A method and system for identifying targets based on radar

Abstract

The application provides a method and system for identifying targets based on radar, the method comprising: determining potential target signals in a radar echo signal; performing waveform feature extraction on the potential target signals based on a waveform feature extraction algorithm to obtain a waveform feature vector; and classifying and identifying the potential targets based on a support vector machine classification model according to the waveform feature vector. The application combines the waveform feature extraction algorithm and the support vector machine classification model, can effectively improve the use performance of the radar in a complex space with serious multipath interference and clutter influence, and can more accurately classify and identify the potential targets from the multipath interference and background clutter.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of radar technology, and in particular to a method and system for identifying targets based on radar. BACKGROUND

[0002] Wall-penetrating radar can realize the detection, identification, positioning and imaging of targets behind various highway surfaces, building fortresses, leafy grasses and thick smoke that are invisible to the human eye through the good low-frequency penetration characteristics of electromagnetic waves. Wall-penetrating radar based on ultra-wideband technology has high range resolution, strong penetration ability, light weight and strong mobility, and will not cause damage to the human body, and is widely used in military and civilian fields. For example, in fire rescue, wall-penetrating radar emits electromagnetic waves, adopts an ultra-wideband radar non-contact life feature extraction technology, and penetrates a non-metallic medium to irradiate a human body. The emitted electromagnetic waves are modulated by the life features of the human body (human motion, heartbeat and breathing) and reflected back. However, the return signal contains not only the signal of the life features, but also the strong return signal reflected by the non-metallic medium, and the strong interference and noise signals caused by the surrounding environment, which affect the target identification effect. SUMMARY

[0003] The present application provides a method and system for identifying targets based on radar to solve the defect that radar cannot accurately identify targets in a complex environment in the prior art. The present application combines a waveform feature extraction algorithm and a support vector machine classification model to effectively improve the use performance of radar in a complex space with serious multipath interference and clutter influence, and can more accurately classify and identify potential targets from multipath interference and background clutter.

[0004] The present application provides a method for identifying targets based on radar, comprising: determining a potential target signal in a radar return signal; performing waveform feature extraction on the potential target signal based on a waveform feature extraction algorithm to obtain a waveform feature vector; and completing classification and identification of the potential target based on a support vector machine classification model according to the waveform feature vector.

[0005] According to the method for identifying targets based on radar provided by the present application, the waveform feature extraction algorithm is used to perform waveform feature extraction on the potential target signal to obtain a waveform feature vector, which comprises: performing waveform feature extraction on the potential target signal in multiple dimensions based on a waveform feature extraction algorithm to obtain a waveform feature vector in multiple dimensions; and the waveform features in multiple dimensions include pulse centroid, pulse skewness, pulse correlation, pulse full width at half maximum, pulse flatness, pulse spectral skewness, pulse spectral kurtosis, pulse spectral centroid and wavelet decomposition one-layer coefficient mean.

[0006] According to the method for identifying a target based on a radar provided in the application, the classification and identification of a potential target based on the waveform feature vector and a support vector machine classification model comprises: setting target classification labels for the waveform feature vector in multiple dimensions to obtain a sample data set of the support vector machine; pre-processing the sample data set of the support vector machine according to a first preset formula to obtain a pre-processed sample data set; the first preset formula is:

[0007]

[0008] wherein x is a pulse feature, x' is a normalized pulse feature, is a mean value of the pulse feature, and A is a standard deviation of the pulse feature;

[0009] The support vector machine classification model is trained by iteratively solving a hyperplane equation of binary classification based on the pre-processed sample data set; the trained support vector machine model is used to perform classification and identification of a potential target to obtain position information of the potential target.

[0010] According to the method for identifying a target based on a radar provided in the application, after the classification and identification of a potential target based on the waveform feature vector and a support vector machine classification model, a classification method based on trajectory tracking is further used to perform re-classification and identification of the potential target identified by the support vector machine classification model.

[0011] According to the method for identifying a target based on a radar provided in the application, the classification method based on trajectory tracking is used to perform re-classification and identification of the potential target identified by the support vector machine classification model, which comprises: based on the position information of the potential target obtained by the support vector machine classification model, target matching is performed based on a Hungarian matching algorithm to obtain a matching result; based on the matching result and a preset matching threshold, Kalman position prediction and updating of a Kalman coefficient matrix are performed to obtain a result of re-classification and identification.

[0012] According to the method for identifying a target based on a radar provided in the application, the determination of a potential target signal in a radar echo signal comprises: acquiring a radar echo signal; pre-processing the radar echo signal to obtain a pre-processed radar echo signal; detecting the pre-processed radar echo signal to obtain the potential target signal.

[0013] According to the method for identifying a target based on a radar provided in the application, the pre-processing of the radar echo signal to obtain a pre-processed radar echo signal comprises: using a combination of adaptive background suppression and a band-pass filter to perform fixed background noise suppression and clutter elimination processing.

[0014] The method for identifying a target based on a radar provided in the application further includes: performing weak potential target signal enhancement processing on the radar echo signal.

[0015] The method for identifying a target based on a radar provided in the application further includes: performing weak potential target signal enhancement processing on the radar echo signal.

[0016]

[0017] wherein, T0 is a constant false alarm rate threshold, γ0 is a weighting coefficient of the determination threshold, k,j is an index of a cell being determined as a threshold, i is an index of a range line considered by the algorithm, W is a total number of range lines processed, and k',j' is an index of a neighboring cell of the threshold.

