectively solves the problem of tracking trajectory interruption caused by instantaneous missed detection and improves the algorithm's tolerance to missed detection; it also introduces a fine-tuning mutation operator, which effectively solves the problem of matching results oscillating near the optimal solution, and improves the accuracy and stability of the matching scheme through local fine search; and it introduces a target switching operator, which effectively solves the matching confusion problem, can actively explore and correct possible erroneous associations, and helps the algorithm escape local optima.

[0140] Furthermore, in one embodiment, step 5 involves correcting the parameters in the Kalman filter algorithm based on the matching results, specifically including:

[0141] Step 5-1: Determine whether the cumulative number of predictions recorded by the tracker has reached the preset cleanup threshold. If yes, proceed to step 5-5; otherwise, proceed to step 5-2.

[0142] Step 5-2: Calculate the Euclidean distance between the predicted value and the observed value, and determine whether the preset threshold condition is met. If so, proceed to step 5-3; otherwise, proceed to step 5-4.

[0143] Step 5-3: Correct the filter parameters, then proceed to step 5-6;

[0144] Step 5-4: Record the observation information without updating the filter parameters, and then proceed to step 5-6.

[0145] Step 5-5, clean the tracker;

[0146] Steps 5-6: Update and output the tracker information.

[0147] In one embodiment, a global target unique identification system is provided, the system comprising:

[0148] The first module is used to: collect target status information data, i.e., target observation status values, through multi-source sensor devices, and remove abnormal data; the status information data includes the target's latitude and longitude location information, velocity, and acceleration;

[0149] The second module is used to calculate the detection overlap area between multi-source sensor devices;

[0150] The third module is used to: based on the detection overlap area, use the Kalman filter algorithm to filter the target state information data detected by each sensor device;

[0151] The fourth module is used to match the target predicted state value obtained by filtering with the target observed state value using a genetic algorithm.

[0152] The fifth module is used to: correct the parameters in the Kalman filter algorithm based on the matching results, and update the Kalman filter algorithm model parameters in multiple iterations;

[0153] The sixth module is used to implement the prediction of the target state based on the updated Kalman filter algorithm model.

[0154] Specific limitations regarding the global target uniqueness identification system based on multi-source sensor data fusion can be found in the limitations of the global target uniqueness identification method based on multi-source sensor data fusion described above, and will not be repeated here. Each module in the aforementioned global target uniqueness identification system based on multi-source sensor data fusion can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0155] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements:

[0156] Step 1, Data Acquisition and Processing: Collect target status information data, i.e., target observation status values, through multi-source sensor devices, and remove abnormal data; the status information data includes the target's latitude and longitude location information, velocity, and acceleration;

[0157] Step 2: Calculate the detection overlap area between multi-source sensor devices;

[0158] Step 3: Based on the detected overlapping area, use the Kalman filter algorithm to filter the target state information data detected by each sensor device;

[0159] Step 4: Use a genetic algorithm to match the target predicted state value obtained from the filtering process with the target observed state value;

[0160] Step 5: Based on the matching results, the parameters in the Kalman filter algorithm are corrected. Through multiple iterations, the Kalman filter algorithm model parameters are updated.

[0161] Step 6: Based on the updated Kalman filter algorithm model, predict the target state.

[0162] For specific limitations on each step, please refer to the limitations on the global target unique identification method based on multi-source sensor data fusion mentioned above, which will not be repeated here.

[0163] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program being implemented when executed by a processor:

[0164] Step 1, Data Acquisition and Processing: Collect target status information data, i.e., target observation status values, through multi-source sensor devices, and remove abnormal data; the status information data includes the target's latitude and longitude location information, velocity, and acceleration;

[0165] Step 2: Calculate the detection overlap area between multi-source sensor devices;

[0166] Step 3: Based on the detected overlapping area, use the Kalman filter algorithm to filter the target state information data detected by each sensor device;

[0167] Step 4: Use a genetic algorithm to match the target predicted state value obtained from the filtering process with the target observed state value;

[0168] Step 5: Based on the matching results, the parameters in the Kalman filter algorithm are corrected. Through multiple iterations, the Kalman filter algorithm model parameters are updated.

[0169] Step 6: Based on the updated Kalman filter algorithm model, predict the target state.

