A training method of a trajectory observation data denoising model for a motorized target

Abstract

The embodiment of the present application provides a kind of training method for trajectory observation data denoising model of motor target, the model includes preprocessing unit and denoising autoencoder, the method comprises: obtaining training set, each sample in training set includes input data and label, input data is the multiple frames of observation data obtained by sensor in predetermined time interval observing target movement track, label is the real movement track of target in corresponding time;Model is trained using training set, and the parameters of model are updated based on the loss calculated, wherein, preprocessing unit carries out denoising preprocessing to input data, and denoising autoencoder carries out coding and decoding to input data after denoising preprocessing, and the movement track after denoising is obtained based on the output of coding and decoding of denoising autoencoder Model, the model capable of effectively denoising observation data is obtained by the training method of the present application, to improve the accuracy of model to motor target movement track estimation.

Description

Technical Field

[0001] This invention relates to the fields of spatiotemporal data analysis and target tracking technology, and more specifically, to a training method for a denoising model of trajectory observation data for maneuvering targets. Background Technology

[0002] Target tracking refers to calculating the true motion of a target from its observation data, playing a crucial role in surveillance, transportation, and national defense. However, target observation data often contains interference such as missed detections, false alarms, and noise. Target tracking algorithms need to calculate the optimal estimate of the target's true motion state from all observation data. Existing target tracking algorithms often model the target's motion using a known motion model (such as a uniform motion model), predict the target's motion state, and then use optimal estimation filtering techniques (such as Kalman filtering) to fuse the observed data with the predicted state to calculate the estimated target motion state. However, for highly maneuverable targets (i.e., mobile targets), modeling their motion state using a single known motion model is often very difficult. Optimization is needed to address the complexity of mobile target motion. Representative existing methods include Interactive Multiple Model (IMM), Hybrid Grid Multiple Model (HGMM), and methods based on Long Short-Term Memory (LSTM) networks.

[0003] Taking the interactive multi-model method in reference [1] as an example, this method uses a set of known motion models, including constant velocity model (CV), constant acceleration model (CA), constant turning rate model (CT), etc., corresponding to different deterministic motion laws such as uniform linear motion, uniformly accelerated linear motion, and uniform circular motion. Then, during the target tracking process, the model probability is updated, that is, the weight of each model is calculated at the current moment to fit the motion law of the maneuvering target. Taking the hybrid grid multi-model method in reference [2] as an example, this method sets two sets of known motion models, which are divided into coarse grid and fine grid according to the complexity of the motion models. The motion law of the maneuvering target is fitted by the update combination of the two sets of motion models, which is a further extension of the interactive multi-model method. The principle of the above target tracking methods is to fit the motion law of the target by a linear combination of a set of known motion models. For maneuvering targets, the motion law of the target is complex and difficult to fit with a set of known motion models.

[0004] Taking the maneuvering target tracking method in reference [3] as an example, this method uses a long short-term memory network to learn the motion law of the maneuvering target from the trajectory dataset in order to fit the motion law of the maneuvering target. This method needs to be combined with the Bayesian filtering method, and can only extract the information of the previous data. Moreover, the motion law of the maneuvering target is complex, and it is impossible to accurately estimate the motion trajectory of the maneuvering target, which leads to poor tracking effect of the maneuvering target.

[0005] Therefore, for maneuvering targets, the motion patterns of the targets are complex. Existing methods for modeling the motion trajectory of maneuvering targets using known motion models have great difficulties. On the other hand, methods that use neural networks to learn the motion patterns of maneuvering targets to predict their motion trajectories have problems such as the complexity and variability of the motion patterns of maneuvering targets, which makes it difficult for neural networks to learn their motion patterns well, resulting in low accuracy of the predicted trajectory data and poor tracking performance.

[0006] References:

[0007] [1]XU L, LI XR, DUAN Z.Hybrid grid multiple-model estimation with application to maneuvering target tracking[J / OL].IEEE Transactions onAerospace and Electronic Systems,2016,52(1):122-136.DOI:10.1109 / TAES.2015.140423.

