False alarm target detection method based on XGBoost algorithm
Through the false alarm object detection method based on the XGBoost algorithm, the model is trained using time sequence data to automatically determine the threshold threshold, solving the problems of low accuracy and high cost of false alarm object detection in the existing technology, and achieving more efficient and accurate false alarm object detection.
Patent Information
- Application Number
- CN202311122867.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-08-31
- Publication Date
- 2025-02-28
AI Technical Summary
In the prior art, false alarm target detection relies on empirical parameter settings, resulting in low accuracy of detection results and a large number of tests are required to obtain reliable thresholds, increasing detection cost and difficulty.
Using the false alarm object detection method based on the XGBoost algorithm, the trained XGBoost model is obtained by building the initial XGBoost model and using the timing data of multiple sample units for training. Then, feature engineering is performed on the target's timing data, input the feature vector into the model, obtain the false alarm probability, and judge whether the target is a false alarm target based on the threshold value determined by the model.
It improves the accuracy of false alarm target detection results, avoids the dependence of empirical parameter settings, and reduces the detection cost and difficulty.
Smart Images
Figure CN119538107A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of perception fusion technology, and in particular to a false alarm target detection method based on an XGBoost algorithm. Background Art
[0002] The perception fusion of autonomous driving is used to integrate data information from multiple different types of sensors. The decision-level target fusion integrates the detection targets from sensors such as millimeter-wave radar, camera, and lidar to output the final perception result. Due to the characteristics of different sensors and their different target detection methods, there are false alarm targets in their detection results. False alarm targets can cause the vehicle to brake incorrectly or perform obstacle avoidance, seriously affecting the driving experience and driving safety. At present, the detection of false alarm targets in perception fusion needs to be judged based on the thresholds of multiple targets set manually. The judgment results rely on empirical parameter settings, resulting in low accuracy of the detection results. In addition, in order to obtain a more reliable threshold, a large number of tests are required, which increases the cost and difficulty of detection. Summary of the invention
[0003] The object of the present invention is to provide a false alarm target detection method based on the XGBoost algorithm to solve the problems existing in the prior art mentioned in the background technology.
[0004] In order to achieve the above object, the technical solution adopted by the present invention is as follows:
[0005] A false alarm target detection method based on XGBoost algorithm, comprising the steps of:
[0006] Build the initial XGBoost model with the goal of minimizing the loss function;
[0007] A trained XGBoost model is obtained according to time series data of a plurality of sample units, wherein the plurality of sample units include at least one false alarm unit and at least one normal unit, and the time series data include observation vectors corresponding to N historical frames, and the observation vectors are composed of a plurality of observation quantities;
[0008] Obtain the time series data of the target;
[0009] Perform feature engineering on the target’s time series data to obtain its eigenvalues and construct feature vectors;
[0010] Input the target’s feature vector into the trained XGBoost model to obtain the false alarm probability of the target;
[0011] The threshold value is determined according to the trained XGBoost model. When the false alarm probability is not less than the threshold value, the target is judged as a false alarm target.
[0012] According to the above technical means, a trained XGBoost model is obtained through the time series data of multiple sample units, and the false alarm probability of the target is obtained based on the time series data of the target and the trained XGBoost model, and then the target is judged whether it is a false alarm target according to the false alarm probability of the target and the threshold value. In this technical solution, the judgment result is obtained through the XGBoost model, which does not rely on empirical parameter settings, thereby improving the accuracy of the detection result; in addition, the threshold value is also obtained through the XGBoost model, and there is no need to conduct a large number of tests to obtain the threshold value, thereby reducing the detection cost and difficulty.
[0013] Furthermore, the step of obtaining a trained XGBoost model according to the time series data of multiple sample units includes:
[0014] Collecting time series data of multiple sample units;
[0015] Perform feature engineering on the time series data of the sample units respectively, obtain their corresponding eigenvalues and construct feature vectors;
[0016] Construct samples based on the feature vectors of multiple sample units, and divide the samples into training samples and validation samples according to a certain ratio;
[0017] The training samples and validation samples are used as input, and the labels of samples from false alarm units are set to 1, and the labels of samples from normal units are set to 0. The XGBoost model is trained. If the model accuracy meets the preset requirements, the training is terminated. Otherwise, the model parameters are adjusted and the number of eigenvalues in the eigenvector is increased to continue training until the model accuracy meets the preset requirements.
