Target tracking method and device, data processing equipment and storage medium

Abstract

The target tracking method comprises the following steps: obtaining a current predicted state of a target according to a history state of the target by using a prediction model, wherein a state vector in the prediction model comprises a first component and a second component, the first component comprises an original state vector of the target, and the second component comprises a bias state vector corresponding to a measurement bias of a corresponding sensor; obtaining observation information of the target affected by the measurement bias of the sensor by using one or more sensors; and correcting the current predicted state of the target according to the observation information of the target to obtain a final estimated state of the target. The method can balance stability and accuracy of target tracking.

Description

Technical Field

[0001] The embodiments in this specification relate to the field of multi-target tracking technology, and in particular to target tracking methods and apparatus, data processing equipment and storage media. Background Technology

[0002] Multi-target tracking is a major task for autonomous driving and intelligent transportation. However, single sensors often have limitations such as small coverage, large blind spots, and harsh working conditions. Therefore, the input for target tracking has naturally been extended from a single sensor to multiple sensors.

[0003] The fundamental technology for multi-sensor target tracking lies in multi-sensor information fusion. Related fusion technologies for multi-sensor target tracking can be divided into two main directions: pre-fusion and post-fusion. Pre-fusion primarily involves fusing the raw data from the sensors, such as point cloud fusion between LiDAR sensors or the fusion of LiDAR point clouds and camera pixels, to form a "super sensor" that is then sent to the downstream detection and tracking module. Post-fusion, on the other hand, involves using each sensor to perform basic detection and tracking tasks, and then sending the perceived results to the backend for fusion.

[0004] The two fusion approaches have their own application scenarios. Among them, in the post-fusion framework, the filtering-based post-fusion scheme is currently the mainstream method, which includes several steps such as data association, multi-sensor input methods, and back-end filters.

[0005] For clarity, the objects maintained and created by the current algorithm are collectively referred to as "targets," often used as inputs, with specific attributes such as position, length, width, height, orientation, velocity, and identification (ID). The inputs to the sensors are called "observations."

[0006] In target tracking and fusion tasks, data association refers to the association between observations and targets. In multi-target tracking, there are often multiple observations at the same time, and the problem of how to correspond one-to-one between observations and currently maintained targets is solved by data association.

[0007] The main function of the back-end filter is to receive the raw input and use motion and observation models to estimate the target's state. The filter's input itself doesn't involve multiple sensors; each input can be considered a single observation. Currently, there are two ways to connect multi-sensor inputs to a single-input filter: sequential input and synchronous parallel input. Sequential input involves alternating data from each sensor, triggering a state update whenever data arrives from any sensor. Parallel input, on the other hand, involves periodically collecting data from all sensors, triggering an update after the collection is complete.

[0008] Sequential input strategies have low latency, making them more suitable for scenarios with high real-time requirements. However, multi-sensor sequential input methods may suffer from continuous filter input jitter. This jitter can prevent the filter from obtaining a stable and accurate output, affecting the accuracy and stability of target tracking. Summary of the Invention

[0009] In view of this, the embodiments of this specification provide a target tracking method and apparatus, a data processing device and a storage medium, which can balance the stability and accuracy of target tracking.

[0010] First, this specification provides a target tracking method, the method comprising:

[0011] A prediction model is used to obtain the current predicted state of one or more targets based on their historical states. The state vector in the prediction model includes a first component and a second component, wherein the first component includes the original state vector of the target, and the second component includes the deviation state vector corresponding to the measurement deviation of the corresponding sensor.

[0012] Acquire observational information of a target from one or more sensors, which is affected by sensor measurement biases;

[0013] The current predicted state of the target is corrected based on the observation information corresponding to the target, so as to obtain the final estimated state of the target.

[0014] Optionally, the step of employing a prediction model to obtain the current predicted state of the target based on the historical states of one or more targets includes:

[0015] Using a preset state transition model, the current predicted state of one or more targets is obtained based on their original historical states.

[0016] Using a preset deviation motion model, the current predicted deviation of the target is obtained based on the historical deviations of one or more targets.

[0017] Optionally, the step of using a preset deviation motion model to obtain the current predicted deviation corresponding to the target based on the historical deviations of one or more targets includes:

[0018] Using a pre-defined proportional attenuation model, the current prediction deviation of a target is obtained based on the historical deviations of one or more targets.