[0008] [2] MAZOR E, AVERBUCH A, BAR-SHALOM Y, et al. Interacting multiple model methods in target tracking: a survey [J / OL]. IEEE Transactions on Aerospace and Electronic Systems, 1998, 34(1): 103-123. DOI: 10.1109 / 7.640267.

[0009] [3] GAO C, YAN J, ZHOU S, et al. Long short-term memory-based deep recurrent neural networks for target tracking[J / OL]. Information Sciences, 2019, 502: 279-296. DOI: 10.1016 / j.ins.2019.06.039. SUMMARY

[0010] Therefore, the purpose of the present application is to overcome the defects of the prior art, and to provide a training method for a trajectory observation data denoising model of a maneuvering target.

[0011] The purpose of the present application is achieved by the following technical solutions:

[0012] According to a first aspect of the present application, a training method for a trajectory observation data denoising model of a maneuvering target is provided, the model comprising a preprocessing unit and a denoising autoencoder, the method comprising: obtaining a training set, each sample in the training set comprising input data and a label, the input data being a plurality of frames of observation data obtained by a sensor observing the target motion trajectory within a predetermined time interval, and the label being the true motion trajectory of the target within the corresponding time; training the model using the training set, and updating the parameters of the model based on the calculated loss, wherein the preprocessing unit performs denoising preprocessing on the input data, the denoising autoencoder performs encoding and decoding on the input data after denoising preprocessing, and the model obtains the denoised motion trajectory based on the output of the denoising autoencoder encoding and decoding.

[0013] In some embodiments of the present application, each frame of observation data comprises a plurality of observation point data observed by the sensor in the corresponding frame, and the preprocessing unit comprises a learnable weight parameter matrix for fusion processing of each predetermined number of observation point data in each frame of observation data, wherein during training, the parameters of the learnable weight parameter matrix of the denoising autoencoder and the preprocessing module are updated based on the loss.

[0014] In some embodiments of the present application, the preprocessing unit uses a predetermined sliding window to read a predetermined number of observation point data from each frame of observation data each time, and uses a learnable weight parameter matrix to perform fusion processing on the observation point data read each time, wherein the predetermined number is greater than or equal to 2.

[0015] In some embodiments of the present application, the observation point data observed in the corresponding frame includes the velocity of the target in the corresponding frame, the acceleration of the target in the corresponding frame, the longitude position of the target in the corresponding frame, and the latitude position of the target in the corresponding frame.

[0016] In some embodiments of the present application, the denoising autoencoder comprises an encoder and a decoder, wherein the encoder is configured to encode the input data after denoising preprocessing to obtain a hidden layer feature vector; and the decoder is configured to decode the hidden layer feature vector to obtain an estimated value of the target motion state, which is the denoised motion trajectory; wherein during training, the loss calculated based on the denoised motion trajectory and the real motion trajectory is used to update the learnable weight parameter matrix, the parameters of the encoder and the decoder.

[0017] In some embodiments of the present application, the denoising autoencoder comprises an encoder and a decoder, and the model further comprises an activation function layer, which is a Maxout activation function layer, wherein the encoder is configured to encode the input data after denoising preprocessing to obtain a hidden layer feature vector; the decoder is configured to decode the hidden layer feature vector to obtain an estimated value of the target motion state; and the Maxout activation function layer is configured to process the estimated value of the target motion state according to the Maxout activation function to obtain the denoised motion trajectory; wherein during training, the loss calculated based on the denoised motion trajectory and the real motion trajectory is used to update the learnable weight parameter matrix, the parameters of the encoder, the decoder and the Maxout activation function layer.

[0018] According to a second aspect of the present application, a method for tracking a maneuvering target is provided, which is used in the fields of security monitoring, intelligent transportation or radar early warning, and the method comprises: using a trajectory observation data denoising model trained based on any one of the methods according to the first aspect of the present application, to obtain a denoised motion trajectory of the target according to the input continuous multiple frames of observation data of the target motion trajectory; and tracking the maneuvering target according to the denoised motion trajectory.