[0018] Based on the above technical means, an optimal method for training the XGBoost model is provided.
[0019] Furthermore, the plurality of observations include velocity, acceleration, length, width, height, position coordinate X, position coordinate Y and heading angle.
[0020] According to the above technical means, the type of observation quantity is optimized to ensure that accurate observation quantities can be obtained through conventional observation means.
[0021] Furthermore, the feature engineering includes: according to the N historical frame data corresponding to each observation, obtaining the corresponding eigenvalue, the eigenvalue includes at least one of the first-order difference mean, the first-order difference maximum, the first-order difference minimum, the first-order difference standard deviation, the maximum, the minimum, the standard deviation and the mean.
[0022] According to the above technical features, the eigenvalue type is optimized to improve the effectiveness of the eigenvalue.
[0023] Furthermore, the eigenvalues of length, width and height include the first-order difference average, maximum, minimum, standard deviation and mean; the eigenvalues of position coordinate X, position coordinate Y and heading angle include the first-order difference average, first-order difference maximum, first-order difference minimum and first-order difference standard deviation; the eigenvalues of velocity and acceleration include the maximum, minimum, standard deviation, mean, first-order difference average and first-order difference standard deviation.
[0024] According to the above technical means, all eigenvalues of the selected observations are optimized to further improve the effectiveness of the eigenvalues.
[0025] Furthermore, the feature vectors of the target and sample units are composed of all the feature values of several observations arranged in a certain order.
[0026] According to the above technical features, the construction rules of the feature vector are limited to ensure that the feature vectors of each target and sample unit have a unified construction rule.
[0027] Furthermore, the method for determining the threshold value according to the trained XGBoost model includes:
[0028] Select multiple units that are determined to be false alarm targets;
[0029] Collect time series data of multiple false alarm targets and perform feature engineering to obtain their eigenvalues and construct feature vectors;
[0030] Input the feature vectors of multiple false alarm targets into the trained XGBoost model to obtain the corresponding multiple model prediction probabilities;
[0031] The minimum value among the predicted probabilities of multiple models is determined as the threshold value.
[0032] According to the above technical means, a threshold value that does not depend on empirical parameters can be obtained based on the XGBoost model, thereby improving the accuracy of the judgment result.
[0033] Furthermore, the method for determining the threshold value according to the trained XGBoost model includes:
[0034] Select multiple units that are judged to be normal targets;
[0035] Collect time series data of multiple normal targets and perform feature engineering to obtain their eigenvalues and construct feature vectors;
[0036] Input the feature vectors of multiple normal targets into the trained XGBoost model to obtain the corresponding multiple model prediction probabilities;
[0037] The maximum value among the predicted probabilities of multiple models is determined as the threshold value.
[0038] According to the above technical means, a threshold value that does not depend on empirical parameters can be obtained based on the XGBoost model, thereby improving the accuracy of the judgment result.
[0039] Furthermore, the sample unit includes a historical target, and the time series data of the sample unit is obtained from a historical database.
[0040] According to the above technical means, the training process of the XGBoost model does not rely on a large amount of real-time observation data that is difficult to obtain, and can be completed before target detection. The XGBoost model can obtain the expected accuracy through sufficient training time, and the training process of the XGBoost model will not affect the timeliness of target detection and discrimination.
[0041] Furthermore, the number of historical frames N is not less than 5.
[0042] According to the above technical means, the accuracy of the characteristic value is improved.
[0043] Beneficial effects of the present invention:
[0044] (1) The judgment result of the present invention is obtained through the XGBoost model, which does not rely on empirical parameter settings, thereby improving the accuracy of the detection result;
[0045] (2) The threshold value is also obtained through the XGBoost model, which eliminates the need for a large number of tests to obtain the threshold value, thereby reducing the detection cost and difficulty. BRIEF DESCRIPTION OF THE DRAWINGS
[0046] Figure 1 It is a flow chart of the steps of the false alarm target detection method based on the XGBoost algorithm of the present invention;
[0047] Figure 2 The flowchart of the present invention is to train the XGBoost model using the time series data of multiple sample units. DETAILED DESCRIPTION
[0048] The following will describe the embodiments of the present invention with reference to the accompanying drawings and preferred embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are only for illustrating the present invention, not for limiting the scope of protection of the present invention.