[0018] The application further provides a system for identifying a target based on a radar, which includes: a potential target signal determination module, configured to determine a potential target signal in a radar echo signal; a waveform feature extraction module, configured to perform waveform feature extraction on the potential target signal based on a waveform feature extraction algorithm to obtain a waveform feature vector; and a potential target classification and identification module, configured to complete classification and identification of the potential target based on a support vector machine classification model according to the waveform feature vector.

[0019] The application provides a method and a system for identifying a target based on a radar, which includes: determining a potential target signal in a radar echo signal; performing waveform feature extraction on the potential target signal based on a waveform feature extraction algorithm to obtain a waveform feature vector; and completing classification and identification of the potential target based on a support vector machine classification model according to the waveform feature vector. The application combines the waveform feature extraction algorithm and the support vector machine classification model, can effectively improve the use performance of the radar in a complex space where multipath interference and clutter influence are relatively serious, and can more accurately classify and identify the potential target from the multipath interference and the background clutter. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor based on these drawings.

[0021] Figure 1 is one of the flowcharts of the method for identifying a target based on a radar provided in the application.

[0022] Figure 2 is a schematic diagram of the principle of the multipath interference model provided by the present application (a is a scene, and b is a simplified model of multipath interference and target echo);

[0023] Figure 3 is a comparison diagram of the simulation of multipath interference and the actual collected data provided by the present application;

[0024] Figure 4 is a comparison diagram of the 18-dimensional feature distribution obtained from the target signal and the multipath interference provided by the present application;

[0025] Figure 5 is a flowchart of a target classification process based on a support vector machine provided by the present application;

[0026] Figure 6 is a data flow diagram of a target classification process based on a support vector machine provided by the present application;

[0027] Figure 7 is a schematic diagram of a target classification result of a support vector machine provided by the present application;

[0028] Figure 8 is a flowchart of a target classification method based on trajectory tracking provided by the present application;

[0029] Figure 9 is a performance diagram of target classification of a support vector machine and target classification based on trajectory tracking provided by the present application;

[0030] Figure 10 is a flowchart of a method for identifying a target based on a radar provided by the present application;

[0031] Figure 11 is an experimental scene diagram provided by the present application (A is a spacious test factory, B is a narrow corridor, and C is an iron house);

[0032] Figure 12 is a target trajectory diagram determined in a spacious laboratory scene provided by the present application;

[0033] Figure 13 is a target trajectory diagram determined in a narrow corridor scene provided by the present application;

[0034] Figure 14 is a target trajectory diagram determined in an iron house scene provided by the present application;

[0035] Figure 15 is a structural diagram of a system for identifying a target based on a radar provided by the present application. DETAILED DESCRIPTION

[0036] In order to make the objects, technical solutions and advantages of the present application clearer, the following will clearly and completely describe the technical solutions in the present application with reference to the drawings in the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based upon the embodiments in the present application, all other embodiments obtained by those ordinarily skilled in the art without creative effort should fall into the scope of the present application.

[0037] The following will be described with reference to the drawings Figures 1-15 A method and system for identifying targets based on radar are described.

[0038] Reference is made to Figure 1 , Figure 1 One of the flowcharts of a method for identifying targets based on radar provided by the present application is shown.

[0039] Reference is made to Figure 2 , Figure 2 The principle diagram of a multipath interference model (a is a scene, and b is a simplified model of multipath interference and target echo) provided by the present application is shown.

[0040] The present application provides a method for identifying targets based on radar, comprising:

[0041] 101: determining potential target signals in radar echo signals;

[0042] According to the discrete radar reflection cross section and the wall reflection surface, a radar echo model of targets and multipath interference can be established for a SISO IR-UWB (simple input simple output Impulse Radio Ultra Wideband Radar) radar. From a fire rescue scene of detecting trapped personnel using a life detection radar, a diagram of radar receiving different reflection mode signals can be evolved, as shown in Figure 1 .

[0043] Figure 2 (a) is a diagram of a scene of deploying a search and rescue radar, including targets, walls, doors and floors and other components. The radar system is placed at the door to simulate the real situation that firefighters forcibly enter a certain area for search and rescue action. In order to simulate a challenging environment, the scene assumes that there is a hot flow and smoke, so that traditional life detection devices (including thermal imaging, optical imaging and laser radar) cannot effectively detect and locate targets.

[0044] Figure 2(b) is a simplified model of the multi-path interference and target echoes, which uses the concept of discretizing the target into isolated points. The received echoes are divided into three different patterns. The model represents echoes originating from a single scattering point on the target or wall. The model includes echoes from multiple scattering points on the target. The model describes echoes from multiple scattering points on the target, which are reflected and then reflected from the wall or floor. These models effectively capture the characteristics of the observed target, clutter, and multi-path ghosting phenomena in a particular scenario, thereby providing a comprehensive understanding of the radar echo characteristics.

[0045] Echo from a single point on the target: Whether the trapped person is in motion or stationary, the life signals containing breathing and heartbeat are the most powerful evidence of the presence of a person. At the same time, such life signals are more difficult to measure than human motion signals. Therefore, we take breathing and heartbeat as the characteristic signals in the process of establishing the mathematical model of radar echoes.