[0019] According to a third aspect of the present application, a method for tracking a maneuvering target is provided, which is used in the fields of security monitoring, intelligent transportation or radar early warning, and the method comprises: dividing multiple input data in the continuous multiple frames of observation data according to a time series data format, wherein each input data comprises a predetermined number of frames of observation data, and adjacent two input data comprise repeated multiple frames of observation data; using a trajectory observation data denoising model trained based on any one of the methods according to the first aspect of the present application, to obtain a denoised motion trajectory of the target according to each input data; and using a trajectory reconstruction module to average two denoised motion trajectories corresponding to the repeated multiple frames of observation data in adjacent two input data to obtain a reconstructed motion trajectory; and tracking the maneuvering target according to the reconstructed motion trajectory.

[0020] According to a fourth aspect of the present application, there is provided an electronic device comprising: one or more processors; and a memory, wherein the memory is configured to store executable instructions; and the one or more processors are configured to implement the steps of the method of any one of the first, second and third aspects of the present application via execution of the executable instructions.

[0021] Compared with the prior art, the present application has the advantages that:

[0022] In the training method of the present application, first, the input data is preprocessed by a preprocessing unit to effectively reduce the influence of noise data and abnormal data, and second, the noise in the input data after preprocessing is removed by a denoising autoencoder to further effectively reduce the noise data of the observation data. The model obtained by the training method has the ability to effectively denoise the observation data, greatly improving the accuracy of the model in estimating the motion trajectory of the maneuvering target. BRIEF DESCRIPTION OF DRAWINGS

[0023] The embodiments of the present application will be further described below with reference to the accompanying drawings, in which:

[0024] Figure 1 Fig. 1 is a schematic diagram of the structure principle of a trajectory observation data denoising model according to an embodiment of the present application;

[0025] Figure 2 Fig. 2 is a schematic diagram of a training method of a trajectory observation data denoising model for a maneuvering target according to an embodiment of the present application;

[0026] Figure 3 Fig. 3 is a schematic diagram of the generation of input data and the processing process principle of the model on the input data according to an embodiment of the present application;

[0027] Figure 4 Fig. 4 is a schematic diagram of generating a motion trajectory of a maneuvering target according to an embodiment of the present application. DETAILED DESCRIPTION

[0028] In order to make the objects, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0029] As mentioned in the background section, due to the complexity of the motion law of the maneuvering target, the existing method of modeling the motion trajectory of the maneuvering target using a known motion model has great difficulties, and the method of using a neural network to learn the motion law of the maneuvering target to predict the motion trajectory has problems such as low accuracy of the predicted trajectory data and poor tracking effect due to the complexity and variability of the motion law of the maneuvering target, which makes it difficult for the neural network to learn the motion law well.

[0030] Based on the above problems, the present application provides a training method for a trajectory observation data denoising model of a maneuvering target, the model comprising a preprocessing unit and a denoising autoencoder (DAE), in which: first, a training set is obtained, and the input data of each sample in the training set is a plurality of frames of observation data obtained by a sensor observing the target motion trajectory within a predetermined time interval; second, the model is trained using the training set, and the parameters of the model are updated based on the calculated loss, wherein the preprocessing unit performs denoising preprocessing on the input data, the observation data of the sensor is preliminarily denoised through the preprocessing unit, effectively reducing the influence of noise data and abnormal data, the denoising autoencoder encodes and decodes the input data after denoising preprocessing, the model obtains the denoised motion trajectory based on the output after encoding and decoding of the denoising autoencoder, and the denoising autoencoder removes the noise in the denoised observation data, further effectively reducing the noise data of the observation data; finally, the model parameters are updated in a training mode based on the loss calculated from the motion trajectory output by the model and the real motion trajectory, so that the model has the ability to effectively denoise the observation data, greatly improving the accuracy of the model in estimating the motion trajectory of the maneuvering target, and improving the target tracking accuracy and tracking efficiency.

[0031] In order to better understand the present application, the present application is described in detail below from four aspects of training set, model structure, model training process and application scenario in combination with the drawings and specific embodiments.