[0019] Optionally, the target tracking method further includes:

[0020] Based on the final estimated state of the target, the historical state corresponding to the target in the target state pool is updated, and the target state pool is adapted to store one or more historical states of the target.

[0021] Optionally, the target tracking method further includes:

[0022] Match the target in the target state pool that corresponds to the currently received observation information;

[0023] When it is determined that there is no target in the target state pool corresponding to the currently received observation information, a target corresponding to the currently received observation information is created in the target state pool.

[0024] Optionally, matching the target in the target state pool that corresponds to the currently received observation information includes:

[0025] Obtain the current predicted state corresponding to each target in the target state pool;

[0026] The observation vector information in the current predicted state of each target is matched with the currently received observation information to determine the target in the target state pool that corresponds to the currently received observation information.

[0027] Optionally, the target tracking method further includes:

[0028] If a target in the target state pool does not have its state updated within a preset time period, the target is deleted from the target state pool.

[0029] Optionally, acquiring observational information of the target from one or more sensors, affected by sensor measurement biases, includes:

[0030] The observation information of the target is obtained sequentially according to the time when the sensor actually observes and obtains the observation information of the target.

[0031] Optionally, acquiring observational information of the target from one or more sensors, affected by sensor measurement biases, includes:

[0032] According to a predetermined time interval, the observation information of the target obtained by each sensor within the time interval is obtained.

[0033] This specification also provides a target tracking device, the device comprising:

[0034] The prediction unit is adapted to use a prediction model to obtain the current predicted state of one or more targets based on their historical states. The state vector in the prediction model includes a first component and a second component, wherein the first component includes the original state vector of the target, and the second component includes the deviation state vector corresponding to the measurement deviation of the corresponding sensor.

[0035] An observation information acquisition unit is adapted to acquire observation information of a target from one or more sensors, which is affected by the measurement bias of the sensors.

[0036] The correction unit is adapted to correct the current predicted state of the target based on the observation information corresponding to the target, so as to obtain the final estimated state of the target.

[0037] This specification also provides a data processing device, the data processing device comprising: a target tracking device adapted to perform target tracking based on input from multiple sensors, the data processing device comprising: a processor and a memory, wherein the memory stores computer instructions executable on the processor, characterized in that the processor executes the steps of the method described in any of the foregoing embodiments when executing the computer instructions.

[0038] This specification also provides a computer-readable storage medium storing computer instructions that, when executed, perform the steps of the method described in any of the foregoing embodiments.

[0039] The target tracking scheme adopted in the embodiments of this specification, in the process of obtaining the current predicted state of the target based on the historical state of the target, since the state vector in the prediction model includes not only the original state vector of the target, but also the deviation state vector corresponding to the measurement deviation of the corresponding sensor, and similarly, the actual observation information is also observation information affected by the measurement deviation of the sensor, the current predicted state of the target can be corrected based on the observation information corresponding to the target, thereby eliminating the influence of the sensor measurement deviation on the target tracking result (i.e., the final estimated state of the target), thus taking into account both the stability and accuracy of target estimation and improving the target tracking quality.

[0040] Furthermore, in the process of obtaining the current predicted state of the target using the prediction model, on the one hand, a preset state transition model is used to obtain the current predicted state of the target based on the original historical states of one or more targets; on the other hand, a preset deviation motion model is used to obtain the current predicted deviation of the target based on the historical deviations of one or more targets. Therefore, the prediction model used can reflect the trend of the target's state change and the influence of the deviation trend of the sensor observing the target on the trend of the target's state change. Therefore, using the prediction model including the state transition model and the deviation motion model to predict the target state can improve the accuracy and stability of the target prediction results and improve the target tracking quality.

[0041] Furthermore, by employing a preset proportional attenuation model, the current predicted deviation of the target is obtained based on the historical deviation of one or more targets. This can reflect the changing trend of the sensor's measurement deviation with relatively small computational load, saving computational resources and improving computational efficiency while ensuring the accuracy and precision of target tracking results. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0043] Figure 1 A schematic diagram illustrating an application scenario for acquiring observation information of a target through multiple sensors is shown.