[0046] Without considering the phase change of the waveform caused by reflection, this set of mathematical models can be considered to express the echo information of a single isolated point on the target under the condition of no interference.

[0047]

[0048] where d(t) is the distance between the single isolated point on the target and the radar. The basic assumption of this formula is that the scattering points caused by breathing and heartbeat move along a sinusoidal trajectory. Where f r , f h , A r and A h are the frequencies and amplitudes of breathing and heartbeat, respectively, and d0 is the distance of the center of sinusoidal motion. Tv is the time delay of the transmitter-target reflection point-receiver, which is actually twice the distance divided by the speed of electromagnetic waves (the speed of light). h(τ, t) is the impulse response function, where a is the response amplitude, similar to the radar cross section, ∑ i a i δ(τ-τ i ) represents the response of all these static objects. s(t) is the transmitted pulse set with reference to the X4M03 radar, where A, f c and B are the transmitted amplitude, center frequency and bandwidth of the radar, respectively.

[0049] This set of equations shows that if the target is considered as an isolated reflection point, the pulse waveform R(τ, t) of the received signal will not be distorted.

[0050] Echo from multiple points of a target: In fact, the model of multiple points of a single target is somewhat similar to the model of multiple targets, except that the distance between multiple targets is particularly close, and is lower than the ranging resolution of ultra-wideband radar, so the received echo of multiple points of a target has a superposition effect.

[0051]

[0052] Compared with equation group (1), equation group (2) is the change of the number of target reflection points at a certain time in the slow time dimension. Specifically, a d n In a certain range, it is randomly changed and added to the distance formula to represent the distance of a certain reflection point of the target. Considering the actual measurement, the number of reflection points Num is also variable in a certain range. The effect of multiple reflection points of the target on the radar echo can be represented by the impulse response equation h(τ, t). Specifically, the response function generated by multiple scattering points is superimposed to form Equation. While the radar transmit waveform s(τ) and the radar calculation formula R(τ, t) remain unchanged. In formula (2), β represents the echo intensity coefficient of the scattering point, which is related to the radar transmission area; and α is the echo coefficient of the fixed object.

[0053] Echo of multipath interference: The reason for multipath ghosting is that the echo generated by the target signal is collected by the receiver after one or more reflections from walls, floors, etc. This part of the signal will produce multiple false targets in the area with only one real target. Of course, in actual measurement, not only multipath ghosting affects the accuracy of radar target recognition. The noise remaining after the radar wave directly reflected from the wall, floor, etc. after background noise suppression will also affect the accuracy of target recognition. The reason for the remaining noise is not necessarily that the algorithm performance for suppressing background noise is poor, but also because there are differences between the radar echo frames returned by the wall, making the pulse phase elimination method not applicable. This phenomenon is very common in handheld radar signal processing that requires fast scene conversion.

[0054] First, the multipath ghosting formula (3) after one reflection can be derived from the multi-point reflection formula of the target.

[0055]

[0056] The physical process of multipath ghosting can be described as: transmitter--one reflection point of the target--one reflection point of the wall--receiver. In general, the whole physical process is: the transmitter sends out a pulse--the target generates multiple reflected pulses--the wall generates more reflected pulses--the pulses return to the receiver. Therefore, the distance from d n (t) to D(t), where d′ n(t) is the sum of the distance from the target reflection point to the wall reflection point and the distance from the wall reflection point to the radar receiver. Correspondingly, the impulse response function is also adjusted to to represent the process that the pulse emitted from one reflection point of the target is reflected by multiple wall reflection points and then received. Wherein, p m is the radar wave reflectivity of the wall, and d′ n (t) is the sum of the target-wall-receiver distance, which can be generally represented by formula (4).

[0057]

[0058] Z1 and Z2 are the impedances of two media (air and concrete in this case). θ ti and θ ir are the incident angle and the transmission angle, respectively. is the vector sum of the incident point to the reflection point, is the normal vector of the reflection surface Ax+By+Cz+D=0. (x t ,y t ,z t ) and (x i ,y i ,z i ) are the spatial coordinates of the incident point (the jth reflection point of the body target) and the ith reflection point, respectively. d ti and d ir are the distance from the reflection point of the human body target to the wall reflection point and the distance from the wall reflection point to the radar receiving point, respectively. (x r ,y r ,z r ) is the spatial coordinates of the radar receiving antenna.

[0059] Please refer to Figure 3 , Figure 3 for the comparison chart of the simulation of multipath interference provided by the present application and the actual collected data.

[0060] T0 represents the complete echo pulse actually collected, T1 represents the transmitted pulse, T2 represents the target pulse, and T3 represents the ghost pulse. The simulation is the simulation signal obtained from the above formula, and measurement 1, measurement 2, and measurement 3 are radar echoes obtained in a spacious laboratory, a narrow corridor, and an iron house environment, respectively. The comparison and analysis of the simulation and the actual collected waveforms show that the direct target echo signal will be distorted, mainly concentrated in the falling edge, and the distortion will increase with the severity of the multipath interference. It is the difference in the degree of distortion between the multipath interference echo waveform and the direct target echo waveform that provides a means for multipath interference suppression based on the waveform.

