Intelligent target tracking method based on big data technology deep learning
By adopting deep learning and long-term memory networks in the UAV radar tracking system, combined with the processing method of gate structure, the problem of high-precision tracking of small targets in complex environments is solved, and the tracking accuracy is significantly improved.
Patent Information
- Application Number
- CN202411446497.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-16
- Publication Date
- 2025-02-18
AI Technical Summary
The existing UAV radar tracking and positioning methods are difficult to ensure high-precision tracking of small targets in complex environments, especially under the influence of external magnetic fields and geographical environment.
Adopting an intelligent target tracking method based on deep learning of big data technology, the target state feature information is extracted from the echo data processed after signal processing, and the gate structure is introduced into the long-term and short-term memory network, and the network is optimized to handle long-term and short-term dependencies. ability.
It improves the tracking accuracy of weak targets in complex scenarios, and avoids the shortcomings of traditional tracking models and recurrent neural networks when dealing with timing information and long-term dependence problems.
Smart Images
Figure CN119471608A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of unmanned aerial vehicle flight position tracking, and specifically to an intelligent target tracking method based on deep learning of big data technology. Background Art
[0002] Radar is an effective means of low-altitude defense. Due to the strong clutter and interference in the complex urban environment, and the weak reflection echo of radar signals from small low-altitude targets, in order to ensure good target detection capability, radar needs to have good anti-interference ability and the ability to detect weak targets and identify them.
[0003] Unmanned aircraft are unmanned aircraft that are controlled by radio remote control equipment and self-contained program control devices, or are operated completely or intermittently autonomously by an onboard computer. Compared with manned aircraft, drones are often more suitable for tasks that are too "dull, dirty or dangerous", so they are widely used in aerial photography, agriculture, plant protection, micro selfies, express delivery, disaster relief, wildlife observation, monitoring infectious diseases, surveying and mapping, news reporting, power inspections, disaster relief, film and television shooting, creating romance, etc.
[0004] When the radar is working, it generally sends radio electromagnetic waves to the space to be detected through the transmitter. When the electromagnetic waves encounter targets or other objects during propagation in space, they will be reflected. When the radar's receiver is in working state, it can receive part of the electromagnetic waves reflected back by the target or other objects, and process these echoes to obtain target information, detect the target, and locate, identify, and track it.
[0005] Traditional radar target tracking algorithms include coordinate transformation, horizontal radial velocity calculation, point cloud clustering, track initiation, association matching, filtering, management and clustering, as well as target recognition such as motion and stillness, length, width and type. In actual application, radar requires a priori estimation ideas. However, for small-volume drones, their ability to generate signals is more easily affected by external magnetic fields, geographical environment, etc., and is limited by the prior information assumption of the known target motion model, making it difficult to ensure high-precision tracking of small targets. Summary of the invention
[0006] In order to overcome the problem that in the actual application of the existing drone radar tracking and positioning method, the radar needs a priori estimation ideas. For small-volume drones, their ability to generate signals is more easily affected by external magnetic fields, geographical environments, etc., and is limited by the prior information assumption of the known target motion model. It is difficult to ensure high-precision tracking of small targets. The embodiment of the present application provides an intelligent target tracking method based on deep learning of big data technology. By extracting the state feature information of the target from the echo data after signal processing, when importing each data into the long and short-term memory network, it can be based on the continuous optimization of the state combined with the processing of the gate structure to finally obtain the estimated parameters, so that the network has the ability to handle long-term and short-term dependencies, and finally improves the tracking accuracy of weak targets in complex scenarios.
[0007] The technical solution adopted by the embodiment of the present application to solve the technical problem is:
[0008] An intelligent target tracking method based on deep learning of big data technology, including reporting data after signal processing, extracting target status feature information from echo data after signal processing based on data analysis, point screening, dimension conversion and coordinate conversion;
[0009] Construct a long short-term memory network as a neural network for small targets moving in a long time dimension to support the matching of motion timing information and scenes, and introduce a gate structure into the long short-term memory network to support the selective memory of long and short-term motion features;
[0010] The construction of the long short-term memory network is completed by data preprocessing, which supports the target estimation state at any time and the measurement information input at the next time, and supports the long short-term memory network to output the optimal estimate of the target state according to the mapping relationship.
[0011] In a possible implementation, the gate structure introduced by the long short-term memory network includes a forget gate, an input gate and an output gate. The forget gate is responsible for determining the number of unit states retained from the previous moment to the current moment, the input gate is responsible for determining the number of unit states retained from the current moment input to the current moment, and the output gate is responsible for determining the output number of the unit state at the current moment.
[0012] In a possible implementation, the long short-term memory network selectively memorizes the long-term and short-term motion features of the target by referring to the motion features corresponding to the timing information in complex scenes through a gate structure.
[0013] In one possible implementation, the extracted state characteristic information of the target includes position, velocity and acceleration. The motion characteristics of small targets are learned to change the motion mode of small targets and to obtain the generated echo data, thereby supporting the extraction of the target's state characteristic information from the echo data after signal processing.