[0032] I. Training set

[0033] When the sensor observes the target motion trajectory, although there is only one target to be observed, there are various interferences and false signals, resulting in error or abnormal noise data in each frame of observation data given, and false alarm data obtained by determining the signal generated by non-target as target signal for subsequent signal processing. Therefore, each frame of observation data observed by the sensor includes a plurality of observation point data, and part of the observation point data is noise data or false alarm data caused by various interferences and false signals (hereinafter, the noise data and the false alarm data are collectively referred to as noise data). According to an embodiment of the present application, the plurality of frames of observation data observed by the sensor within a predetermined time interval includes a plurality of observation point data observed by the sensor in each frame of observation, and the observation point data in the corresponding frame includes the velocity of the target in the corresponding frame, the acceleration of the target in the corresponding frame, the longitude position of the target in the corresponding frame and the latitude position of the target in the corresponding frame.

[0034] According to an embodiment of the present application, the M frames of observation data observed by the sensor within a predetermined time interval are denoted as Z = [Z1, Z2, …, ZM], wherein Z1 represents the observation data of the 1st frame, Z2 represents the observation data of the 2nd frame, and ZM represents the observation data of the Mth frame. M M ​represents the observation data of the Mth frame, and each frame of observation data includes a plurality of observation point data, such as the kth frame of observation data represented as represents the 1st observation point data of the kth frame, represents the 2nd observation point data of the kth frame, represents the n observation point data of the kth frame, such as the i observation point data of the kth frame represented as represents the i observation point data of the kth frame, indicating the longitude position of the target in the corresponding frame, represents the i observation point data of the kth frame, indicating the latitude position of the target in the corresponding frame, represents the i observation point data of the kth frame, indicating the speed of the target in the corresponding frame, represents the i observation point data of the kth frame, indicating the acceleration of the target in the corresponding frame. The following gives a specific example to illustrate the construction of each sample in the training set:

[0035] Example 1: Taking a predetermined time interval of 1 second as an example, the sensor observes 6 frames of observation data every 1 second, and each frame of observation data includes a plurality of observation point data. The label is the real motion trajectory of the target in the corresponding time, wherein the real motion trajectory includes 6 position coordinates (each position coordinate is the longitude position and latitude position of the target in the corresponding frame) corresponding to 6 frames of observation data. It should be understood that this is only illustrative, and the specific number of frames of observation data included in the input data can be set according to the specific needs, or the following examples two or three can be used to construct the sample, and the present application does not make any limitation on this.

[0036] Example 2: Taking a predetermined time interval of 1.5 seconds, and the sensor observes 6 frames of observation data every second, then the input data of each sample is 9 frames of observation data, and the label is the real motion trajectory of the target in the corresponding 1.5 seconds, and the real motion trajectory includes 9 position coordinates corresponding to 9 frames of observation data.

[0037] Example 3: Or taking a predetermined time interval of 1 second, the sensor observes 12 frames of observation data, then the input data of each sample is 12 frames of observation data, and the label is the real motion trajectory of the target in the corresponding 1 second, and the real motion trajectory includes 12 position coordinates corresponding to 12 frames of observation data.

[0038] II. Model structure

[0039] According to an embodiment of the present application, referring to Figure 1The trajectory observation data denoising model comprises a preprocessing unit, a denoising autoencoder unit and an activation function layer. The preprocessing unit performs denoising preprocessing on input data in the training set to obtain denoised input data. The denoising autoencoder comprises an encoder and a decoder. The encoder is configured to encode the denoised input data to obtain a hidden layer feature vector. The decoder is configured to decode the hidden layer feature vector to obtain an estimated value of the target motion state. Both the encoder and the decoder adopt a double-layer long short-term memory network or a double-layer gated recurrent neural network. The activation function layer is configured to process the estimated value of the target motion state according to an activation function to obtain a denoised motion trajectory. The activation function layer can adopt a Maxout activation function layer or a ReLU activation function layer. The technical scheme of the embodiment can achieve at least the following beneficial technical effects: The preprocessing unit of the application performs initial denoising preprocessing on the observation data, reduces the influence of noise data and abnormal data in the observation data, and further removes noise in the initial denoised observation data through the denoising autoencoder, thereby effectively denoising and obtaining a denoised motion trajectory, and improving the accuracy of the final obtained motion trajectory.