[0044] Figures 2A to 2D This diagram illustrates the tracking results obtained under different filtering parameters for multi-sensor input.

[0045] Figure 3 A flowchart of a target tracking method according to an embodiment of this specification is shown;

[0046] Figure 4 This diagram illustrates the relationship between target tracking results and observation results when the attenuation model is not used in the prediction model corresponding to sensor deviation in a specific observation scenario.

[0047] Figure 5 It shows in Figure 4 The diagram shows the relationship between the target tracking results and the observation results obtained when using the attenuation model of the embodiment in this specification in the observation scenario shown.

[0048] Figure 6The experimental data curves of two lidars in a specific application scenario of this specification are shown;

[0049] Figure 7A A flowchart of a target state pool maintenance method according to an embodiment of this specification is shown;

[0050] Figure 7B A flowchart of another target state pool maintenance method in an embodiment of this specification is shown;

[0051] Figure 8 A flowchart of a target matching method according to an embodiment of this specification is shown;

[0052] Figure 9A A schematic diagram illustrating the input method when obtaining observation information is shown;

[0053] Figure 9B A schematic diagram illustrating the input method when acquiring observation information from multiple sensors is shown.

[0054] Figure 10 A schematic diagram of the structure of a target tracking device according to an embodiment of this specification is shown;

[0055] Figure 11 A schematic diagram of the structure of a data processing device according to an embodiment of this specification is shown. Detailed Implementation

[0056] As mentioned in the background section, during the target tracking fusion process, when tracking targets based on observation information input sequentially from multiple sensors, there may be a phenomenon of continuous jitter in the multi-sensor input, which will cause the filter to fail to obtain a stable and accurate output, affecting the accuracy and stability of target tracking.

[0057] Specifically, research and experiments have revealed the following difficulties when sensors exhibit bias: For the same target, different sensors may have different ranges due to object movement or variations in sensor sensing range, resulting in inconsistent input information from different sensors. If information from multiple sensors is input sequentially and alternately, it causes input jitter, preventing the filter from obtaining a stable and accurate output.

[0058] Reference Figure 1The diagram illustrates an application scenario where multiple sensors acquire observational information about a target. Sensor D1 has a detection range of F1, and sensor D2 has a detection range of F2. For the actual target F0, the observational information received by sensor D1 represents the information detected by sensor D1 within region F1, while the observational information received by sensor D2 represents the information detected by sensor D2 within region F2. Therefore, as the target F0 moves, if the observational information from sensors D1 and D2 is sequentially and alternately input, it will cause jitter in the filter input. This jitter results in the filter failing to obtain a smooth and accurate output. Figures 2A to 2D The diagram shows the tracking results obtained under different filtering parameters in a multi-sensor input scenario. The merged-mea curve represents the sequential alternating input after fusion of two sensors. Figures 2A to 2D As can be seen from the merge-mea curve, due to bias, the input value jumps back and forth between the two observation sources; the p-est curve is the filter's estimate of the input observation; the v-est curve is the first derivative of the input observation; R represents the observation covariance in the Kalman filter, where R takes values ​​of 100, 1, 0.01, and 0.0001. A larger R indicates greater trust in the motion model, resulting in greater output hysteresis; conversely, a smaller R indicates greater trust in the observations, leading to output jitter. Figures 2A to 2D It is evident that simply adjusting the filter parameters cannot produce a stable and accurate output.

[0059] To address the aforementioned issues, the target tracking method employed in the embodiments of this specification, in the process of obtaining the current predicted state of the target based on its historical state, uses a prediction model whose state vector includes not only the original state vector of the target but also the deviation state vector corresponding to the measurement deviation of the corresponding sensor. As the foregoing analysis shows, the actual observation information obtained is also affected by the measurement deviation of the sensor. Therefore, by correcting the current predicted state of the target based on the observation information corresponding to the target, the influence of the sensor's measurement deviation on the target tracking result (i.e., the final estimated state of the target) can be eliminated, thereby balancing the stability and accuracy of target estimation and improving the quality of target tracking.

[0060] To enable those skilled in the art to better understand the technical concept, principles, and advantages of the present invention, the following detailed description is provided with reference to the accompanying drawings and specific application scenarios through specific embodiments.