[0170] For specific limitations on each step, please refer to the limitations on the global target unique identification method based on multi-source sensor data fusion mentioned above, which will not be repeated here.

[0171] As a specific example, the invention will be further verified and illustrated in one embodiment.

[0172] In this embodiment, a 100GHz millimeter-wave radar and an 8-megapixel monocular camera are used as sensors to simultaneously acquire road target position coordinates relative to the ground, target speed data, acceleration data, target identification number, and relative coordinate information.

[0173] The global target unique identification method based on multi-source sensor data fusion in this embodiment includes the following processes: data acquisition and reception, data relationship construction, filtering, target matching, filtering parameter correction, and target prediction.

[0174] Step 1, Data Acquisition and Processing: Target state information data, i.e., target observation state values, is acquired through multi-source sensor equipment, and abnormal data is removed. The state information data includes the target's latitude and longitude location, velocity, and acceleration. Specifically, this includes:

[0175] Step 1-1: Build an application system including computing nodes and distribution nodes, and deploy sensor devices in a traffic scenario, while establishing a network cascading relationship between the sensor devices; if there are overlapping detection areas between sensor devices, these sensor devices must be in the same network (to eliminate the influence of the network on different sensor devices); the computing nodes deploy various computing algorithm models for data processing; the distribution nodes deploy data routing rules for initializing data and various computing algorithm models, and also for storing and sharing external inputs;

[0176] Steps 1-2: Collect target status information data through sensor devices and transmit it to the distribution node;

[0177] Steps 1-3: The distribution node performs dirty data processing on the received data to remove abnormal data.

[0178] Combination Figure 2 The entire domain is equipped with front-end devices, which are millimeter-wave radar detectors (radar-visual integrated machines). The devices are connected to switches to upload data to the center. The central switch communicates with each processing node. All devices in the domain upload all detection information and aggregate it to the central computer room. The computer room service distributes the data to each computing node.

[0179] Combination Figure 3 The entire scenario includes relay fusion scenarios such as ordinary roads, road sections, and elevated roads. For devices in the scenario, angles are selected and configured to observe target vehicles and traffic conditions from the best perspective. Devices detect targets in the same direction, and the device pitch and orientation are adjusted to ensure that the devices form an effective overlapping area. The effective movement length range of the target within the overlapping area is defined. Unify the coordinate system between devices to ensure that the accuracy error of target detection between devices is within a threshold range. This is to control the target tracking prediction in the model within a reasonable range, thereby enabling the target to be detected simultaneously by multiple devices in the detection area in traffic scenarios, thus forming a multi-device relay tracking.

[0180] Combination Figure 4 The full-domain scenario includes ordinary intersection scenarios. Equipment is deployed to intersection poles to form an effective detection area. In this embodiment, the equipment is deployed at the intersection entrance, with the camera facing the rear of the vehicles to maximize coverage of the intersection area, creating overlapping intersection cross-sections. A unified coordinate system is used between devices to ensure that the accuracy error of target detection between devices is within a threshold range. .

[0181] Combination Figure 5 The application system provides a message bus. The device needs to upload the detection target information to the message bus at the same time. The message reporting delay is less than 10ms. The distribution service obtains the bus message and distributes it to each computing node. The data distribution service divides the deployment relationship of the devices in the whole domain: (1) By default, the same intersection and the same road segment are grouped together. The whole domain adopts the default grouping method and is divided into the following groups: Grouping, generating group codes, performing HASH or modulo calculations on the group codes, distributing the device detection data of each group to each computing node, and each computing node completing the calculation task; (2) Manual configuration method, where system personnel configure the grouping relationship, dividing the entire domain devices into The system generates group codes, and users manually choose between hashing or modulo calculation. The detection data from each group is distributed to each computing node, which then completes its computational task. This step establishes the cascading relationship between the devices, ensuring stable transmission of sensor data. For sensor-detected target data, outliers and non-motorized vehicle target data must be removed. The target types detected by the sensors include: car, bus, bike, truck, person, and motor.

[0182] Step 2: Calculate the detection overlap area between multi-source sensor devices. This step constructs relationships between sensor data, device installation associations, and computing node information, using device groups as the dimension, generates a task allocation strategy, and distributes data to each computing node according to the strategy. Figure 6 The build process includes:

[0183] (1) The distribution node completes the construction of the cascade relationship of the devices in the whole domain and forms the device group relationship. The device group is a structured relationship from the upstream to the downstream of the detection area. The distribution node completes the construction of the detection data task of the device group and forms the distribution strategy. The distribution node saves the data according to the device group construction information, either by default strategy or manual configuration strategy. At the same time, the sensor data of the same device group is framed.