[0040] Since there is noise data in the multiple observation point data corresponding to each frame, directly using the observation data of each frame for motion trajectory estimation is inaccurate. According to an embodiment of the application, the preprocessing unit performs denoising preprocessing on the observation data of each frame in a fusion processing manner. The preprocessing unit comprises a learnable weight parameter matrix configured to perform fusion processing on each predetermined number of observation point data in each frame of observation data.

[0041] According to an embodiment of the application, the preprocessing unit performs denoising preprocessing by using a predetermined sliding window to read a predetermined number of observation point data from each frame of observation data at a time, and using a learnable weight parameter matrix to perform fusion processing on the observation point data read at a time, to achieve denoising preprocessing, wherein the predetermined number is greater than or equal to 2. For example, taking the fusion processing manner of reading five observation point data from the kth frame of observation data at a time as an example, the fusion processing manner is as follows:

[0042]

[0043] wherein X k1 represents the 1st observation point data of the kth frame after fusion processing, represents the 1st observation point data of the kth frame observed by the sensor, represents the 2nd observation point data of the kth frame observed by the sensor, represents the 3rd observation point data of the kth frame observed by the sensor, represents the 4th observation point data of the kth frame observed by the sensor, The 5th observation point data of the kth frame observed by the sensor is represented, and A represents a learnable weight parameter matrix. The observation data of the kth frame after fusion processing is denoted as The input data is processed by fusion processing to obtain the input data after noise reduction preprocessing. The input data after noise reduction preprocessing includes multiple frames of observation data after noise reduction preprocessing, and the ith input data after noise reduction preprocessing is denoted as The observation data of the 1st frame after noise reduction preprocessing is represented, The observation data of the 2nd frame after noise reduction preprocessing is represented, The observation data of the n th frame after noise reduction preprocessing is represented. Taking the training set obtained in the above example 1 as an example, the input data of each sample in the training set includes 6 frames of observation data observed by the sensor within a predetermined time interval of 1 second. The 6 frames of observation data observed within the predetermined time interval of 1 second are preprocessed by the preprocessing unit to obtain the input data after noise reduction preprocessing. Since the probability distribution of noise data is generally zero mean or approximately zero, the positive and negative offsets after addition will make the noise value after addition closer to the mean. Therefore, the input data is preprocessed by the fusion processing method to effectively reduce the noise data in the input data with zero mean or approximately zero mean distribution, and to weaken the influence of abnormal data. Furthermore, the input data after noise reduction preprocessing is encoded and decoded by the denoising autoencoder to remove the noise in the observation data after noise reduction, and to further effectively reduce the noise data in the input data. The model obtains the denoised motion trajectory based on the output after encoding and decoding by the denoising autoencoder, and improves the accuracy of the model in estimating the motion trajectory of the maneuvering target.

[0044] III. Model training process

[0045] According to an embodiment of the present application, a method for training a trajectory observation data denoising model of a maneuvering target is provided, as shown in Figure 2 The method comprises one or more rounds of iterative training of the model according to steps S1 and S2. Each round is trained one or more times until the model converges or the preset number of iterations is reached. In step S1, a training set is obtained, each sample in the training set including input data and labels. The input data is multiple frames of observation data of the target motion trajectory observed by the sensor within a predetermined time interval, and the labels are the true motion trajectory of the target within the corresponding time. In step S2, the model is trained using the training set, and the parameters of the model are updated based on the calculated loss. The preprocessing unit pre-processes the input data, the denoising autoencoder encodes and decodes the input data after noise reduction preprocessing, and the model obtains the denoised motion trajectory based on the output after encoding and decoding by the denoising autoencoder.