[0014] In a possible implementation, the data preprocessing includes reading, reorganization and normalization, and provides support for long short-term memory network simulation training.
[0015] In a possible implementation, the long short-term memory network maintains a continuous optimization state after construction is completed, and the optimization categories include changing the number of neurons, changing the number of network layers, and using different regularization methods.
[0016] In a possible implementation, data set generation is performed between the data preprocessing and the long short-term memory network, and data set segmentation is performed during the data set generation process.
[0017] In a possible implementation, step 1: extracting position, velocity and acceleration state feature information;
[0018] Step 2: Perform data preprocessing on the extracted state features;
[0019] Step 3: Generate a data set and split it according to requirements;
[0020] Step 4: Build and optimize the long short-term memory network;
[0021] Step 5: Output estimated parameters.
[0022] In a possible implementation, the process of extracting the state feature information in step 1 includes:
[0023] S1: Report data during information processing;
[0024] S2: The data parsing module parses the data packets reported by the signal center into radar status parameters and point trace information according to the protocol;
[0025] S3: The point trace screening module retains the points whose sum difference ratio amplitude modulus, azimuth ratio amplitude and signal-to-noise ratio meet the requirements, whose speed and distance are within the range of several pre-processing parameters, and removes the points that do not meet the processing requirements;
[0026] S4: The dimension conversion module converts the point trace azimuth angle measurement amplitude into the point trace azimuth angle, converts the point trace pitch and difference ratio amplitude into the point trace pitch angle, and converts the point trace amplitude and noise amplitude into the point trace signal-to-noise ratio;
[0027] S5: Convert the point data in polar coordinates into point data in rectangular coordinates through a coordinate conversion module.
[0028] The beneficial effects of this application are:
[0029] First, in this scheme, by extracting the state feature information of the target from the echo data after signal processing, and importing each data into the long short-term memory network, the estimated parameters can be finally obtained based on the processing of the gate structure based on the continuous optimization state, so that the network has the ability to process long-term and short-term dependencies, and finally improves the tracking accuracy of weak targets in complex scenes;
[0030] Second, in this scheme, a data set is generated by preprocessing after state feature information extraction, and a long short-term memory network is built, so that it is in a state of continuous optimization according to the input data, which is convenient for obtaining the optimal estimate of the target state based on the input state feature information, and is conducive to ensuring the maintenance-free characteristics and training authority of the long short-term memory network after construction. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] Figure 1 It is a schematic diagram of the process of radar target tracking based on deep learning of big data technology of the present invention;
[0032] Figure 2 It is a schematic diagram of the process of state feature extraction of the present invention;
[0033] Figure 3 It is a schematic diagram of the process of the long short-term memory network of the present invention;
[0034] Figure 4 It is a schematic diagram of the door structure of the present invention. DETAILED DESCRIPTION
[0035] The technical solution in the embodiment of the present application is to solve the problems of the above-mentioned background technology, and the overall idea is as follows:
[0036] Embodiment 1:
[0037] This embodiment introduces an intelligent target tracking method based on deep learning of big data technology. Figure 1-Figure 4 As shown, it includes reporting data after signal processing, extracting target status feature information from the echo data after signal processing based on data analysis, point trace screening, dimension conversion and coordinate conversion;
[0038] Among them, after data analysis, the data packets reported by information processing are parsed into radar state parameters and point trace information according to the protocol, and then after point trace screening and dimension conversion, the point trace data in polar coordinates are finally converted into point trace data in rectangular coordinate system through the coordinate conversion module;
[0039] Construct a long short-term memory network (its internal situation is as follows Figure 3As shown in Figure 1, it is a neural network for the movement of small targets in a long time dimension, supporting the matching of motion timing information and scenes, and introducing a gate structure (including forget gate, input gate and output gate) in the long short-term memory network. Figure 4 As shown), it supports the selective memory of long-term and short-term motion features;
[0040] The forget gate is responsible for determining the number of cell states retained from the previous moment to the current moment, the input gate is responsible for determining the number of cell states retained from the current moment to the current moment, and the output gate is responsible for determining the number of outputs of the current moment cell state;
[0041] The long short-term memory network can selectively memorize the long-term and short-term motion features of the target by referring to the motion features corresponding to the temporal information in complex scenes through the gate structure;
[0042] The long short-term memory network is built by data preprocessing, which supports the target estimation state at any time and the input of measurement information at the next time, and supports the long short-term memory network to output the optimal estimation of the target state according to the mapping relationship;
[0043] Among them, data preprocessing is the last step before network construction and training, and its purpose is to process the filtered target state information to be more suitable for network training. After the long short-term memory network forms a specific mapping relationship, when the target estimation state at any time and the measurement information at the next time are input, the long short-term memory network outputs the optimal estimation of the target state based on the current network according to the mapping relationship, thereby realizing weak target tracking under deep neural network.