[0061] 102: Perform waveform feature extraction on the potential target signal based on a waveform feature extraction algorithm to obtain a waveform feature vector;

[0062] As a preferred embodiment, the waveform feature extraction algorithm is used to extract waveform features of the potential target signal to obtain a waveform feature vector, including: using the waveform feature extraction algorithm to extract waveform features of the potential target signal in multiple dimensions to obtain waveform feature vectors in multiple dimensions; the waveform features in multiple dimensions include pulse centroid, pulse skewness, pulse correlation, pulse full width at half maximum, pulse flatness, pulse spectral skewness, pulse spectral kurtosis, pulse spectral centroid, and wavelet decomposition one-layer coefficient mean.

[0063] Specifically, in order to comprehensively analyze the waveform features of the radar echo, so as to more accurately screen out the target echo from the multipath interference and background clutter, 18 pulse feature parameters are selected from the time domain, frequency domain and wavelet domain. The meanings and calculation methods of these parameters are shown in Table 1.

[0064] Table 1 Feature extraction algorithm

[0065]

[0066]

[0067] Table 1 provides a variety of 18 pulse feature extraction algorithms, which meet the complex requirements of extracting target features from radar echoes. In order to illustrate its significance, the data of a walker can be comprehensively analyzed using an X4M03 UWB radar. This radar model provides an efficient method to highlight the feature distribution under different laboratory conditions, thereby widening the range of possibilities for data analysis.

[0068] Please refer to Figure 4 , Figure 4 The comparison chart of the 18-dimensional feature distribution obtained from the target signal and the multipath interference provided by the present application.

[0069] The 18-dimensional features of the potential target in the waveband are obtained using the algorithms provided in Table 1. In this case, the left and right sides are the features of the target and the features of the multipath or clutter, respectively. The three cases, i.e. a wide laboratory, a narrow corridor and an iron room, basically cover various cases where multipath interference and clutter have different degrees of influence on the target. Considering the concentration of the feature distribution and the contrast between the target features and the multipath features, features 2, 4, 6, 7, 8, 10, 11, 14 and 18 can be used as feature vectors to distinguish multipath interference from targets.

[0070] 103: According to the waveform feature vector, the classification and recognition of the potential target are completed based on a support vector machine classification model.

[0071] As a preferred embodiment, according to the waveform feature vector, the classification and identification of the potential target are completed based on a support vector machine classification model, including: setting a target classification label for the waveform feature vector in multiple dimensions to obtain a sample data set of the support vector machine; pre-processing the sample data set of the support vector machine according to a first preset formula to obtain a pre-processed sample data set; the first preset formula is:

[0072]

[0073] Wherein, x is a pulse feature, x' is a normalized pulse feature, is the mean of the pulse feature, and sigma A is the standard deviation of the pulse feature.

[0074] Based on the pre-processed sample data set, a hyperplane equation of binary classification is solved iteratively to train the support vector machine classification model; the trained support vector machine model is used for classification and identification of the potential target to obtain position information of the potential target.

[0075] Specifically, the SVM (support vector machine) is a supervised learning model, which aims to construct a hyperplane or a set of hyperplanes in a high-dimensional or infinite-dimensional space by training a data set with specified labels, so as to separate data points of different categories. In the present application, the training and use of the model are directly using the SVM optimization program of Matlab.

[0076] The SVM is a binary classification model, which is a linear classifier with the largest interval defined in the feature space, and the largest interval makes it different from the perceptron; the SVM also includes a kernel trick, which makes it a substantial nonlinear classifier. The learning strategy of the SVM is to maximize the interval, which can be formalized as a solution to a convex quadratic programming problem, and is equivalent to the minimization problem of a regularized hinge loss function. The learning algorithm of the SVM is an optimization algorithm for solving the convex quadratic programming problem.

[0077] The SVM training is to solve a separation hyperplane that can correctly divide the training data set and has the largest geometric interval. The algorithm includes inputting the training data set into the SVM model, and the SVM model outputs the separation hyperplane and the classification decision function. An appropriate kernel function, such as a Gaussian kernel function, needs to be selected, a penalty parameter needs to be selected, a convex quadratic programming problem needs to be constructed and solved to obtain an optimal solution, and finally the classification decision function is calculated.

[0078] Please refer to Figure 5 , Figure 5 The support vector machine-based target classification process flowchart provided by the present application.

[0079] The training process: radar echo data of different scenes such as corridors, metal rooms, empty laboratories, etc. are collected, the radar echo data is preprocessed including noise reduction, signal enhancement, etc., potential target detection is carried out by using a 1-dimensional mean constant false alarm rate detector, the wave band where the detected potential target is located is extracted, 9 characteristic values of the wave band are calculated, the labeled samples are imported into the support vector machine training model, and the trained model is obtained;

[0080] Real-time measurement: receiving radar real-time echo signal, carrying out noise reduction and signal enhancement preprocessing, detecting potential target signal by using a 1-dimensional mean constant false alarm detector, extracting the wave band of the target signal, calculating 9 characteristic values of the wave band, inputting the calculated 9 characteristic values into the support vector machine classifier, analyzing the potential target signal based on the training model, and identifying 'false target' or 'true target'.

[0081] Please refer to Figure 6 , Figure 6 The target classification process data flow diagram based on the support vector machine provided by the application.

