High-precision target tracking method capable of preventing perception loss

Abstract

The invention belongs to the technical field of target tracking, and discloses a high-precision target tracking method capable of preventing perception loss, which comprises the steps of recording target structured data by using a camera before a target enters a basement, making a feature vector, detecting a target object by using a laser point cloud feature map, making the feature map into the feature vector, and tracking the target object by using the feature vector. And the target client sends a positioning position to the basement, clock synchronization is carried out on the camera feature vector, the laser feature vector and the positioning position, cross multiplication is carried out on the camera feature vector and the laser feature vector, and the similarity is calculated. According to the application, an active positioning system is combined, original isolated systems are integrated, the anti-lost capability is improved, active positioning data are not added with features, checking is facilitated, three-dimensional data of laser is used for preventing lost, the tracking capability is remarkably improved, a laser radar feature map is directly utilized, multiple sensors are used, and target feature description is richer.

Description

Technical Field

[0001] The present application relates to the technical field of target tracking, and more specifically, to a high-precision target tracking method capable of preventing perception loss. Background Art

[0002] The existing underground garage target anti-loss technology mainly uses surveillance cameras to identify license plates and track them as unique identifiers, or by laying UWB positioning base stations and then configuring positioning beacons for targets for high-precision active positioning, or by using manual monitoring.

[0003] The existing methods for preventing basement targets from being lost have the following drawbacks:

[0004] High cost: A large number of surveillance cameras or positioning base stations need to be deployed, and manual monitoring cannot be done 24 hours a day;

[0005] Poor anti-loss capability: When manually monitoring, the anti-loss capability decreases significantly as the number of targets increases. Using license plates as the only identifier will also be affected by the recognition rate. If a vehicle has a duplicate license plate, it cannot be processed.

[0006] High backtracking cost: When backtracking the target trajectory through videos, a large amount of videos needs to be reviewed, which is costly.

[0007] Centralized computing: It is necessary to deploy centralized computing units with large computing power. If the computing units fail, the overall perception effect will be lost. Centralized computing requires synchronization of all data, which takes time and increases the overall perception delay.

[0008] Poor target feature extraction capability: General solutions only use simple vision or positioning hardware, which has weak target feature expression capability and is not conducive to cross-regional anti-loss;

[0009] Low accuracy: It is difficult to achieve centimeter-level positioning using surveillance cameras, and most of them remain at meter-level positioning. Summary of the invention

[0010] In order to solve the above problems, the present application provides a high-precision target tracking method that can prevent perception loss.

[0011] The present application provides a high-precision target tracking method capable of preventing perception loss, which adopts the following technical solutions:

[0012] A high-precision target tracking method capable of preventing perception loss comprises the following steps:

[0013] S1. Use a camera to record the target structured data before the target enters the basement and make it into a feature vector;

[0014] S2, using the laser point cloud feature map to detect the target object, and at the same time making the feature map into a feature vector;

[0015] S3. If the target uses indoor and outdoor navigation technology such as Bluetooth or GPS navigation, the target client sends the location to the basement;

[0016] S4, performing clock synchronization on the camera feature vector, the laser feature vector, and the positioning position;

[0017] S5, cross-multiply the camera feature vector and the laser feature vector, and calculate the similarity;

[0018] S6, performing feature extraction transformation on the continuous asynchronous time function data outputted by the sensor structure, and extracting a feature vector Yi representing the observed data;

[0019] S7, recording the characteristic vector Yi to the historical data;

[0020] S8, draw a circle with the initial target position as the reference and r as the radius, and then calculate the circumscribed rectangle of the circle;

[0021] S9, bringing the feature vector Yi+1 of the next moment into the historical data for matching;

[0022] S10, performing iterative filtering on the feature vector Yi to fuse the active positioning data, and completing the target fusion of each sensor;

[0023] S11, using a fusion algorithm to synthesize the sensor data of each target to obtain a consistent interpretation and description of the target;

[0024] S12. Combine the feature vector with the structured data to extract the main point cloud image data and store them in the database.

[0025] Furthermore, the similarity calculation in S5 includes the following formula:

[0026]

[0027] Furthermore, the correlation feature used for matching in S9 adopts CIoU, and the formula is:

[0028]

[0029] Furthermore, the iterative filter fusion of the feature vector Yi into active positioning data in S10 includes:

[0030] Calculate the L2 norm based on Yi active positioning point and Yi+1 active positioning point;

[0031] Determine whether the calculated result meets the distance. If so, Yi+1 will cover the Yi active positioning point. If not, Yi will be eliminated.