[0184] (2) Complete the construction of equipment group relationship, the distribution node completes the calculation of overlapping area of ​​equipment group, completes the calculation of overlapping area between equipment groups, and stores the data in the data sharing area. The overlapping area calculation method adopts the convex hull algorithm, and the calculation result is the set of regional feature points.

[0185] (3) Initialize the mutation operator confidence interval, confidence level, termination condition, and filter parameter threshold for equipment groups with overlapping regions. This initialization can be performed in parallel. Specifically:

[0186] (3.1) Complete the data initialization for the mutation operator;

[0187] (3.2) Initialize the termination condition, the constraints of which include: (1) For each batch of synchronous frame data passed in at a 300ms cycle, the algorithm dynamically adjusts the single iteration time threshold of the population sample between 500 and 1000ms (500-700ms when the number of targets in the overlapping area is ≤5, and 800-1000ms when it is >5) to ensure that the matching result is output in a timely manner to support the real-time performance of the holographic intersection; (2) The number of population iterations should be less than 50 times to avoid the iteration not ending when the target has crossed the overlapping area; (3) The iteration time should be synchronously less than 20% of the time taken for the target to cross the overlapping area to further ensure the consistency between the matching result and the actual movement state of the target. In the overlapping area, according to the actual scenario of the current patent, select, cross over, and mutate samples, generate a new population, and continue to calculate the fitness function. After triggering the termination condition, output the optimal sample. If not triggered, continue to iterate to generate new samples.

[0188] (3.3) Initialize the filter threshold parameter, and set the threshold to 0.2;

[0189] (3.4) Store the initialized data and merge and distribute the data collected by the devices in the device group.

[0190] Step 3: Based on the detected overlapping region, the Kalman filter algorithm is used to filter the target state information data detected by each sensor device, combined with... Figure 7 Specifically, it includes:

[0191] Step 3-1: Determine whether the target is within a certain detection overlap area. If so, proceed to step 3-2.

[0192] Step 3-2: Determine if the current input is the first input. If it is the first input, proceed to step 3-3; otherwise, proceed to step 3-4.

[0193] Step 3-3: Construct the Kalman filter model and initialize the effective state vector and Kalman filter model parameters;

[0194] Return to step 3-1;

[0195] Steps 3-4 involve using Kalman filtering to process the target's state information data, thereby enabling the prediction of the next frame's data based on the previous frame's data.

[0196] Step 4: Use a genetic algorithm to match the target predicted state value obtained from the filtering process with the target observed state value. Combined with... Figure 8 The process includes:

[0197] (1) Calculate node preheating data and obtain the confidence interval, confidence level, and termination iteration information of the equipment group configuration mutation operator;

[0198] (2) Obtain the target matching cost. During initialization, perform pairwise matching on the targets, calculate the Euclidean distance between the targets, use this cost as the input parameter of the genetic algorithm, calculate the sample fitness, initialize the sample population, and perform genetic operations such as selection, crossover, and mutation at the current time.

[0199] (3) Perform Time-based matching, device input One goal, 1 moment formed One tracker, If no matching object is found during the time-matter fusion process, the historical trajectory is stored in the tracker, forming a relationship to be associated between K tracker predictions and Q target detection values. This generates the initial sample population for the genetic algorithm, with samples represented using symbolic encoding. In each iteration, the fitness of the samples is calculated to find the globally optimal matching solution. Before triggering the iteration termination condition, the sample fitness is calculated in each iteration, and the selection, crossover, and mutation processes are repeated until the algorithm iteration termination condition is met. Finally, the target matching fusion result is output, ending step 4.

[0200] Step 5: Based on the matching results, correct the parameters in the Kalman filter algorithm, specifically including:

[0201] Step 5-1: Determine whether the cumulative number of predictions recorded by the tracker has reached the preset cleanup threshold. If yes, proceed to step 5-5; otherwise, proceed to step 5-2.

[0202] Step 5-2: Calculate the Euclidean distance between the predicted value and the observed value, and determine whether the preset threshold condition is met. If so, proceed to step 5-3; otherwise, proceed to step 5-4.

