A target detection method, system and storage medium

Abstract

The application discloses a target detection method and system and a storage medium, and the method comprises the following steps: preprocessing millimeter wave radar point cloud data to obtain pedestrian target point cloud data; setting initial values of a particle filtering algorithm; initializing a Dirac particle array in the particle filtering algorithm by using radar precision, tracking frame number and number; predicting the pedestrian target and outputting a prediction value; performing target matching by using particle filtering, updating particle weights and normalizing; and updating a target pool to output a target. The application solves the problem of visual target jumping in insufficient light and the problem of misidentification or missed identification caused by excessive clutter in a relatively closed space of the millimeter wave radar, and can obviously improve the identification accuracy of the pedestrian target, thereby improving the safety of driving.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, and more specifically, to a target detection method, system, and storage medium. Background Technology

[0002] Intelligent electric drive systems (battery, motor, and electronic control), intelligent cockpits, and autonomous driving represent the future development trend of automobiles. The software system for autonomous driving is generally defined by six modules: perception, prediction, high-precision mapping, localization, decision-making and planning, and control. The perception module typically refers to obstacle and traffic light recognition, and the sensors used mainly include LiDAR, millimeter-wave radar, and cameras. These three types of sensors have different advantages in different environments.

[0003] Among these, cameras are inexpensive and possess advantages in object detection and classification, making them widely used in ADAS (Advanced Driver Assistance Systems). However, their sensitivity to changes in lighting and reliance on deep learning network models and training datasets make them susceptible to failure. LiDAR, as an active sensor, offers advantages such as high accuracy and strong 3D perception capabilities, but its high cost and susceptibility to weather conditions make it a popular choice. Millimeter-wave radar, on the other hand, is relatively inexpensive, possesses 3D point cloud and speed perception capabilities, and can operate in all weather conditions with high reliability. Therefore, millimeter-wave radar is widely used in ADAS systems at various levels.

[0004] ADAS (Advanced Driver Assistance Systems) places higher demands on perception systems. When tracking pedestrians in underground parking garages, the system needs to accurately perceive the pedestrian's position and trajectory to facilitate the decision-making and planning modules in calculating reasonable decisions and paths, thus avoiding potential safety risks. Due to dim lighting, visual pedestrian target detection is prone to missed detections, misidentifications, or significant movement, making the tracking and matching of visual sensors in parking garages less than satisfactory. For example, Chinese invention patent application CN202010995314.5 discloses "A Short-Range Person Target Tracking Method Based on Millimeter-Wave Radar." While the Kalman-type millimeter-wave radar target processing method used in this solution can effectively address the problem of visual target movement due to conventional tracking, the linear tracking method is ineffective due to the instability of pedestrian movement trajectories. Summary of the Invention

[0005] To overcome the problems existing in the prior art, the present invention provides a target detection method, system, and storage medium.

[0006] The primary objective of this invention is to solve the aforementioned technical problems. The technical solution of this invention is as follows:

[0007] The first aspect of this invention provides a target detection method, comprising the following steps:

[0008] Preprocess millimeter-wave radar point cloud data to obtain point cloud data of pedestrian targets;

[0009] Set the initial values ​​for the particle filter algorithm;

[0010] The Dirac particle array in the particle filter algorithm is initialized using radar accuracy, tracking frame number, and number. Specifically, a two-dimensional array with a preset number of rows m and columns n is constructed for each tracking frame number TrackID. The first items of the two-dimensional array are TrackID and TrackFrame, and the distance of each cell in the two-dimensional array is denoted as DmDn, where the value of DmDn is the radar distance accuracy.

[0011] Predict pedestrian targets and output the predicted values;

[0012] Particle filtering is used for target matching, and particle weights are updated and normalized. Specifically, the pedestrian target obtained from the preprocessed point cloud data is used as the new round of observation value; the Euclidean distance between the new target corresponding to the new round of observation value and each target in the target pool is calculated, and the target in the target pool corresponding to the smallest Euclidean distance is used as the first new target. The first new target is matched and verified by energy threshold to obtain the second new target.

[0013] The particle weights in the two-dimensional array of the target TrackID are modified using a second new target, and the modified weights are normalized.

[0014] Update the target pool output target.