[0046] According to an embodiment of the present application, as shown inFigure 3 , Figure 3 The input data is generated and the model processes the input data, wherein the input data is based on the above-mentioned example one, and 6 frames of observation data are observed at a predetermined time interval of 1 second. Each circle in the input data corresponds to one frame of observation data, and each circle in the output of the decoder is the estimated value of the target motion state of the corresponding frame. The input data includes true values and noise data. In one training, the model is trained based on the training set obtained in the above-mentioned manner. First, the input data is preprocessed by the preprocessing unit to obtain the denoised preprocessed input data. Second, the input data is encoded by the encoder to obtain the hidden layer feature vector. Finally, the hidden layer feature vector is decoded by the decoder to obtain the estimated value of the target motion state, which is the denoised motion trajectory. In each training, the loss calculated based on the denoised motion trajectory and the true motion trajectory is used to update the learnable weight parameter matrix, the parameters of the encoder and the decoder. The mean square error loss function is used to calculate the loss. The loss calculation method is as follows:

[0047]

[0048] wherein Loss DAE represents the loss calculated in one training, n represents the total number of samples used in one training, i represents the i-th sample in the total number of samples, X i represents the label of the i-th sample in the total number of samples, f θ represents the decoder, f θ (·) represents the estimated value of the target motion state (i.e. the denoised motion trajectory), and θ represents the parameters of the decoder, represents the encoder, represents the hidden layer feature vector, represents the parameters of the encoder, represents the denoised preprocessed input data according to the i-th sample.

[0049] According to one embodiment of the present application, the model uses a Maxout activation function layer. In one training, the encoder is used to encode the denoised preprocessed input data to obtain the hidden layer feature vector. The decoder is used to decode the hidden layer feature vector to obtain the estimated value of the target motion state. The Maxout activation function layer is used to process the estimated value of the target motion state using the Maxout activation function to obtain the denoised motion trajectory. In each training, the loss calculated based on the denoised motion trajectory and the true motion trajectory is used to update the learnable weight parameter matrix, the parameters of the encoder, the decoder and the Maxout activation function layer.

[0050] According to one embodiment of the present application, the hidden layer of the Maxout activation function layer has k neurons (k can be manually set according to requirements, for example, 6 or 8 neurons), and one neuron corresponds to a set of network parameters. In each training, the result output by the decoder is obtained through the k neurons to obtain k data, and the maximum data in the k data is selected as the output to obtain the denoised motion trajectory.

[0051] According to one embodiment of the present application, the model adopts a ReLU activation function layer, and in each training, the ReLU activation function layer is used to process the estimated value of the target motion state to obtain the denoised motion trajectory. In each training, the loss calculated based on the denoised motion trajectory and the real motion trajectory is used to update the learnable weight parameter matrix, the encoder, and the decoder.

[0052] According to one embodiment of the present application, in each training, the parameters of the learnable weight parameter matrix in the preprocessing unit can be continuously updated with the model, or can be locked after one or more trainings or manually adjusted in subsequent specified trainings. For example, after one training, the parameters of the current preprocessing unit are updated according to the loss, or the parameters in the preprocessing unit are manually set and adjusted according to specific conditions, so as to adaptively adjust the parameters of the preprocessing unit, better denoising preprocessing of the observation data, and better denoising preprocessing of the input data in the next training. The parameters of the learnable weight parameter matrix in the preprocessing unit can also be adjusted to a better condition before training the model, and the fixed-parameter preprocessing unit is directly used for denoising preprocessing of the observation data. The present application can reduce the influence of noise data and improve the training efficiency of the model by denoising preprocessing of the input data through the preprocessing unit.

[0053] IV. Application scenarios

[0054] According to one embodiment of the present application, a method for tracking a maneuvering target is provided, which is used in the fields of security monitoring, intelligent transportation, or radar early warning. The method comprises: using a trajectory observation data denoising model trained based on the method described in the above embodiments of the present application, obtaining a denoised motion trajectory of a target according to inputted continuous multiple frames of observation data of the motion trajectory of the target; and tracking the maneuvering target according to the denoised motion trajectory. The target tracking method of the present application can directly track the target through the motion trajectory obtained by the trajectory observation data denoising model, and is applicable to various fields and has strong flexibility.