[0203] Step 5-3: Correct the filter parameters, then proceed to step 5-6;

[0204] Step 5-4: Record the observation information without updating the filter parameters, and then proceed to step 5-6.

[0205] Step 5-5, clean the tracker;

[0206] Steps 5-6: Update and output the tracker information.

[0207] Step 6: Based on the updated Kalman filter algorithm model, predict the target state.

[0208] The fusion results of the road scene obtained by the method of the present invention are as follows: Figure 9 As shown, the fusion result of the road scene obtained by the method of the present invention is as follows: Figure 10 As shown, this invention maximizes the utilization of computing resources while ensuring the accuracy and robustness of actual target tracking, effectively solving the target fusion problem in holographic scenes.

[0209] In summary, this invention integrates sensor information, overcoming problems such as inaccurate target detection by a single sensor, poor static target detection, missegmentation of large targets, and susceptibility to environmental interference in video detection. It achieves high precision, high robustness, and trajectory continuity in target detection in holographic traffic scenarios.

[0210] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention without departing from its spirit and scope should be included within the protection scope of the present invention.

Claims

1. A global target unique identification method based on multi-source sensor data fusion, characterized in that, The method realizes periodic updating of overlapping area calculation and node calculation tasks, and comprises the following steps: Step 1, data acquisition and processing: acquiring state information data of a target, i.e. target observation state values, through a plurality of source sensor devices, and eliminating abnormal data; the state information data comprises longitude and latitude position information, speed and acceleration of the target; Step 2, calculating detection overlapping areas between the plurality of source sensor devices; Step 3, based on the detection overlapping areas, filtering the target state information data detected by each sensor device by using a Kalman filtering algorithm; Step 4, matching the target prediction state values obtained by filtering with the target observation state values by using a genetic algorithm; Step 5, according to the matching result, correcting parameters in the Kalman filtering algorithm, and completing model parameter updating of the Kalman filtering algorithm in multiple iterations; Step 6, based on the updated Kalman filtering algorithm model, realizing prediction of the target state. 2.The global target unique identification method based on multi-source sensor data fusion according to claim 1, wherein, Step 1 specifically comprises: Step 1-1, building an application system comprising a computing node and a distribution node, deploying sensor devices in a traffic scene, and establishing network cascading relationships between the sensor devices; if there are detection overlapping areas between the sensor devices, the sensor devices must be in the same network; the computing node is used to deploy various calculation algorithm models for calculating and processing data; the distribution node is used to deploy data routing rules for initializing data and various calculation algorithm models, and is also used to store and share external inputs; Step 1-2, acquiring state information data of a target through a sensor device, and transmitting the data to the distribution node; Step 1-3, the distribution node performs dirty data processing on the received data to remove abnormal data. 3.The global target unique identification method based on multi-source sensor data fusion according to claim 2, characterized in that, In step 2, the detection overlapping areas between the plurality of source sensor devices are calculated, specifically comprising: Step 2-1, the distribution node constructs a global sensor device cascading relationship to form a device group relationship; each device group is a structured relationship from upstream to downstream of a detection area; Step 2-2, the distribution node completes device group detection data task construction, establishes a corresponding relationship between each device group and each computing node, forms a distribution strategy, and ensures that there is no overlap between the overlapping areas of different device groups; at the same time, the distribution node initializes parameters of the Kalman filtering algorithm and the genetic algorithm; Step 2-3, for each device group, performing frame combining processing on data collected by all sensor devices in the device group; Step 2-4, for each device group, calculating overlapping areas, specifically comprising: calculating the overlapping area formed by the device group by traversing target position information collected by all sensor devices in the device group by using a convex hull algorithm; Step 2-5, judging whether there is overlap between the overlapping areas corresponding to each device group; if yes, returning to step 2-2, otherwise ending the whole process.

4. The method of claim 3, wherein the method further comprises: Step 2-4 specifically comprises: calculating the overlapping area formed by the device group by traversing target position information collected by all sensor devices in the device group by using a convex hull algorithm.