[0015] Furthermore, the preprocessing of the millimeter-wave radar point cloud data specifically involves:

[0016] The first frame of millimeter-wave radar point cloud data is filtered using a preset first threshold to remove misidentified points;

[0017] Calculate the position and velocity of each point in the point cloud data;

[0018] Based on the position and velocity of the points, the point cloud data is sequentially clustered, RCS pedestrian and vehicle distinguished, and second threshold filtering is performed to obtain the point cloud data of the pedestrian target.

[0019] Furthermore, the initial value of the particle filter algorithm is set as follows: the number of tracking frames of the pedestrian target is denoted as TrackFrame, TrackFrame is set to a set value and numbered, the number is denoted as TrackID, and the pedestrian target is stored in the target pool TargetPool as the initial value of the particle filter algorithm.

[0020] Furthermore, when the TrackID is 1, the initial particle weight is 0.01 for each item.

[0021] Furthermore, the pedestrian target is predicted and the predicted value is output, specifically as follows:

[0022] Set the state equations for the points in the point cloud in the X and Y directions as follows:

[0023] Xk+1 = Vx * frame time + Xk

[0024] Yk+1=Vy*frametime+Yk, where k is the number of iterations, and the value of k is equal to the number of tracking frames TrackFrame. The updated target is denoted as PredictTarget, and TrackID and other attributes are retained from the initial pedestrian target.

[0025] Furthermore, update the target pool output target, specifically as follows:

[0026] Update the TrackID target position using the updated particle weights, and output the target when the number of tracking frames TrackFrame is greater than the set threshold TrackFrameThresh.

[0027] A second aspect of the present invention provides a target detection system, the system comprising: a memory and a processor, wherein the memory includes a target detection method program, and the target detection method program, when executed by the processor, performs the following steps:

[0028] Preprocess millimeter-wave radar point cloud data to obtain point cloud data of pedestrian targets;

[0029] Set the initial values ​​for the particle filter algorithm;

[0030] The Dirac particle array in the particle filter algorithm is initialized using radar accuracy, tracking frame number, and number. Specifically, a two-dimensional array with a preset number of rows m and columns n is constructed for each tracking frame number TrackID. The first items of the two-dimensional array are TrackID and TrackFrame, and the distance of each cell in the two-dimensional array is denoted as DmDn, where the value of DmDn is the radar distance accuracy.

[0031] Predict pedestrian targets and output the predicted values;

[0032] Particle filtering is used for target matching, and particle weights are updated and normalized. Specifically, the pedestrian target obtained from the preprocessed point cloud data is used as the new round of observation value; the Euclidean distance between the new target corresponding to the new round of observation value and each target in the target pool is calculated, and the target in the target pool corresponding to the smallest Euclidean distance is used as the first new target. The first new target is matched and verified by energy threshold to obtain the second new target.

[0033] The particle weights in the two-dimensional array of the target TrackID are modified using a second new target, and the modified weights are normalized.

[0034] Update the target pool output target.

[0035] A third aspect of the present invention provides a computer-readable storage medium including a target detection method program, wherein when the target detection method program is executed by a processor, it implements the steps of the target detection method.

[0036] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0037] This invention is based on a particle filtering algorithm. It initializes the Dirac particle array using radar accuracy, tracking frame count, and numbering, and then updates the weights in the particle array through prediction iteration. This effectively reduces the uncertainty of pedestrian target detection caused by the randomness of particle generation in the particle filtering algorithm. This makes pedestrian state detection no longer dependent on linear behavior prediction, and solves the problem of visual target hopping in low light conditions and the problem of false or missed identification caused by excessive clutter in relatively enclosed spaces by millimeter-wave radar. It can significantly improve the recognition accuracy of pedestrian targets, thereby improving driving safety. Attached Figure Description

[0038] Figure 1 This is a flowchart of a target detection method according to an embodiment of the present invention.

[0039] Figure 2 Flowchart for initializing the Dirac particle array in an embodiment of the present invention.