[0055] According to another embodiment of the present application, a method for tracking a maneuvering target is provided, which is used in the fields of security monitoring, intelligent transportation, or radar early warning.Figure 4 , Figure 4 To generate a schematic diagram of the motion trajectory of the maneuvering target, the trajectory observation data denoising model and the trajectory reconstruction module trained by the method of the above embodiment are used to process the input data to obtain the motion trajectory. The maneuvering target tracking method comprises: dividing a plurality of input data in the continuous multiple frames of observation data in a time sequence data format, wherein the input data comprises a predetermined number of frames of observation data, and adjacent two input data comprise multiple frames of repeated observation data; using the trajectory observation data denoising model trained based on the method of the above embodiment of the present application, obtaining the denoised motion trajectory of the target corresponding to each input data; using the trajectory reconstruction module to average the two denoised motion trajectories corresponding to the multiple frames of repeated observation data in the adjacent two input data to obtain the reconstructed motion trajectory; and tracking the maneuvering target according to the reconstructed motion trajectory. The present application can repeatedly input multiple frames of observation data into the model, and the corresponding frames of observation data can obtain multiple motion trajectories through the model. The average of the multiple motion trajectories is obtained through the trajectory reconstruction module, and the accuracy of the motion trajectory is further improved.

[0056] According to an embodiment of the present application, a plurality of input data is divided in a time sequence data format. Taking the case that the model input data contains 6 frames of observation data each time as an example, there are 5 frames of overlapping data between the adjacent two divided input data. For example, the first input data is the first, second, third, fourth and fifth frames of observation data, and the second input data is the second, third, fourth, fifth and sixth frames of observation data. Among the two input data, the second, third, fourth and fifth frames of observation data are repeatedly inputted. The average of the two denoised motion trajectories corresponding to the second, third, fourth and fifth frames is obtained between the adjacent two input data, and the reconstructed motion trajectory is obtained.

[0057] According to an embodiment of the present application, for example, in the field of security monitoring, the target tracking in the monitoring video is realized in combination with a target detection algorithm. The target detection algorithm can give multiple frames of observation data, and each frame of observation data comprises the vertex position or other position of the target detection box in the frame image, such as the coordinates of the two vertices on the diagonal line of the target detection box (bounding box) in the frame image. The training set is generated by the method of the above embodiment, and the trajectory observation data denoising model is trained. The trajectory observation data denoising model and the trajectory reconstruction module process the observation data in the field of security monitoring to obtain each frame of motion trajectory. Each frame of motion trajectory is taken as the center point coordinates of the target detection box in each frame of image. Multiple continuous motion trajectories correspond to multiple continuous center point coordinates of the target detection box, and the multiple continuous center point coordinates are inputted into the target tracking algorithm for target tracking.

[0058] It is to be understood that even though various embodiments can be described in conjunction with a specific order, the order is not absolute and the steps can be performed in other orders.

[0059] The present application can be a system, a method, and / or a computer program product. The computer program product can include a computer readable storage medium (or media) having computer readable program instructions thereon for causing a processor to carry out aspects of the present application.

[0060] The computer readable storage medium can be a tangible device that can retain and store instructions for use by an instruction execution device. The computer readable storage medium can be, for example, but is not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer readable storage medium include the following: a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanically encoded device such as punch-cards or raised structures in a groove having instructions recorded thereon, and any suitable combination of the foregoing. A computer readable storage medium, as used herein, is not to be construed as being transitory signals per se.

[0061] Embodiments of the application have been described above with the aid of example embodiments described above. The description above is intended to be illustrative, and not exhaustive, of the embodiments disclosed. Many modifications and variations of the described embodiments are possible and are within the scope of the described embodiments, as those skilled in the relevant art will recognize. The embodiments were chosen and described in order to best explain the principles of the application and its practical application, and to thereby enable others skilled in the art to best utilize the application and various embodiments with various modifications as are suited to the particular use contemplated.