[0061] Reference Figure 3 The flowchart of the target tracking method described above may include steps S31, S32 and S33 in a specific implementation.

[0062] In step S31, a prediction model is used to obtain the current predicted state of the target based on the historical states of one or more targets. The state vector in the prediction model includes a first component and a second component, wherein the first component includes the original state vector of the target, and the second component includes the deviation state vector corresponding to the measurement deviation of the corresponding sensor.

[0063] In specific implementation, prediction models that conform to their own change trends can be adopted for the first component and the second component respectively. Specifically, a preset state transition model can be used to obtain the current predicted state of the target based on the original historical state of one or more targets; and a preset deviation motion model can be used to obtain the current predicted deviation of the target based on the historical deviation of one or more targets.

[0064] The historical state of the one or more targets may include only the state at a single historical moment, thereby reducing the amount of data storage and computation.

[0065] As a specific example, assuming the number of sensors is L, the state vector in the prediction model can be represented as:

[0066]

[0067] Where x′ is the state vector in the prediction model, x is the first component, and b is the second component, the current predicted state of the target object can be represented as:

[0068]

[0069] in:

[0070] b k =[b 11 ..., b 1N′ , ..., b L1 ..., b LN′ , T (3)

[0071] In the second component, the first subscript {1, ..., L} indicates the sensor to which the deviation state quantity belongs, and the second subscript {1, ..., N′} corresponds to the number of state components in the original state quantity whose deviation needs to be estimated.

[0072] In practical implementation, for the first component, the current predicted state corresponding to the target can be obtained using the following state transition equation:

[0073] x k =A k x k-1 +ωk (4)

[0074] Among them, A k Let ω be the state transition matrix. k This is process noise, and ω k ∈R N ~N(0, Q) k That is, assuming it has a mean of zero and a covariance matrix of Q. k It follows a multivariate normal distribution.

[0075] The following state transition equation can then be used to predict the target state in the prediction model:

[0076]

[0077] Among them, B k Let β be the deviation state transition matrix. k It is the process noise of the deviation state quantity, and how the deviation evolves is determined according to the actual motion model.

[0078] For scenarios with a small rate of change in deviation, a proportional decay model can be used. Based on the historical deviations of one or more targets, the current predicted deviation corresponding to the target can be obtained. Then, the deviation state transition matrix can be:

[0079]

[0080] Among them, b k =0 < η 11 , ..., η 1N′ , ..., η L1 , ..., η LN , <1.

[0081] In step S32, observation information of the target from one or more sensors, which is affected by the measurement bias of the sensors, is obtained.

[0082] During the observation process, the observation information obtained by any sensor may be affected by measurement deviation. Furthermore, as the target state changes, at any given moment, the target may be within the field of view of at least one sensor. Therefore, at any given moment, observation information of the target by one or more sensors, affected by the sensor's measurement deviation, can be obtained.

[0083] In practical implementation, an observation vector can be constructed based on the sensor's observation information. For ease of expression, the observation vector is represented as y. Since the observation bias of each sensor is usually different, each sensor corresponds to a different observation equation. For observations of sensor i∈[1,L], the following observation equation can be used to obtain the observation information:

[0084]

[0085] Where: n k Let n represent the observation noise vector, and n k ∈R M ~N(0, R) k That is, assuming that its mean is zero and its covariance matrix is ​​R. k It follows a multivariate normal distribution.

[0086]

[0087] The above observation equations can be used to obtain observation information about the target from one or more sensors, which is affected by the measurement bias of the sensors.

[0088] In practice, the observed information may have the same parameters as the original state vector of the target in the prediction model, or it may have fewer state parameters than the original state vector of the target in the prediction model, for example, it may have fewer dimensions. As a specific example, the original state vector of the target is:

[0089] x = [pvsq] T

[0090] =[p x p y p z v x v y v z l ω hq ω q x q y q z ] T ;

[0091] The observation vector of the target:

[0092] y = [psq] T =[p x p y p z l ω hq ω q x q y q z ] T ;

[0093] Where p represents position, v represents velocity, s represents length, width, and height, and q represents attitude. More specifically, p x p y and p z These represent the positions along the x, y, and z axes in the coordinate system, respectively. x vy and v z Let x, y, and z represent the velocities along the x, y, and z axes in the coordinate system, respectively; l, ω, and h represent the length, width, and height of the target, respectively; and q... ω q x q y q z Represents the rotation quaternion of the target.