5. The method of claim 4, wherein, The overlapping region is saved in the form of a feature point set, and the point set is connected in sequence to form a final effective overlapping polygon region. 6.The global target unique identification method based on multi-source sensor data fusion according to claim 1, wherein, In step 3, based on the detected overlapping region, the target state information data detected by each sensor device is filtered by using a Kalman filtering algorithm, and specifically includes the following steps: Step 3-1, judge whether the target is in a certain detection overlapping region, if yes, jump to step 3-2; Step 3-2, judge whether the current input is the first input, if yes, go to step 3-3, otherwise, go to step 3-4; Step 3-3, construct a Kalman filtering model, and initialize the effective state vector and the Kalman filtering model parameters; where the effective state vector is: ； wherein is the longitude, is the latitude, is is the velocity in the direction, is is the velocity in the direction, is is the acceleration in the direction, is is the acceleration in the direction; wherein the target is set in the positive north direction as is the direction, set in the positive east direction as is the direction; Return to step 3-1; Step 3-4, filter the state information data of the target by using the Kalman filter, thereby realizing the prediction of the next frame data based on the previous frame data.

7. The method of claim 1, wherein the method further comprises: In step 4, the target predicted state value obtained by filtering is matched with the target observed state value by using a genetic algorithm, and specifically includes the following steps: Step 4-1, initial population size, generate initial population; Step 4-2, judge whether the termination iteration condition is met, if yes, jump to step 4-5, otherwise, execute step 4-3; Step 4-3, calculate the fitness function: ； ； wherein, is the fitness of the sample to the target, is the cost of matching the detected target to the tracker state prediction, M is the number of elements in the observation set input to the detector, N is the number of elements in the tracker set produced by the filter algorithm, is the Euclidean distance between the observation and the prediction, i, j are the tracker and observation target indices respectively, is the Euclidean distance between the observation and the tracker prediction, is the observation vector of the jth target, is the predicted state vector of the ith tracker, i.e. the prediction produced by the tracker in the Kalman filter; parameters is the missed detection tolerance in the sample, when the Euclidean distance exceeds this range, i.e. considered as unreasonable matching, the sample formed by this matching scheme will be mutated in the subsequent iteration according to the mutation operator; Then execute step 4-4; Step 4-4, the population sample is selected, crossed, and mutated to generate a new population, and then jump to step 4-2; Step 4-5, end the process and output the global optimal individual as the optimal matching result. 8.The global target unique identification method based on multi-source sensor data fusion according to claim 7, wherein, Step 4-4 introduces the missed detection operator, target switching operator, and target mutation operator to mutate the sample. The confidence and confidence interval are defined as follows As shown: ； wherein is a missed detection operator, is a confidence of is a target switch operator, is a confidence of is a target variation operator, is a confidence of 9.The global target unique identification method based on multi-source sensor data fusion according to claim 7, wherein, In step 5, according to the matching result, the parameters in the Kalman filtering algorithm are corrected, and specifically includes the following steps: Step 5-1, judge whether the prediction number cumulative value recorded by the tracker reaches a preset cleaning threshold, if yes, execute step 5-5, otherwise, execute step 5-2; Step 5-2, calculate the Euclidean distance between the predicted value and the observed value, and judge whether it meets the preset threshold condition, if yes, execute step 5-3, otherwise, execute step 5-4; Step 5-3, correct the filtering parameters, and then execute step 5-6; Step 5-4, record the observed value information, and do not update the filtering parameters, and then execute step 5-6; Step 5-5, clean the tracker; Step 5-6, update and output the tracker information.

10. A global object unique identification system based on the method according to any one of claims 1 to 9, characterized in that, The system includes: A first module for realizing: collecting target state information data, i.e. target observed state value, by multiple source sensor devices, and eliminating abnormal data; the state information data includes the latitude and longitude position information, speed, and acceleration of the target; A second module for realizing: calculating the detection overlapping region between multiple source sensor devices; A third module for realizing: based on the detection overlapping region, filtering the target state information data detected by each sensor device by using a Kalman filtering algorithm; A fourth module for realizing: matching the target predicted state value obtained by filtering with the target observed state value by using a genetic algorithm; A fifth module for realizing: correcting the parameters in the Kalman filtering algorithm according to the matching result, and updating the Kalman filtering algorithm model parameters in multiple iterations; The sixth module is used for realizing the prediction of the target state based on the updated Kalman filtering algorithm model. The sixth module is used for realizing the prediction of the target state based on the updated Kalman filtering algorithm model.

Images