[0040] Figure 3 This is a flowchart of target matching using particle filtering in an embodiment of the present invention. Detailed Implementation

[0041] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0042] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0043] Example 1

[0044] like Figure 1 As shown, the first aspect of the present invention provides a target detection method, comprising the following steps:

[0045] S1. Preprocess millimeter-wave radar point cloud data to obtain point cloud data of pedestrian targets;

[0046] It should be noted that in this invention, point cloud data from millimeter-wave radar can be received via Ethernet. First, the first frame of point cloud data undergoes physical value transformation. Based on the vehicle coordinate system definition and trigonometric relationships, distance and azimuth are converted into X and Y coordinates. Then, the coordinate values ​​are transformed back to the vehicle coordinate system and stored in global variables. Next, a preset first threshold is used for filtering to remove falsely identified points. This first threshold is a coarse threshold; for example, an energy threshold can be set to remove falsely identified points, and noise can also be filtered out using the signal-to-noise ratio and X and Y values ​​to improve the quality of the point cloud. After the first threshold filtering, the point cloud data is sequentially clustered, RCS (Radar Cross Section) distinguished between vehicles and pedestrians, and filtered again using a second threshold to obtain the point cloud data for pedestrian targets. The second threshold in this invention can be distance, speed, angle, etc.

[0047] S2. Set the initial values ​​for the particle filter algorithm;

[0048] It should be noted that for the first frame of point cloud data, the tracking frame number of the pedestrian target is denoted as TrackFrame. TrackFrame is set to a set value and numbered, and the number is denoted as TrackID. The pedestrian target is stored in the target pool TargetPool as the initial value of the particle filter algorithm.

[0049] S3. Initialize the Dirac particle array in the particle filter algorithm using radar accuracy, tracking frame count, and numbering.

[0050] like Figure 2 As shown, in a specific embodiment, the initialization process is as follows:

[0051] S301. Construct a two-dimensional array with a preset number of rows m and a preset number of columns n for each tracked frame number TrackID;

[0052] S302, the first items of the two-dimensional array are TrackID and TrackFrame, respectively, and the distance of each cell in the two-dimensional array is denoted as DmDn, where the value of DmDn is the radar's distance accuracy.

[0053] It should be noted that, considering that a large array would place a heavy burden on computing power, while a small array would lead to inaccurate target prediction, a two-dimensional array DiracArray with m and n of size 101 (this is an adjustable parameter) is created for each TrackID. The m and n values ​​of the first item of DiracArray are TrackID and TrackFrame, respectively, to facilitate the matching of targets and particles. Considering the need for a stable particle method to reduce computing power and randomness, the distance DmDn of each cell is the radar's distance accuracy Dmin. Since it is not initialized, TrackID is 1, so the weight of each particle is set to 1 / 100 = 0.01.

[0054] S4. Predict pedestrian targets and output the predicted values;

[0055] In a specific embodiment, the state equations for the points in the point cloud in the X and Y directions are set as follows:

[0056] Xk+1 = Vx * frame time + Xk

[0057] Yk+1=Vy*frametime+Yk, where k is the number of iterations, and the value of k is equal to the number of tracking frames TrackFrame. The updated target is denoted as PredictTarget, and TrackID and other attributes are retained from the initial pedestrian target.

[0058] S5. Use particle filtering for target matching, update particle weights and normalize them.

[0059] like Figure 3 As shown, in a specific embodiment, the specific steps are as follows:

[0060] S501 uses the pedestrian targets obtained from the preprocessed point cloud data as the new round of observations;

[0061] S502 calculates the Euclidean distance between the new target corresponding to the new round of observations and each target in the target pool, and selects the target in the target pool with the smallest Euclidean distance as the first new target.

[0062] S503 obtains a second new target after matching and verifying the first new target through an energy threshold;

[0063] S504 modifies the particle weights in the two-dimensional array of the target TrackID using the second new target and normalizes the modified weights.

[0064] The modification of particle weights is achieved by traversing a two-dimensional array. Taking the X direction as an example, a simplified version of the likelihood probability formula is used and normalized to achieve the effect of the update step.

[0065] The simplified expression for the likelihood probability formula is as follows:

[0066] DiracArray(i)=exp(-((NewTarget.x-(PredictTarget.x)2)2 / (2*0.001))

[0067] S6. Update the target pool output target.

[0068] In one specific embodiment, the TrackID target position is updated using the updated particle weights, and the target is output when the number of tracking frames TrackFrame is greater than the set threshold TrackFrameThresh.