[0049] It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and thus the drawings only show components related to the present invention rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.
[0050] Figure 1 The figure is a flow chart of the steps of the false alarm target detection method based on the XGBoost algorithm of the present invention, and the false alarm target detection method based on the XGBoost algorithm comprises the following steps:
[0051] Step S1: Build the initial XGBoost model with the goal of minimizing the loss function;
[0052] Step S2: obtaining a trained XGBoost model according to time series data of a plurality of sample units, wherein the plurality of sample units include at least one false alarm unit and at least one normal unit, and the time series data include observation vectors corresponding to N historical frames, and the observation vectors are composed of a plurality of observation quantities;
[0053] Step S3: Obtaining target time series data;
[0054] Step S4: Perform feature engineering on the target time series data to obtain its feature value and construct a feature vector;
[0055] Step S5: Input the feature vector of the target into the trained XGBoost model to obtain the false alarm probability of the target;
[0056] Step S6: Determine the threshold value of the false alarm probability according to the trained XGBoost model. When the false alarm probability is not less than the threshold value, the target is judged to be a false alarm target.
[0057] First, execute step S1 to build an initial XGBoost model with the goal of minimizing the loss function. The loss function is selected as the Logistic function. The initial XGBoost model includes initial model parameters, which obtain the expected optimal solution or relatively optimal solution after model training. The XGBoost model is an additive model that continuously fits the residual of the base model after the previous iteration in each round of iteration to achieve the purpose of minimizing the loss function, thereby converging to a relatively optimal solution. It has the characteristics of strong interpretability, ability to handle mixed-type features, and robustness to outliers; further, the XGBoost model adds a regularization term to the loss function to reduce the risk of overfitting, and uses the Taylor second-order expansion loss function to make the convergence result more accurate, improve training speed and scalability.
[0058] Then, step S2 is executed to obtain a trained XGBoost model based on the time series data of multiple sample units. Figure 2 The flowchart of the present invention is to train the XGBoost model using the time series data of multiple sample units, and step S2 further includes the following steps:
[0059] S21: Collecting time series data of multiple sample units;
[0060] S22: Perform feature engineering on the time series data of the sample units respectively, obtain the corresponding feature values and construct feature vectors;
[0061] S23: construct samples according to the feature vectors of multiple sample units, and divide the samples into training samples and validation samples according to a certain ratio;
[0062] S24: Take the training samples and the verification samples as input, and set the labels of the samples from the false alarm unit to 1, and the labels of the samples from the normal unit to 0, and train the XGBoost model. If the model accuracy meets the preset requirements, the training is terminated. Otherwise, the model parameters are adjusted and the number of eigenvalues in the eigenvector is increased to continue training until the model accuracy meets the preset requirements.
[0063] In the present embodiment, the sample unit can be a historical target, and the historical target is an object that has been subjected to target detection and has recorded measured values and determined detection results. The time series data of the sample unit can be obtained from a historical database, and the training process of the XGBoost model does not rely on a large amount of real-time observation data that is difficult to obtain, and can be completed before starting target detection. The XGBoost model can obtain the expected accuracy through sufficient training time, and the training process of the XGBoost model will not affect the timeliness of target detection and discrimination. The sample unit can also be a reference unit around the target to be detected, and the time series data of the reference unit needs to be obtained instantly by the sensor. Multiple sample units include at least one false alarm unit and at least one normal unit. The more the number of sample units, the more accurate the result of model training, but the larger the amount of calculation, therefore, the number of sample units can be determined according to the computing power and the demand for the accuracy of the training results. Further, the number ratio of false alarm units to normal units can also be controlled to make the obtained samples more reasonable. Furthermore, the ratio of training samples to validation samples can be controlled, for example, between 9:1 and 5:5, to ensure that the number of training samples meets the model training requirements, while the number of validation samples can verify the model accuracy; further, the ratio of training samples to validation samples is controlled to 8:2.
[0064] In this embodiment, the observed quantities include velocity, acceleration, length, width, height, position coordinate X, position coordinate Y and heading angle. These observed quantities and time series data based on the observed quantities can be easily obtained through sensors. The observed quantities of the sample unit or target constitute its observation vector.