[0032] Merge the positioning point with the feature vector to generate a new Yi and import it into the historical database.

[0033] Furthermore, the step of synthesizing the sensor data of each target by using a fusion algorithm in S11 includes:

[0034] Obtain active positioning data and determine the confidence level. If the confidence level meets the requirements, use Yi positioning data to update. If not, use Yi+1 fused positioning data.

[0035] Determine whether the positioning data is updated. If so, merge the positioning data with the image feature vector and the laser feature vector through neural network fusion. If not, set the not updated flag and merge the neural network features.

[0036] Furthermore, the step of extracting the main point cloud image data by combining the feature vector with the structured data and storing the extracted data in the storage includes:

[0037] Get the original image data, original point cloud data and positioning data, and segment them using structured data; extract the pixels at the target location in the image, extract the data at the target location in the point cloud, merge the positioning points, and write the data to disk.

[0038] In summary, the present application includes at least one of the following beneficial technical effects:

[0039] (1) Combined with the active positioning system, the original isolated system is integrated to improve the anti-loss capability, and features are added to the active positioning data for easy viewing;

[0040] (2) Using laser three-dimensional data to prevent loss and significantly improve tracking capabilities;

[0041] (3) Pre-fusion, making the best use of raw data and reducing output latency;

[0042] (4) Directly use the lidar feature map;

[0043] (5) Using multiple sensors, the target features are described more richly;

[0044] (6) Reduce total costs. Data storage only needs to store structured data and feature vectors. There is no need to store large amounts of video data or manual monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 A schematic diagram of the method flow of this application;

[0046] Figure 2It is a schematic diagram of the process of iteratively filtering the feature vector Yi to fuse the active positioning data;

[0047] Figure 3 A schematic diagram of the process of synthesizing the sensor data of each target using the fusion algorithm;

[0048] Figure 4 A schematic diagram of the process of extracting the main point cloud image data and storing it in the database by combining the feature vector with the structured data. DETAILED DESCRIPTION

[0049] The technical solutions 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; obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments, and all other embodiments obtained by ordinary technicians in this field based on the embodiments in the present application without making creative work are within the scope of protection of the present application.

[0050] In the description of the present application, it should be noted that the terms "upper", "lower", "inner", "outer", "top / bottom" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific position, be constructed and operated in a specific position, and therefore cannot be understood as limiting the present application. In addition, the terms "first" and "second" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance.

[0051] In the description of this application, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "provided with", "set / connected", "connected", etc. should be understood in a broad sense. For example, "connected" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in this application can be understood according to specific circumstances.

[0052] Embodiment 1:

[0053] The present application embodiment discloses a high-precision target tracking method capable of preventing perception loss, comprising the following steps:

[0054] S1. Use a camera to record the target structured data before the target enters the basement and make it into a feature vector;

[0055] S2, using the laser point cloud feature map to detect the target object, and at the same time making the feature map into a feature vector;

[0056] S3. If the target uses indoor and outdoor navigation technology such as Bluetooth or GPS navigation, the target client sends the location to the basement;

[0057] S4, performing clock synchronization on the camera feature vector, the laser feature vector, and the positioning position;

[0058] S5. Perform cross multiplication of the camera feature vector and the laser feature vector, and calculate the similarity. The formula is as follows:

[0059]

[0060] S6, performing feature extraction transformation on the continuous asynchronous time function data outputted by the sensor structure, and extracting a feature vector Yi representing the observed data;

[0061] S7, recording the characteristic vector Yi to the historical data;

[0062] S8, draw a circle with the initial target position as the reference and r as the radius, and then calculate the circumscribed rectangle of the circle;

[0063] S9, bring the feature vector Yi+1 of the next moment into the historical data for matching, and use CIoU as the correlation feature, the formula is as follows:

[0064]

[0065] S10, iteratively filter the feature vector Yi to fuse the active positioning data, and complete the target fusion of each sensor. The steps include:

[0066] Calculate the L2 norm based on Yi active positioning point and Yi+1 active positioning point;

[0067] Determine whether the calculated result meets the distance. If so, Yi+1 will cover the Yi active positioning point. If not, Yi will be eliminated.