[0082] The overall situation of the data training step is that the measured radar data of three scenes, an empty room S1, a narrow concrete corridor S2 and an iron room S3, are used as data input, a 1-dimensional mean constant false alarm detector is used to detect potential targets, echo signal samples are collected, and then the echo signal samples are screened, the 'true target' samples, i.e. positive samples, and the 'false target' samples, i.e. negative samples, are screened; further, the waveforms of the potential targets are extracted from the determined potential target echo signal samples, the waveforms of the potential targets are used as samples, nine time-frequency features of the waveforms are extracted as pulse features to form a feature vector; generally, a feature vector composed of nine characteristic values can be obtained for each sample. These feature vectors are used as the basis for classifying potential target signals and are sent to the SVM classifier for model training.

[0083] Please refer to Figure 7 , Figure 7 The target classification result diagram of the support vector machine provided by the application.

[0084] In order to analyze the influence of data generated by different scenes on the classification performance, the waveform data in three training scenes are trained according to three steps; the three training scenes are an empty room sample scene, an empty room sample and a narrow corridor sample mixed scene, and an empty room sample, a narrow corridor sample and an iron room sample mixed scene. The number of training samples in the above three cases is 5336, 9180 and 12464 respectively, and the number of target samples and ghost samples is equal. Figure 7(a) represents an empty room sample scene, (b) represents a mixed scene of a narrow corridor sample and an iron room sample, and (c) represents a mixed scene of an empty room sample and a narrow corridor sample and an empty room sample scene.

[0085] Figure 7 It is shown that as the severity of clutter and multipath interference increases, the performance of the trained model also decreases. In view of the fact that in fire rescue, false negatives are more deadly than false positives, the cost function is adjusted during training, so that the cost of false negatives is three times that of false positives;

[0086] Figure 7 (a) shows that based on the data collected from the spacious room scene less affected by multipath interference, the trained model can achieve an overall validation accuracy of 91.3%. When testing the model using the reserved 10% test data, the test accuracy can reach 90.1%, the false negative rate is 4.4%, and the false positive rate is 13.1%.

[0087] Figure 7 (b) shows that the mixture of samples from spacious rooms and samples from narrow concrete corridors is more severely affected by multipath interference and clutter, resulting in slightly poorer performance of the trained model. The overall classification accuracy can reach 84.2%, when using a 10% test data set to test the model, the test accuracy can reach 83.5%, the false negative rate is 12.9%, and the false positive rate is 18.8%.

[0088] Figure 7 (c) shows the training results of sample data containing three scenes, which perform the worst. The overall classification accuracy can reach 79.1%, when using a 10% test data set to test the model, the test accuracy can reach 78.2%, the false negative rate is 19.3%, and the false positive rate is 22.5%. This is because the radar target signal in the iron room is most severely interfered by multipath interference and clutter.

[0089] As a preferred embodiment, after the classification and identification of the potential target based on the waveform feature vector are completed based on the support vector machine classification model, the method further comprises: performing re-classification and identification of the potential target identified by the support vector machine classification model based on a trajectory tracking classification method.

[0090] As a preferred embodiment, the re-classification and identification of the potential target identified by the support vector machine classification model based on the trajectory tracking classification method comprises: performing target matching based on the Hungarian matching algorithm according to the position information of the potential target obtained by the support vector machine classification model to obtain a matching result; and performing Kalman position prediction and updating of a Kalman coefficient matrix according to the matching result and a preset matching threshold to obtain a re-classification and identification result.

[0091] Generally, the SVM-based target classification model can remove most of the multipath ghosting and clutter, but in some harsh environments, such as metal spaces with more clutter, the ability of the classification method based on waveform characteristics will be inhibited, so that some ghosting and clutter are mixed into the target signal. However, these mixed targets have defects in time continuity, that is, the uncertainty of the target reflection point causes the change of the echo path, but the path change of the direct echo is smaller than that of the wall twice reflected echo, resulting in insufficient continuity of the ghosting. Based on this, the embodiment further adopts the method of trajectory tracking forward, taking the target trajectory as a feature parameter of target recognition, so as to further improve the accuracy of target recognition.

[0092] Reference is made to Figure 8 , Figure 8 The flowchart of the target classification method based on trajectory tracking provided by the present application is shown.

[0093] SORT (Simple Online and Realtime Tracking) algorithm is a target tracking calculation method, the core of which is Kalman filter algorithm and Hungarian algorithm. The main function of Kalman filter algorithm is to initialize the position of each object of Kalman filter by the result obtained by the first target detection network, and then to predict the position of each object in the next frame by using the position of each object in the current frame. The Hungarian algorithm is mainly used to associate the predicted position with the detection result in the current frame; the predicted position and the related detection result are then used to update the state of the Kalman filter, so as to achieve the effect of tracking. There is an assignment problem between the detection and tracking results in the SORT algorithm, and the Hungarian algorithm is used to calculate an optimal assignment between the detection results of the detector and the tracking trajectories of the tracker, and the cost of the assignment is the minimum. The SORT algorithm uses a weighted Hungarian algorithm to associate the tracking targets frame by frame. The weight of the algorithm is the IOU distance.

[0094] The target tracking method is used to further identify potential targets and improve detection accuracy. The target screening process based on target tracking includes:

[0095] The position information of the potential target obtained by the support vector machine classification model is input to the IOU (Intersection-over-Union) for target matching, and is divided into three categories according to the matching threshold: an unmatched label, an unmatched point and a matched label;

[0096] The minimum visibility threshold and the target age threshold parameter are added as the basis for rejecting unmatched labels and new tracking, so as to determine the tracking; the unmatched label will be deleted according to the target age threshold; the unmatched point will be a new label according to the minimum visibility threshold; and the matched label indicates that the tracking is successful between the two frames.