[0065] In this embodiment, the time series data is composed of observation vectors of N historical frames; further, the number of historical frames N is not less than 5 to ensure the accuracy of the eigenvalues; when the number of historical frames N is less than 5, the data is completed by filling the previous frame data or the next frame data. After obtaining the time series data of the sample unit or the target, the time series data of the sample unit or the target is subjected to feature engineering, so that its eigenvalues can be obtained and the eigenvectors can be constructed. Feature engineering includes: according to the N historical frame data corresponding to each observation, the corresponding eigenvalue is obtained, and the eigenvalue includes at least one of the first-order difference mean, the first-order difference maximum, the first-order difference minimum, the first-order difference standard deviation, the maximum, the minimum, the standard deviation and the mean. The type and number of eigenvalues of each observation can be selected as needed to improve the efficiency of feature engineering. As a specific embodiment, the eigenvalues of length, width and height include the first-order difference average, maximum, minimum, standard deviation and mean; the eigenvalues of position coordinate X, position coordinate Y and heading angle include the first-order difference average, first-order difference maximum, first-order difference minimum and first-order difference standard deviation; the eigenvalues of speed and acceleration include the maximum, minimum, standard deviation, mean, first-order difference average and first-order difference standard deviation. After taking all the eigenvalues of several observation vectors as required, all the eigenvalues can be arranged in a certain order to construct the eigenvector corresponding to the sample unit or target.
[0066] In this embodiment, the model learning rate and the number of model iterations are selected to improve the efficiency of the algorithm. For example, the model learning rate is 0.1 and the number of model iterations is 50. To prevent overfitting, the maximum model depth and the subsampling rate of the model are selected, for example, the maximum model depth is 4 and the subsampling rate of the model is 0.8.
[0067] Afterwards, step S3 is executed to obtain the time series data of the target. The time series data of the target can be collected using multiple sensors.
[0068] Next, step S4 is executed to perform feature engineering on the target time series data to obtain its feature value and construct a feature vector. The method of obtaining the feature value and constructing the feature vector in this step is the same as the method in step S2.
[0069] Afterwards, step S5 is executed to input the feature vector of the target into the trained XGBoost model to obtain the false alarm probability of the target.
[0070] Finally, step S6 is performed to determine whether the target is a false alarm target according to the false alarm probability. Specifically, the threshold of the false alarm probability is determined according to the trained XGBoost model; when the false alarm probability is not less than the threshold, the target is determined to be a false alarm target.
[0071] Methods for determining thresholds based on a trained XGBoost model include:
[0072] S61: Select multiple units determined to be false alarm targets;
[0073] S62: collecting the time series data of these false alarm targets, performing feature engineering on the time series data of the false alarm targets respectively, obtaining their feature values and constructing feature vectors;
[0074] S63: Inputting feature vectors of multiple false alarm targets into the trained XGBoost model to obtain their corresponding model prediction probabilities;
[0075] S64: Determine the minimum value among the model prediction probabilities of the multiple false alarm targets as the threshold value.
[0076] The method for determining the threshold value based on the trained XGBoost model may also include:
[0077] Select multiple units that are judged to be normal targets;
[0078] Collect time series data of multiple normal targets and perform feature engineering to obtain their eigenvalues and construct feature vectors;
[0079] Input the feature vectors of multiple normal targets into the trained XGBoost model to obtain the corresponding multiple model prediction probabilities;
[0080] The maximum value among the predicted probabilities of multiple models is determined as the threshold value.
[0081] In step S6, the unit determined as a false alarm target and the unit determined as a normal target can be a historical target, and its time series data can also be directly obtained from a historical database, and the threshold value can be determined by the above method before starting target detection. The determination of the threshold value of the above method does not rely on empirical parameter settings, and the threshold value can be obtained by the XGBoost model based on historical data, which improves the accuracy of the judgment result.
[0082] In summary, the present invention obtains a trained XGBoost model through the time series data of multiple sample units, obtains the false alarm probability of the target based on the time series data of the target and the trained XGBoost model, and then judges whether the target is a false alarm target according to the false alarm probability of the target and the threshold value. The judgment result of the present invention is obtained by the XGBoost model, which does not rely on empirical parameter settings, thereby improving the accuracy of the detection result; in addition, the threshold value is also obtained by the XGBoost model, and there is no need to perform a large number of tests to obtain the threshold value, thereby reducing the detection cost and difficulty.