[0094] In the example above, the observed vector lacks the velocity dimension of the target's original state vector. The prediction model can obtain the velocity dimension of the target based on its position changes at multiple time points.

[0095] In step S33, the current predicted state of the target is corrected based on the observation information corresponding to the target, so as to obtain the final estimated state of the target.

[0096] After obtaining the observation information corresponding to the target through step S32, the observation information corresponding to the target can be used to correct the current predicted state of the target, thereby obtaining the final estimated state of the target.

[0097] In practice, Bayesian filtering can be used. As a specific example, Kalman filtering can be used for correction, or the target tracking process described above can be performed by a Kalman filter.

[0098] Even with the addition of sensor measurement bias, the entire target tracking prediction model still falls within the Bayesian filtering framework if Bayesian filtering or Kalman filtering is used. Therefore, it can be solved using the classic Bayesian filter solution method, which will not be described in detail here.

[0099] In the above embodiment, during the process of obtaining the current predicted state of the target based on its historical state, the state vector in the prediction model includes not only the original state vector of the target but also the deviation state vector corresponding to the measurement deviation of the sensor. Furthermore, the actual observation information obtained is also affected by the measurement deviation of the sensor. The current predicted state of the target is then corrected using the observation information affected by the measurement deviation of the sensor. Therefore, the influence of the sensor measurement deviation on the target tracking result (i.e., the final estimated state of the target) can be eliminated, thereby balancing the stability and accuracy of target estimation and improving the quality of target tracking.

[0100] In the foregoing embodiments, as a preferred example, a proportional attenuation model was used to predict the current measurement deviation of the sensor. Its function is illustrated below through an example in a specific scenario.

[0101] The inventors discovered through research that, without employing an attenuation model, the bias persists even when the target transitions from being observed by multiple sensors simultaneously to being observed by only one sensor. Figure 4 As shown, the prediction model used predicts the current state of the sensor's measurement deviation. Without the attenuation model, the diagram illustrates the relationship between the output target tracking result and the observation result. Curve Sen1 represents which sensor the observation information at the current moment comes from. Specifically, before time t1, only the observation information from sensor 1 is available. During the time interval t1 to t2, sensor 1 and sensor 2 are input alternately. After time t2, only the observation information from sensor 2 is available. The sawtooth-shaped observation curve d1 represents the sequential observation input, and the output curve c1 is the final estimated state curve of the target. As can be seen from the magnified view of region Z1, when switching to only the observation information input from sensor 2, the deviation still persists.

[0102] The reason why the bias persists is that the solution to the observation equation is not unique. For example, "the true value of 1.0 plus the bias of 0" and "the true value of 0.9 plus the bias of 0.1" both conform to the observation equation. Therefore, during the estimation process, the estimator assumes that the bias always exists, which is reasonable.

[0103] In the specific implementation of this specification, if the observation information is input in real time, the update time of the observation information is very short. In this case, a constant model can be used to predict the deviation. In reality, the measurement deviation of the sensor can only be perceived by the simultaneous observation input of multiple sensors. If there is no observation from multiple sensors for a long time, then the observation of the sensor's measurement deviation is actually in an unobserved state. Over time, the constant model will no longer hold true and will not match the actual measurement deviation of the sensor.

[0104] To address the aforementioned issues, the embodiments in this specification may employ a proportional attenuation model. As a preferred example, the attenuation coefficient in the proportional attenuation model can be 0.9. When only one sensor is present for actual observation, the deviation can be gradually reduced to zero. Figure 5 As shown, the prediction model used in this study predicts the current state of the sensor's measurement deviation. When using the attenuation model, the diagram illustrates the relationship between the output target tracking result and the observation result. Curve Sen2 represents which sensor the observation information at the current moment comes from. Specifically, before time t3, only the observation information from sensor 1 is available. During the period from t3 to t4, sensor 1 and sensor 2 are input alternately. After time t4, only the observation information from sensor 2 is input. The sawtooth curve d2 represents the sequential observation input, and the output curve c2 is the final estimated state curve of the target. As can be seen from the magnified view of region Z2, when switching to only the observation information input from sensor 2, the deviation state gradually returns to zero, and the output curve c2 and the observation curve d2 coincide.