[0097] Kalman position prediction and Kalman state transition matrix update are performed;

[0098] The potential target of the track is determined to be a real target, i.e., the track is a real target;

[0099] The result is output and recycled, and the potential target identification is continuously performed.

[0100] Reference is made to Figure 9 , Figure 9 The performance chart of the target classification of the support vector machine and the target classification based on track tracking provided by the present application.

[0101] In order to fully study the performance of the classification method based on the support vector machine and the classification method based on track tracking, data of three detection scenes are used for processing, i.e., a moving target S1 in a spacious room, a stationary target S2 in a narrow cement corridor and a stationary target S3 in an iron house. Through calculation, the original target position and track image, the first-level classification image (based on the SVM classifier) and the second-level classification image (based on the track tracking classifier) are obtained.

[0102] Figure 9 In Fig. A, the initial target track image is affected by a large amount of clutter and multi-path ghost accompanying the target position change; after the first identification of the potential target based on the support vector machine, most of the clutter can be effectively suppressed, but some multi-path ghosts still remain in the image, as shown in Fig. B; finally, after the second identification of the target through track tracking, the multi-path ghost is also well suppressed, as shown in Fig. C. Figure 9 Figure 9

[0103] Reference is made to Figure 10 , Figure 10 The flowchart of the method for identifying a target based on a radar provided by the present application.

[0104] The main flow of the radar target identification designed by the present application includes six parts: stationary background noise suppression, target signal enhancement, initial peak positioning, feature extraction, first-level target identification based on the SVM and second-level target identification based on track tracking. The above content introduces the feature extraction, the first-level target screening and the second-level target screening method. Then, the three parts of the stationary noise suppression, the weak signal enhancement and the preliminary peak positioning are mainly discussed.

[0105] As a preferred embodiment, the potential target signal in the radar echo signal is determined, including: acquiring the radar echo signal; pre-processing the radar echo signal to obtain a pre-processed radar echo signal; detecting the pre-processed radar echo signal to obtain the potential target signal.

[0106] ​​As a preferred embodiment, the radar echo signal is preprocessed to obtain a preprocessed radar echo signal, comprising: adopting a combination of adaptive background suppression and a band-pass filter to perform fixed background noise suppression and clutter elimination processing.

[0107] Specifically, target signals, Gaussian noise, time reference unstable noise, fixed background noise, and non-fixed object clutter exist in the received radar signals. The fixed background noise greatly contributes to improving target detection accuracy. The application adopts a combination of adaptive background suppression and a band-pass filter to suppress the fixed background noise. The stray signal ratio a is set to an adaptive value, and the value of a determines the dependence of environmental changes. In detail, a larger value of a is suitable for a fixed radar scene and cannot significantly remove the echoes of slightly changed objects in the scene, while a smaller value of a is suitable for a scene with a more dramatic change in the environment, such as a moving radar, and can also remove the echoes of small moving objects in the environment, but the disadvantage is that it causes attenuation of the target signal.

[0108]

[0109] After a certain measurement interval (such as 1 minute), five frames of pulse echoes are used to calculate four correlation coefficients, x i+1,s and x i,s are the amplitudes of two adjacent pulses of each point s, and are their average values), and the average value Because a higher correlation coefficient indicates a higher similarity of the scene, the change in the echo signal produced by the fixed object is weaker. In this case, the larger the value of a, the more obvious the fixed clutter in the scene, and the more significant the signal of the moving object after removal. Conversely, the smaller the value of a, the stronger the ability to remove non-target clutter or multipath interference signals, but the weaker the target echo signal. The band-pass filter can be a 160-order Hanning filter.

[0110] As a preferred embodiment, the radar echo signal is preprocessed to obtain a preprocessed radar echo signal, further comprising: performing weak potential target signal enhancement processing.

[0111] In order to solve the problem that weak targets in the radar echo signal are covered, signal enhancement processing can be performed on the weak potential target signal. For example, time gain method, automatic gain control, and early normalization.

[0112] As a preferred embodiment, the preprocessed radar echo signal is detected to obtain a potential target signal, comprising: adopting a one-dimensional constant false alarm rate detector to detect the preprocessed radar echo signal, and obtaining the potential target signal according to a second preset formula; the second preset formula is:

[0113]

[0114] wherein T0 is a constant false alarm rate threshold, γ0 is a weighting coefficient for determining the threshold, k,j is an index of the cell being determined for the threshold, i is an index of the range line considered by the algorithm, W is the total number of range lines processed, and k',j' is an index of the adjacent cell for which the threshold is calculated.

[0115] Considering the accuracy of the initial peak screening and the computing power required by the algorithm, 1D-CFAR (Constant False Alarm Rate Detector) is selected as the pretreatment program for the initial position locking. The core formula of the algorithm is the second preset formula. In order to prevent the problem of small targets being covered in the multi-target detection process, through the analysis of some experimental data, the optimal parameters obtained are set as follows: the weighting coefficient γ is 2.9, the threshold determination unit k,j is 30, and the adjacent unit k',j' is 40.

[0116] The algorithm verification experiment is as follows:

[0117] Please refer to Figure 11 , Figure 11 The experimental scene provided by the present application (A is a spacious test factory, B is a narrow corridor, and C is an iron house).