[0083] The above embodiments are only preferred embodiments for fully illustrating the present invention, and the protection scope of the present invention is not limited thereto. Any equivalent substitution or change made by a person skilled in the art based on the present invention is within the protection scope of the present invention.
Claims
1. A false alarm target detection method based on XGBoost algorithm, characterized in that: Includes steps: Build the initial XGBoost model with the goal of minimizing the loss function; Obtaining a trained XGBoost model according to time series data of a plurality of sample units, wherein the plurality of sample units include at least one false alarm unit and at least one normal unit, the time series data include observation vectors corresponding to N historical frames, and the observation vectors are composed of a plurality of observation quantities; Obtain the time series data of the target; Perform feature engineering on the time series data of the target to obtain its feature value and construct a feature vector; Input the feature vector of the target into the trained XGBoost model to obtain the false alarm probability of the target; The threshold value is determined according to the trained XGBoost model, and when the false alarm probability is not less than the threshold value, the target is judged to be a false alarm target.
2. The false alarm target detection method based on the XGBoost algorithm according to claim 1, characterized in that: The step of obtaining a trained XGBoost model according to the time series data of multiple sample units includes: Collecting time series data of multiple sample units; Performing feature engineering on the time series data of the sample units respectively, obtaining corresponding feature values and constructing feature vectors; Constructing samples according to the feature vectors of the plurality of sample units, and dividing the samples into training samples and verification samples according to a certain ratio; The training samples and the verification samples are used as input, and the labels of the samples from the false alarm unit are set to 1, and the labels of the samples from the normal unit are set to 0, and the XGBoost model is trained. If the model accuracy meets the preset requirements, the training is terminated; otherwise, the model parameters are adjusted and the number of eigenvalues in the eigenvector is increased to continue the training until the model accuracy meets the preset requirements.
3. The false alarm target detection method based on the XGBoost algorithm according to claim 1 or 2, characterized in that, The plurality of observed quantities include velocity, acceleration, length, width, height, position coordinate X, position coordinate Y and heading angle.
4. The false alarm target detection method based on the XGBoost algorithm according to claim 3, characterized in that: The feature engineering includes: according to N historical frame data corresponding to each observation quantity, obtaining the corresponding eigenvalue, and the eigenvalue includes at least one of the first-order difference average value, the first-order difference maximum value, the first-order difference minimum value, the first-order difference standard deviation, the maximum value, the minimum value, the standard deviation and the mean value.
5. The false alarm target detection method based on XGBoost algorithm according to claim 4, characterized in that: The eigenvalues of length, width and height include the first-order difference average, maximum, minimum, standard deviation and mean; the eigenvalues of position coordinate X, position coordinate Y and heading angle include the first-order difference average, first-order difference maximum, first-order difference minimum and first-order difference standard deviation; the eigenvalues of velocity and acceleration include the maximum, minimum, standard deviation, mean, first-order difference average and first-order difference standard deviation.
6. The false alarm target detection method based on the XGBoost algorithm according to claim 1 or 2, characterized in that: The characteristic vectors of the target and the sample unit are both composed of all characteristic values of a number of the observation quantities arranged in a certain order.
7. The false alarm target detection method based on XGBoost algorithm according to claim 1, characterized in that: The method for determining the threshold value according to the trained XGBoost model includes: Select multiple units that are determined to be false alarm targets; Collecting time series data of a plurality of false alarm targets and performing feature engineering to obtain feature values and construct feature vectors; Inputting the feature vectors of the multiple false alarm targets into the trained XGBoost model to obtain the corresponding multiple model prediction probabilities; The minimum value among the multiple model prediction probabilities is determined as the threshold value.
8. The false alarm target detection method based on XGBoost algorithm according to claim 1, characterized in that: The method for determining the threshold value according to the trained XGBoost model includes: Select multiple units that are judged to be normal targets; Collecting time series data of a plurality of normal targets and performing feature engineering to obtain feature values and construct feature vectors; Inputting the feature vectors of the plurality of normal targets into the trained XGBoost model to obtain the corresponding plurality of model prediction probabilities; The maximum value among the multiple model prediction probabilities is determined as the threshold value.
9. The false alarm target detection method based on XGBoost algorithm according to claim 1, characterized in that: The sample unit includes a historical target, and the time series data of the sample unit is obtained from a historical database.
10. The false alarm target detection method based on XGBoost algorithm according to claim 1, characterized in that: The number of historical frames N is not less than 5.