[0069] A second aspect of the present invention provides a target detection system, the system comprising: a memory and a processor, wherein the memory includes a target detection method program, and the target detection method program, when executed by the processor, performs the following steps:

[0070] S1 preprocesses millimeter-wave radar point cloud data to obtain point cloud data of pedestrian targets;

[0071] It should be noted that in this invention, point cloud data from millimeter-wave radar can be received via Ethernet. First, the first frame of point cloud data undergoes physical value transformation. Based on the vehicle coordinate system definition and trigonometric relationships, distance and azimuth are converted into X and Y coordinates. Then, the coordinate values ​​are transformed back to the vehicle coordinate system and stored in global variables. Next, a preset first threshold is used for filtering to remove misidentified points. This first threshold is a coarse threshold; for example, an energy threshold can be set to remove misidentified points, and noise can also be filtered out using signal-to-noise ratio and XY values ​​to improve the quality of the point cloud. After the first threshold filtering, the point cloud data is sequentially clustered, RCS-based for pedestrian and vehicle differentiation, and filtered again using a second threshold to obtain the point cloud data for pedestrian targets. The second threshold in this invention can be distance, speed, angle, etc.

[0072] S2. Set the initial values ​​for the particle filter algorithm;

[0073] It should be noted that for the first frame of point cloud data, the tracking frame number of the pedestrian target is denoted as TrackFrame. TrackFrame is set to a set value and numbered, and the number is denoted as TrackID. The pedestrian target is stored in the target pool TargetPool as the initial value of the particle filter algorithm.

[0074] S3. Initialize the Dirac particle array in the particle filter algorithm using radar accuracy, tracking frame count, and numbering.

[0075] In one specific embodiment, a two-dimensional array of size X and Y is constructed for each TrackID number of the tracking frame, wherein X represents the number of distance points in the X direction and Y represents the number of distance points in the Y direction;

[0076] The first items of the two-dimensional array are TrackID and TrackFrame, respectively. The distance of each cell in the two-dimensional array is denoted as DxDy, and the value of DxDy is the radar's distance accuracy.

[0077] It should be noted that, considering that a large array would place a heavy burden on computing power, while a small array would lead to inaccurate target prediction, a two-dimensional array DiracArray with m and n of size 101 (this is an adjustable parameter) is created for each TrackID. The m and n values ​​of the first item of DiracArray are TrackID and TrackFrame, respectively, to facilitate the matching of targets and particles. Considering the need for a stable particle method to reduce computing power and randomness, the distance DmDn of each cell is the radar's distance accuracy Dmin. Since it is not initialized, TrackID is 1, so the weight of each particle is set to 1 / 100 = 0.01.

[0078] S4. Predict pedestrian targets and output the predicted values;

[0079] In a specific embodiment, the state equations for the points in the point cloud in the X and Y directions are set as follows:

[0080] Xk+1 = Vx * frame time + Xk

[0081] Yk+1=Vy*frametime+Yk, where k is the number of iterations, and the value of k is equal to the number of tracking frames TrackFrame. The updated target is denoted as PredictTarget, and TrackID and other attributes are retained from the initial pedestrian target.

[0082] S5. Use particle filtering for target matching, update particle weights and normalize them.

[0083] In one specific embodiment, the pedestrian target obtained from the preprocessed point cloud data is used as the new round of observations;

[0084] Calculate the Euclidean distance between the new target corresponding to the new round of observations and each target in the target pool. Select the target in the target pool with the smallest Euclidean distance as the first new target.

[0085] A second new target is obtained by matching and verifying the first new target using an energy threshold.

[0086] The particle weights in the two-dimensional array of the target TrackID are modified using a second new target, and the modified weights are normalized.

[0087] The modification of particle weights is achieved by traversing a two-dimensional array. Taking the X direction as an example, a simplified version of the likelihood probability formula is used and normalized to achieve the effect of the update step.

[0088] The simplified expression for the likelihood probability formula is as follows:

[0089] DiracArray(i)=exp(-((NewTarget.x-(PredictTarget.x)2)2 / (2*0.001))

[0090] S6. Update the target pool output target.

[0091] In one specific embodiment, the TrackID target position is updated using the updated particle weights, and the target is output when the number of tracking frames TrackFrame is greater than the set threshold TrackFrameThresh.