Claims

1. A training method for a denoising model of trajectory observation data for maneuvering targets, characterized in that, The model includes a preprocessing unit and a denoising autoencoder, and the method includes: Obtain the training set. Each sample in the training set includes input data and a label. The input data consists of multiple frames of observation data obtained by the sensor observing the target's motion trajectory within a predetermined time interval. The label is the target's actual motion trajectory within the corresponding time period. Each frame of observation data includes multiple observation points observed by the sensor in the corresponding frame. The model is trained using a training set, and the model parameters are updated based on the calculated loss. A preprocessing unit performs denoising preprocessing on the input data, and a denoising autoencoder encodes and decodes the denoised input data. The model obtains the denoised motion trajectory based on the output of the denoising autoencoder. The preprocessing unit includes a learnable weight parameter matrix for fusing data from a predetermined number of observation points in each frame of observation data. During training, the parameters of the learnable weight parameter matrix of the denoising autoencoder and the preprocessing module are updated based on the loss.

2. The method according to claim 1, characterized in that, The preprocessing unit uses a predetermined sliding window to read a predetermined number of observation point data from each frame of observation data each time, and uses a learnable weight parameter matrix to perform fusion processing on the observation point data read each time, wherein the predetermined number is greater than or equal to 2.

3. The method according to claim 1, characterized in that, The observation point data for the corresponding frame includes the target's velocity, acceleration, longitude, and latitude in the corresponding frame.

4. The method according to any one of claims 1-3, characterized in that, The denoising autoencoder includes an encoder and a decoder, wherein, The encoder is used to encode the noise-reduced preprocessed input data to obtain hidden layer feature vectors; and The decoder is used to decode the hidden layer feature vector to obtain an estimate of the target motion state, and the estimate of the target motion state is used as the denoised motion trajectory. During training, the learnable weight parameter matrix, encoder, and decoder parameters are updated based on the loss calculated from the denoised motion trajectory and the real motion trajectory.

5. The method according to any one of claims 1-3, characterized in that, The denoising autoencoder includes an encoder and a decoder. The model also includes an activation function layer, which is a Maxout activation function layer. The encoder is used to encode the noise-reduced preprocessed input data to obtain hidden layer feature vectors; The decoder is used to decode the hidden layer feature vectors to obtain an estimate of the target's motion state; and The Maxout activation function layer is used to process the estimated value of the target motion state based on the Maxout activation function to obtain the denoised motion trajectory. During training, the learnable weight parameter matrix, encoder, decoder, and Maxout activation function layer parameters are updated based on the loss calculated from the denoised motion trajectory and the real motion trajectory.

6. A method for tracking a maneuvering target, characterized in that, For applications in security monitoring, intelligent transportation, or radar early warning systems, the methods include: Using a trajectory observation data denoising model trained based on the method described in any one of claims 1-5, the denoised motion trajectory of the target is obtained based on continuous multi-frame observation data of the input target motion trajectory. Among them, the moving target is tracked based on the denoised motion trajectory.

7. A method for tracking a maneuvering target, characterized in that, For applications in security monitoring, intelligent transportation, or radar early warning systems, the methods include: Multiple input data are divided into a time-series data format from continuous multi-frame observation data. The input data includes a predetermined number of observation data frames, and two adjacent input data frames include repeated multi-frame observation data. Using a trajectory observation data denoising model trained based on the method described in any one of claims 1-5, the denoised motion trajectory of the target is obtained according to each input data. The trajectory reconstruction module calculates the average of two denoised motion trajectories corresponding to repeated multi-frame observation data in two adjacent input data to obtain the reconstructed motion trajectory. Among these methods, the reconstructed motion trajectory is used to track the moving target.

8. A computer-readable storage medium, characterized in that, It stores a computer program that can be executed by a processor to implement the steps of the method according to any one of claims 1 to 7.

9. An electronic device, characterized in that, include: One or more processors; as well as Memory, wherein the memory is used to store executable instructions; The one or more processors are configured to implement the steps of the method according to any one of claims 1 to 7 by executing the executable instructions.

Images