[0105] The following example illustrates this through a specific application scenario.

[0106] Two lidar sensors simultaneously observe a small car, obtaining its position q from the sensor observations. However, there is a discrepancy between the observations from the two sensors. Through target tracking, its position p and velocity v0 can be estimated.

[0107] First, the state vector in the prediction model can be represented as:

[0108] x k =[pv b1 b2] T (9)

[0109] Where, p k It is a state component with bias, b 1k and b 2k The deviation in the position of the car is caused by the observations of sensor 1 and sensor 2 respectively.

[0110] The following state transition equation can be used for prediction:

[0111]

[0112] Where Δt is the time interval between the k-th frame and the (k-1)-th frame, η1 and η2 are attenuation coefficients, for example, they can be 0.9, ω pk and ω vk These are the process noises for position and velocity, respectively.

[0113] The observation equations corresponding to the two lidar sensors are as follows:

[0114] q k (1) = [1 0 1 0] + n k (11)

[0115] q k (2) = [1 0 0 1] + n k (12)

[0116] Then, the position p and velocity v0 of the vehicle can be estimated by using the standard Kalman filter method.

[0117] Reference Figure 6The experimental data curves for the two lidar sensors are shown below. The sawtooth curve merged-mea represents the sequential input of position observations from the two lidar sensors. Curve p-est is the filter-estimated output of position, curve v-est is the velocity-estimated output, and curves pb0-est and pb1-est are the deviation estimates of position between the two sensors, respectively. R is set to 0.0001. Figure 6 It can be seen intuitively that, compared to Figures 2A to 2D The stability and accuracy of the output position estimation have been greatly improved.

[0118] In specific applications, the aforementioned embodiments can be further expanded and optimized. The following detailed description is based on the accompanying drawings and specific application scenarios.

[0119] As a specific application example, a target state pool can be used to store and maintain the state corresponding to the target, as shown in the following example. Figure 7A and Figure 7B The flowchart of the target state pool maintenance method shown may include step S7A in some embodiments of this specification.

[0120] In step S7A, the historical state corresponding to the target in the target state pool is updated according to the final estimated state of the target. The target state pool is adapted to store one or more historical states of the target.

[0121] As described in the previous embodiment, the final estimated state of the target can be obtained through steps S31 to S33. After that, the historical state corresponding to the target in the target state pool can be updated, so that the target state pool always maintains the latest historical state of the target, thereby improving the target accuracy.

[0122] In other embodiments of this specification, step S7B may be included, as described below. Figure 7B The flowchart shown is for the target state pool maintenance method.

[0123] In step S7B, when the state of a target in the target state pool is not updated within a preset time period, the target is deleted from the target state pool.

[0124] In practice, if the state of a target in the target state pool is not updated within a preset time period, it indicates that the state of the target may not have changed or may no longer be within the observation field of view. In order to save storage resources, the target can be deleted from the target state pool.

[0125] It should be noted that in specific implementation, steps S7A and S7B can be selected to be executed as needed, or both can be executed in a target state pool maintenance scheme, and can be selected to be executed according to the state of the corresponding target in the target state pool.

[0126] In actual target tracking, to achieve more accurate target tracking and prediction, multi-target prediction is usually employed. Accordingly, the target state pool stores the states of multiple targets. Therefore, after acquiring observation information, it is necessary to match the observation information with the targets in the target state pool before using the observation information corresponding to that target to correct the current predicted state of that target. (Refer to...) Figure 8 The flowchart shown below illustrates the target matching method. The target matching process is described in detail below.

[0127] S81, Match the target in the target state pool that corresponds to the currently received observation information.

[0128] In some embodiments of this specification, matching can be performed in the following manner: First, the current predicted state corresponding to each target in the target state pool is obtained. Then, the observation vector information in the current predicted state of each target is matched with the currently received observation information to determine the target in the target state pool that corresponds to the currently received observation information.

[0129] S82, when it is determined that there is no target in the target state pool corresponding to the currently received observation information, a target corresponding to the currently received observation information is created in the target state pool.