[0118] Construction of experimental scene: the scene faced in fire rescue is very complex, which may include a room-dense office building, a spacious factory, a narrow corridor and an iron house with dense offices. Some scenes have serious multipath interference and clutter effects, which pose a great challenge to the weak life perception of fire rescue. In order to verify the effectiveness and reliability of the algorithm designed in the above-mentioned environment, three scenes of spacious laboratory, narrow corridor and iron house are selected for experimental test.

[0119] In addition, the radar used by the present application is Novalda X4M03, the center frequency of which is 8.748 GHz, the bandwidth is 2.95 GHz, the fixed measurement range is 0.4-9 meters, and the pulse repetition rate is 20 Hz.

[0120] Performance: for the three types of scenes: spacious laboratory, narrow corridor and iron house, different numbers of single and multiple targets and different states of stillness and movement are tested for each type of scene.

[0121] (1) spacious laboratory

[0122] Please refer to Figure 12 , Figure 12 The target trajectory determined by the present application in the spacious laboratory scene is provided.

[0123] To study the robustness and reliability of the designed algorithm under different measurement conditions, three scenes were set up for target detection in an open laboratory scene: a single person walking (as shown in A1 of Figure 12 ), two people walking (as shown in A2 of Figure 12 ), and two people standing still (as shown in A3 of Figure 12 ). After radar signal preprocessing, target retrieval, target primary classification based on the SVM trained model, and target secondary classification based on the tracking algorithm, the target trajectory map can be obtained. Due to the low level of clutter and multipath interference (mainly caused by ground reflection) in the inherent open laboratory, the accuracy of target detection is high. The vital sign signals can also be dynamically extracted according to the real-time trajectory data.

[0124] (2) Long and narrow corridor

[0125] Please refer to Figure 13 , Figure 13 the target trajectory map determined under the long and narrow corridor scene provided by the present application.

[0126] Four groups of experiments were set up in a narrow corridor test scene, including one blank scene (as shown in B1 of Figure 13 ), two people crossing and moving (as shown in B2 of Figure 13 ), one person standing still (as shown in B3 of Figure 13 ), and one person lying on one side and squatting (as shown in B4 of Figure 13 ). Without targets, false positives can be avoided with a high probability, and for those existing targets, their trajectories are also more complete.

[0127] (3) Iron house

[0128] Please refer to Figure 14 , Figure 14 the target trajectory map determined under the iron house scene provided by the present application.

[0129] The iron house produces the most clutter and multipath ghosting, which is the most serious scene for target detection interference. The algorithm verification experiment sets up four experiments: one blank scene (as shown in C1 of Figure 14 ), two people walking across (as shown in C2 of Figure 14 ), two people standing still (as shown in C3 of Figure 14 ), and one moving radar searching for one moving person (as shown in C4 of Figure 14 ). Whether it is a stationary target or a moving target, whether it is a single target or multiple targets, the proposed algorithm is basically effective in identifying targets from clutter and multipath interference and effectively tracking their trajectories. However, due to the severity of multipath interference and clutter influence, false positives and false negatives cannot be completely eliminated.

[0130] In general, Figures 12-14 The robustness and reliability of the proposed algorithm are verified from three scenarios and two states. The comparison of A, B and C shows that the radar echo signal is significantly affected by the environment, but the proposed algorithm can effectively suppress the influence of environmental clutter, system thermal noise, etc., and can largely suppress the interference of multipath ghosting. The comparison of A2 and A3, B2 and B3, and C2 and C3 shows that the moving target is easier to identify than the stationary target. When observing the automatically generated anti-cancellation parameter a, it is found that the value of a in the moving scene is lower, allowing the anti-cancellation to suppress most of the stationary clutter and highlight the features of the moving target. The reason is that the motion amplitude of the moving target is greater than the amplitude of the breathing and heartbeat, making them more distinguishable from the echoes of stationary objects. It is particularly noted that the data of C4 is collected when a handheld radar is searching for a moving person (irregular motion of the radar). The processing result shows that the algorithm proposed in the present application can also cope with this situation, effectively identifying the target and tracking its trajectory.

[0131] The radar-based target identification method provided by the present application can effectively suppress environmental clutter interference, overcome the influence of multipath ghosting on target identification, and can be applied to radar with different numbers of targets and different motion states to effectively identify targets, and is suitable for the application requirements of a weak life sensing radar system in a complex fire rescue environment.

[0132] Reference is made to Figure 15 , Figure 15 A structural schematic diagram of a radar-based target identification system provided by the present application.

[0133] The present application also provides a radar-based target identification system, comprising: a potential target signal determination module 1501 for determining potential target signals in a radar echo signal; a waveform feature extraction module 1502 for performing waveform feature extraction on the potential target signals based on a waveform feature extraction algorithm to obtain a waveform feature vector; and a potential target classification and identification module 1503 for completing classification and identification of the potential target based on a support vector machine classification model according to the waveform feature vector.

[0134] For the radar-based target identification system provided by the present application, please refer to the above method embodiments, which will not be described here again.