[0068] Merge the positioning point with the feature vector to generate a new Yi and import it into the historical database;

[0069] S11, using a fusion algorithm to synthesize the sensor data of each target to obtain a consistent interpretation and description of the target, the steps include:

[0070] Obtain active positioning data and determine the confidence level. If the confidence level meets the requirements, use Yi positioning data to update. If not, use Yi+1 fused positioning data.

[0071] Determine whether the positioning data is updated. If so, merge the positioning data with the image feature vector and the laser feature vector by neural network fusion. If not, set the not updated flag and merge the neural network features.

[0072] S12, extracting the main point cloud image data by combining the feature vector with the structured data and storing it in the database, the steps include:

[0073] Get the original image data, original point cloud data and positioning data, and segment them using structured data; extract the pixels at the target location in the image, extract the data at the target location in the point cloud, merge the positioning points, and write the data to disk.

[0074] This application combines the active positioning system to integrate the original isolated system, improve the anti-loss capability, and add features to the active positioning data for easy viewing; the three-dimensional data of the laser is used for anti-loss, which significantly improves the tracking capability; this application is a front fusion, which makes the best use of the original data and reduces output delay; this application directly uses the laser radar feature map and uses multiple sensors to provide a richer target feature description; the method described in this application can further reduce the total cost, and data storage only needs to store structured data and feature vectors, without the need to store large amounts of video data or manual monitoring, which is more practical.

[0075] The above are all preferred embodiments of the present application, and the protection scope of the present application is not limited thereto. Therefore, any equivalent changes made according to the structure, shape, and principle of the present application should be included in the protection scope of the present application.

Claims

1. A high-precision target tracking method capable of preventing perception loss, characterized in that: The following steps are involved: S1. Use a camera to record the target structured data before the target enters the basement and make it into a feature vector; S2, using the laser point cloud feature map to detect the target object, and at the same time making the feature map into a feature vector; S3. If the target uses indoor and outdoor navigation technology such as Bluetooth or GPS navigation, the target client sends the location to the basement; S4, performing clock synchronization on the camera feature vector, the laser feature vector, and the positioning position; S5, cross-multiply the camera feature vector and the laser feature vector, and calculate the similarity; S6, performing feature extraction transformation on the continuous asynchronous time function data outputted by the sensor structure, and extracting a feature vector Yi representing the observed data; S7, recording the characteristic vector Yi to the historical data; S8, draw a circle with the initial target position as the reference and r as the radius, and then calculate the circumscribed rectangle of the circle; S9, bringing the feature vector Yi+1 of the next moment into the historical data for matching; S10, performing iterative filtering on the feature vector Yi to fuse the active positioning data, and completing the target fusion of each sensor; S11, using a fusion algorithm to synthesize the sensor data of each target to obtain a consistent interpretation and description of the target; S12. Combine the feature vector with the structured data to extract the main point cloud image data and store them in the database.

2. A high-precision target tracking method capable of preventing perception loss according to claim 1, characterized in that: The similarity calculation in S5 includes the following formula:

3. A high-precision target tracking method capable of preventing perception loss according to claim 1, characterized in that: The correlation feature used for matching in S9 is CIoU, and the formula is:

4. A high-precision target tracking method capable of preventing perception loss according to claim 1, characterized in that: The step of performing iterative filtering on the feature vector Yi to fuse the active positioning data in S10 includes: Calculate the L2 norm based on Yi active positioning point and Yi+1 active positioning point; Determine whether the calculated result meets the distance. If so, Yi+1 will cover the Yi active positioning point. If not, Yi will be eliminated. Merge the positioning point with the feature vector to generate a new Yi and import it into the historical database.

5. The high-precision target tracking method capable of preventing perception loss according to claim 1, characterized in that: The S11 uses a fusion algorithm to synthesize the sensor data of each target, including: Obtain active positioning data and determine the confidence level. If the confidence level meets the requirements, use Yi positioning data to update. If not, use Yi+1 fused positioning data. Determine whether the positioning data is updated. If so, merge the positioning data with the image feature vector and the laser feature vector through neural network fusion. If not, set the not updated flag and merge the neural network features.

6. A high-precision target tracking method capable of preventing perception loss according to claim 1, characterized in that: In S12, the feature vector is combined with the structured data to extract the main point cloud image data and store them in the database, including: Get the original image data, original point cloud data and positioning data, and segment them using structured data; extract the pixels at the target location in the image, extract the data at the target location in the point cloud, merge the positioning points, and write the data to disk.

Images