[0130] In practical implementation, if no corresponding target is found in the target pool for the currently received observation information, it indicates that a new target has appeared within the observation field of view. Therefore, a target corresponding to the currently received observation information can be created in the target state pool based on the currently received observation information. The prediction model can then be used to further predict its state, and the current predicted state in the target pool can be corrected based on subsequent observation information to obtain the final estimated state of the target. Similarly, the entire prediction process considers the impact of sensor measurement bias on the newly created target, and the target tracking method described in the foregoing embodiments of this specification can be used to eliminate the impact of sensor measurement bias on the final tracking result of the target.

[0131] When adding a new target to the target state pool, the initial state vector of the target is assigned a value using the observation vector of the target. When there are vector dimensions that are not present in the observation vector, such as the second component in the target's state vector used to represent the measurement deviation, a default value can be assigned to it. For example, the default value of the measurement deviation of each dimension in the second component can be set to zero (i.e., assuming there is no deviation at present), etc.

[0132] Through the above process, the currently received observation information can be quickly matched with the corresponding target in the target pool, thereby reducing the time delay between the currently received observation information and the state information of the target in the target pool, and further improving the prediction accuracy.

[0133] As described in the previous embodiment, observation information of the target affected by the measurement deviation of the sensor can be obtained through one or more sensors. It should be noted that during the entire target tracking process, depending on the different motion characteristics of the target, or due to the limitation of the sensor distribution location, or due to external interference or sensor failure, observation information from multiple sensors may be received simultaneously at certain times or periods, or only observation information from one sensor may be received.

[0134] Combined with reference Figure 9A and Figure 9B The diagram illustrates an input method for acquiring observation information from one or more sensors. Sensor 1 and Sensor 2 can detect observation information in real time at the times indicated by the corresponding arrows. In practice, the observation information of the target can be obtained sequentially according to the actual time at which the sensors acquire the target's observation information. Figure 9A As shown, the moment when sensor 1 and sensor 2 actually obtain the observation information of the target is the moment when one observation of the target is obtained.

[0135] In practical implementation, observation information can also be acquired non-real-time. For example, observation information of the target obtained by each sensor within a predetermined time interval can be obtained. Figure 9B As shown, sensor 1 and sensor 2 may have actually obtained multiple observation information. They can acquire observation information of one or more sensors on the target generated within a time interval, or they can acquire only the observation information of one or more targets on the target generated at the most recent moment.

[0136] This specification also provides corresponding target tracking devices in its embodiments, see below. Figure 10 The diagram shown is a structural schematic of a target tracking device. In the embodiments of this specification, as shown... Figure 10 As shown, the target tracking device 100 may include: a prediction unit 101, an observation information acquisition unit 102, and a correction unit 103, wherein,

[0137] The prediction unit 101 is adapted to use a prediction model to obtain the current predicted state of the target based on the historical state of one or more targets. The state vector in the prediction model includes a first component and a second component, wherein the first component includes the original state vector of the target, and the second component includes the deviation state vector corresponding to the measurement deviation of the corresponding sensor.

[0138] The observation information acquisition unit 102 is adapted to acquire observation information of a target from one or more sensors that is affected by the measurement deviation of the sensors;

[0139] The correction unit 103 is adapted to correct the current predicted state of the target based on the observation information corresponding to the target, so as to obtain the final estimated state of the target.

[0140] The specific tracking method, tracking principle, specific implementation examples, application scenarios and effects of the target tracking device can be referred to the foregoing embodiments, and will not be repeated here.

[0141] This specification also provides a data processing device suitable for target tracking based on input from multiple sensors, such as... Figure 11 As shown, the data processing device 110 may include a processor 111 and a memory 112, wherein the memory 112 stores computer instructions that can be executed on the processor 111, and the processor 111 executes the steps of the target tracking method described in any of the foregoing embodiments when executing the computer instructions. Specific steps can be found in the foregoing embodiments and will not be repeated here.

[0142] This specification also provides corresponding computer-readable storage media storing computer instructions that, when executed, perform the steps of the target tracking method described in any of the foregoing embodiments. Specific steps can be found in the foregoing embodiments and will not be repeated here.