[0135] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method of identifying a target based on radar, characterized in that, The method comprises the steps of: determining a potential target signal in a radar echo signal; the radar is a single-input single-output pulse ultra-wideband through-wall radar, which is applied to a fire rescue scene with multipath interference; the potential target signal is a life signal of a stationary human body breathing and heartbeat; waveform feature extraction is performed on the potential target signal based on a waveform feature extraction algorithm to obtain a waveform feature vector; classification and identification of the potential target are completed based on a support vector machine classification model according to the waveform feature vector; after the classification and identification of the potential target based on the support vector machine classification model according to the waveform feature vector, the method further comprises the steps of: performing re-classification and identification of the potential target identified by the support vector machine classification model based on a trajectory tracking classification method; the re-classification and identification of the potential target identified by the support vector machine classification model based on the trajectory tracking classification method comprises the steps of: target matching is performed based on a Hungarian matching algorithm according to position information of the potential target obtained by the support vector machine classification model to obtain a matching result; Kalman position prediction and updating of a Kalman coefficient matrix are performed according to the matching result and a preset matching threshold to obtain a re-classification and identification result; the method of determining a potential target signal in a radar echo signal comprises the steps of: obtaining a radar echo signal; preprocessing the radar echo signal to obtain a preprocessed radar echo signal; detecting the preprocessed radar echo signal to obtain the potential target signal; the preprocessing of the radar echo signal to obtain the preprocessed radar echo signal comprises the steps of: fixed background noise suppression and clutter elimination processing are performed by combining adaptive background suppression and a band-pass filter; the preprocessing of the radar echo signal to obtain the preprocessed radar echo signal further comprises the step of: performing weak potential target signal enhancement processing; the detection of the preprocessed radar echo signal to obtain the potential target signal comprises the step of: a one-dimensional constant false alarm rate detector is used to detect the preprocessed radar echo signal, and the potential target signal is obtained according to a second preset formula; the second preset formula is: ； wherein, is a constant false positive rate threshold, is a weighting factor for the determined threshold, is an index of the cell for which the threshold is being determined, i is an index of the range line under consideration by the algorithm, W is a total number of range lines processed, is an index of the adjacent cell for which the threshold is being calculated.

2. The method of claim 1, wherein, the waveform feature extraction of the potential target signal based on the waveform feature extraction algorithm to obtain a waveform feature vector comprises the steps of: waveform feature extraction of multiple dimensions is performed on the potential target signal based on the waveform feature extraction algorithm to obtain a waveform feature vector of multiple dimensions; the waveform features of multiple dimensions include pulse centroid, pulse skewness, pulse correlation, pulse full width at half maximum, pulse flatness, pulse spectral skewness, pulse spectral kurtosis, pulse spectral centroid and wavelet decomposition one-layer coefficient mean.

3. The method of claim 2, wherein, the classification and identification of the potential target based on the support vector machine classification model according to the waveform feature vector comprises the steps of: a target classification label is set for the waveform feature vector of multiple dimensions to obtain a sample data set of a support vector machine; the sample data set of the support vector machine is preprocessed according to a first preset formula to obtain a preprocessed sample data set; the first preset formula is: ； wherein, is the pulse feature, is the normalized pulse feature, is the mean of the pulse feature, is the standard deviation of the pulse feature; Based on the pre-processed sample data set, a hyperplane equation of binary classification is solved iteratively to train the support vector machine classification model; The trained support vector machine classification model is used for classification and identification of potential targets to obtain position information of the potential targets.

4. A system for identifying targets based on radar, characterized in that, Comprise: A potential target signal determination module for determining potential target signals in radar echo signals; The radar is a single-input single-output pulse ultra-wideband through-wall radar, and is applied to a fire rescue scene with multipath interference; the potential target signal is a life signal of static human body breathing and heartbeat; A waveform feature extraction module for performing waveform feature extraction on the potential target signals based on a waveform feature extraction algorithm to obtain a waveform feature vector; A potential target classification and identification module for completing classification and identification of potential targets based on a support vector machine classification model according to the waveform feature vector; After the classification and identification of potential targets based on the support vector machine classification model according to the waveform feature vector, further comprising: A classification method based on trajectory tracking is used to re-classify and identify the potential targets identified by the support vector machine classification model; The classification method based on trajectory tracking for re-classifying and identifying the potential targets identified by the support vector machine classification model comprises: Based on the position information of the potential targets obtained by the support vector machine classification model, a target matching is performed based on a Hungarian matching algorithm to obtain a matching result; According to the matching result and a preset matching threshold, a Kalman position prediction and updating of a Kalman coefficient matrix are performed to obtain a re-classification result; The determination of the potential target signals in the radar echo signals comprises: Obtaining a radar echo signal; Preprocessing the radar echo signal to obtain a pre-processed radar echo signal; Detecting the pre-processed radar echo signal to obtain the potential target signals; The preprocessing of the radar echo signal to obtain the pre-processed radar echo signal comprises: A combination of adaptive background suppression and a band-pass filter is used for fixed background noise suppression and clutter elimination processing; The preprocessing of the radar echo signal to obtain the pre-processed radar echo signal further comprises: Weak potential target signal enhancement processing is performed; The detection of the pre-processed radar echo signal to obtain the potential target signals comprises: A one-dimensional constant false alarm rate detector is used to detect the pre-processed radar echo signal, and the potential target signals are obtained according to a second preset formula; The second preset formula is: ； wherein, is a constant false positive rate threshold, is a weighting factor for the determined threshold, is an index of the cell for which the threshold is being determined, i is an index of the range line under consideration by the algorithm, W is the total number of range lines processed, is an index of the adjacent cell for which the threshold is being calculated.

Images