[0092] A third aspect of the present invention provides a computer-readable storage medium including a target detection method program, wherein when the target detection method program is executed by a processor, it implements the steps of the target detection method.

[0093] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A target detection method, characterized in that, Includes the following steps: Preprocess millimeter-wave radar point cloud data to obtain point cloud data of pedestrian targets; Set the initial values ​​for the particle filter algorithm; The Dirac particle array in the particle filter algorithm is initialized using radar accuracy, tracking frame number, and number. Specifically, a two-dimensional array with a preset number of rows m and columns n is constructed for each tracking frame number TrackID. The first items of the two-dimensional array are TrackID and TrackFrame, and the distance of each cell in the two-dimensional array is denoted as DmDn, where the value of DmDn is the radar distance accuracy. Predict pedestrian targets and output predicted values; Particle filtering is used for target matching, and particle weights are updated and normalized. Specifically, the pedestrian target obtained from the preprocessed point cloud data is used as the new round of observation value; the Euclidean distance between the new target corresponding to the new round of observation value and each target in the target pool is calculated, and the target in the target pool corresponding to the smallest Euclidean distance is used as the first new target. The first new target is matched and verified by energy threshold to obtain the second new target. The particle weights in the two-dimensional array of the target TrackID are modified using a second new target, and the modified weights are normalized. Update the target pool output target.

2. The target detection method according to claim 1, characterized in that, The preprocessed millimeter-wave radar point cloud data specifically includes: The first frame of millimeter-wave radar point cloud data is filtered using a preset first threshold to remove misidentified points; Calculate the position and velocity of each point in the point cloud data; Based on the position and velocity of the points, the point cloud data is sequentially clustered, RCS pedestrian and vehicle distinguished, and second threshold filtering is performed to obtain the point cloud data of the pedestrian target.

3. The target detection method according to claim 1, characterized in that, The initial value setting of the particle filter algorithm is specifically as follows: the number of tracking frames of the pedestrian target is denoted as TrackFrame, TrackFrame is set to a set value and numbered, the number is denoted as TrackID, and the pedestrian target is stored in the target pool TargetPool as the initial value of the particle filter algorithm.

4. The target detection method according to claim 3, characterized in that, When TrackID is 1, the initial particle weight is 0.01 for each item.

5. The target detection method according to claim 3, characterized in that, The process involves predicting pedestrian targets and outputting the predicted values, specifically: Set the state equations for the points in the point cloud in the X and Y directions as follows: Xk+1 = Vx * frame time + Xk Yk+1=Vy*frametime+Yk, where k is the number of iterations, and the value of k is equal to the number of tracking frames TrackFrame. The updated target is denoted as PredictTarget, and TrackID and other attributes are retained from the initial pedestrian target.

6. The target detection method according to claim 3, characterized in that, Update the target pool output target, specifically: Update the TrackID target position using the updated particle weights, and output the target when the number of tracking frames TrackFrame is greater than the set threshold TrackFrameThresh.

7. A target detection system, characterized in that, The system includes: a memory and a processor. The memory includes a target detection method program, which, when executed by the processor, performs the following steps: Preprocess millimeter-wave radar point cloud data to obtain point cloud data of pedestrian targets; Set the initial values ​​for the particle filter algorithm; The Dirac particle array in the particle filter algorithm is initialized using radar accuracy, tracking frame number, and number. Specifically, a two-dimensional array with a preset number of rows m and columns n is constructed for each tracking frame number TrackID. The first items of the two-dimensional array are TrackID and TrackFrame, and the distance of each cell in the two-dimensional array is denoted as DmDn, where the value of DmDn is the radar distance accuracy. Predict pedestrian targets and output predicted values; Particle filtering is used for target matching, and particle weights are updated and normalized. Specifically, the pedestrian target obtained from the preprocessed point cloud data is used as the new round of observation value; the Euclidean distance between the new target corresponding to the new round of observation value and each target in the target pool is calculated, and the target in the target pool corresponding to the smallest Euclidean distance is used as the first new target. The first new target is matched and verified by energy threshold to obtain the second new target. The particle weights in the two-dimensional array of the target TrackID are modified using a second new target, and the modified weights are normalized. Update the target pool output target.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a target detection method program, which, when executed by a processor, implements the steps of a target detection method as described in any one of claims 1 to 6.

Images