[0143] The computer-readable storage medium may include, for example, any suitable type of memory cell, memory device, memory article, memory medium, storage device, storage article, storage medium and / or storage unit, such as memory, removable or non-removable medium, erasable or non-erasable medium, writable or rewritable medium, digital or analog medium, hard disk, floppy disk, optical disc read-only memory (CD-ROM), recordable optical disc (CD-R), rewritable optical disc (CD-RW), optical disc, magnetic medium, magneto-optical medium, removable memory card or disk, various types of digital universal optical disc (DVD), magnetic tape, cassette tape, etc.

[0144] Computer instructions may include any suitable type of code implemented using any appropriate high-level, low-level, object-oriented, visual, compiled, and / or interpreted programming language, such as source code, compiled code, interpreted code, executable code, static code, dynamic code, encrypted code, etc.

[0145] While the embodiments of the present invention have been disclosed above, the present invention is not limited thereto. Any person skilled in the art can make various modifications and alterations without departing from the spirit and scope of the present invention; therefore, the scope of protection of the present invention should be determined by the scope defined in the claims.

Claims

1. A target tracking method, characterized in that, include: A prediction model is used to obtain the current predicted state of one or more targets based on their historical states. This includes: using a preset state transition model to obtain the current predicted state of one or more targets based on their original historical states; and using a proportional attenuation model to obtain the current predicted deviation of one or more targets based on their historical deviations. The state vector in the prediction model includes a first component and a second component, wherein the first component includes the original state vector of the target, and the second component includes the deviation state vector corresponding to the measurement deviation of the corresponding sensor. Acquire observational information of a target from one or more sensors, which is affected by sensor measurement biases; The current predicted state of the target is corrected based on the observation information corresponding to the target, so as to obtain the final estimated state of the target.

2. The target tracking method according to claim 1, characterized in that, Also includes: Based on the final estimated state of the target, the historical state corresponding to the target in the target state pool is updated, and the target state pool is adapted to store one or more historical states of the target.

3. The target tracking method according to claim 1 or 2, characterized in that, Also includes: Match the target in the target state pool that corresponds to the currently received observation information; When it is determined that there is no target in the target state pool corresponding to the currently received observation information, a target corresponding to the currently received observation information is created in the target state pool.

4. The target tracking method according to claim 3, characterized in that, The matching of targets in the target state pool that correspond to the currently received observation information includes: Obtain the current predicted state corresponding to each target in the target state pool; The observation vector information in the current predicted state of each target is matched with the currently received observation information to determine the target in the target state pool that corresponds to the currently received observation information.

5. The target tracking method according to claim 2, characterized in that, Also includes: If a target in the target state pool does not have its state updated within a preset time period, the target is deleted from the target state pool.

6. The target tracking method according to claim 1, characterized in that, The acquisition of observational information of the target from one or more sensors, affected by sensor measurement biases, includes: The observation information of the target is obtained sequentially according to the time when the sensor actually observes and obtains the observation information of the target.

7. The target tracking method according to claim 1, characterized in that, The acquisition of observational information of the target from one or more sensors, affected by sensor measurement biases, includes: According to a predetermined time interval, the observation information of the target obtained by each sensor within the time interval is obtained.

8. A target tracking device, characterized in that, include: The prediction unit is adapted to use a prediction model to obtain the current predicted state of one or more targets based on their historical states. This includes: using a preset state transition model to obtain the current predicted state of one or more targets based on their original historical states; and using a proportional attenuation model to obtain the current predicted deviation of one or more targets based on their historical deviations. The state vector in the prediction model includes a first component and a second component, wherein the first component includes the original state vector of the target, and the second component includes the deviation state vector corresponding to the measurement deviation of the corresponding sensor. An observation information acquisition unit is adapted to acquire observation information of a target from one or more sensors, which is affected by the measurement bias of the sensors. The correction unit is adapted to correct the current predicted state of the target based on the observation information corresponding to the target, so as to obtain the final estimated state of the target.

9. A data processing device, characterized in that, include: Suitable for target tracking based on input from multiple sensors, the data processing device includes a processor and a memory, wherein the memory stores computer instructions executable on the processor, characterized in that the processor executes the steps of the method according to any one of claims 1 to 7 when executing the computer instructions.

10. A computer-readable storage medium storing computer instructions thereon, characterized in that, When the computer instructions are executed, they perform the steps of the method according to any one of claims 1 to 7.

Images