[0044] By adopting the above technical solution:
[0045] The above design learns the motion characteristics of small targets to change their motion mode, generates a large amount of echo data, and then extracts the state feature information of the target from the echo data after signal processing. When the data are imported into the long short-term memory network, the estimated parameters can be finally obtained based on the processing of the gate structure based on continuous optimization of the state, so that the network has the ability to handle long-term and short-term dependencies, and finally improves the tracking accuracy of small targets in complex scenarios, so as to avoid the problem that the traditional tracking model does not contain timing information and is difficult to match the tracking scenario, and the problem that the recurrent neural network is prone to gradient explosion and gradient disappearance when facing long-term dependency problems.
[0046] Embodiment 2:
[0047] Based on Example 1, this example introduces the implementation steps of an intelligent target tracking method based on deep learning of big data technology, such as Figure 1As shown, step 1: extracting the position, velocity and acceleration state feature information;
[0048] Among them, Figure 2 As shown, the extraction of state feature information requires reporting data during information processing. The data parsing module parses the data packets reported by the signal source into radar state parameters and trace information according to the protocol. The trace screening module retains the traces whose sum difference ratio amplitude modulus, azimuth ratio amplitude and signal-to-noise ratio meet the requirements, and whose speed and distance are within the range of several preprocessing parameters, and removes the traces that do not meet the processing requirements. The dimension conversion module converts the trace azimuth angle measurement amplitude into the trace azimuth angle, converts the trace pitch and difference ratio amplitude into the trace pitch angle, and converts the trace amplitude and noise amplitude into the trace signal-to-noise ratio. Then, the coordinate conversion module converts the trace data in polar coordinates into the trace data in rectangular coordinates, and finally outputs the estimated parameters.
[0049] Secondly, based on learning the motion characteristics of small targets, the motion mode of small targets is changed to obtain the generated echo data, which supports the extraction of target state feature information from the echo data after signal processing;
[0050] Step 2: Perform data preprocessing on the extracted state features;
[0051] Among them, data preprocessing includes reading, reorganization and normalization, and provides support for long short-term memory network simulation training;
[0052] Step 3: Generate a data set and split it according to requirements;
[0053] Step 4: Build and optimize the long short-term memory network;
[0054] Among them, the long short-term memory network maintains a continuous optimization state after the completion of the construction, and the optimization categories include changing the number of neurons, changing the number of network layers and using different regularization methods to avoid the situation where the target state information is put into the traditional tracking model (which does not contain time series information and is therefore difficult to match the tracking scene) or the recurrent neural network (which is prone to gradient explosion and gradient vanishing problems when facing long-term dependency problems) due to its mobility in the short time dimension, resulting in poor performance in the short time dimension or the long time dimension;
[0055] Secondly, between data preprocessing and long short-term memory network is data set generation, and the data set generation process is involved in data set segmentation;
[0056] Step 5: Output estimated parameters.
[0057] By adopting the above technical solution:
[0058] The above design generates a data set by preprocessing after extracting state feature information, and builds a long short-term memory network, so that it is in a state of continuous optimization according to the input data, which is convenient for supporting the target estimation state at any time and the measurement information input at the next time, thereby supporting the long short-term memory network to output the optimal estimate of the target state according to the mapping relationship, which is conducive to ensuring the maintenance-free characteristics and training authority of the long short-term memory network after construction.
[0059] Finally, it should be noted that: Obviously, the above embodiments are only examples for clearly explaining the present invention, and are not intended to limit the implementation methods. For ordinary technicians in the relevant field, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived from this are still within the scope of protection of the present invention.
Claims
1. An intelligent target tracking method based on deep learning of big data technology, characterized in that: include: Report data after signal processing, extract target status feature information from echo data after signal processing based on data analysis, point screening, dimension conversion and coordinate conversion; Construct a long short-term memory network as a neural network for small targets moving in a long time dimension to support the matching of motion timing information and scenes, and introduce a gate structure into the long short-term memory network to support the selective memory of long and short-term motion features; The construction of the long short-term memory network is completed by data preprocessing, which supports the target estimation state at any time and the measurement information input at the next time, and supports the long short-term memory network to output the optimal estimate of the target state according to the mapping relationship.
2. The intelligent target tracking method based on deep learning of big data technology as claimed in claim 1, characterized in that: The gate structure introduced by the long short-term memory network includes a forget gate, an input gate and an output gate. The forget gate is responsible for determining the number of unit states retained from the previous moment to the current moment, the input gate is responsible for determining the number of unit states retained from the current moment to the current moment, and the output gate is responsible for determining the output number of the unit state at the current moment.
3. The intelligent target tracking method based on deep learning of big data technology as claimed in claim 2, characterized in that: The long short-term memory network selectively memorizes the long-term and short-term motion features of the target by referring to the motion features corresponding to the timing information in complex scenes through a gate structure.
4. The intelligent target tracking method based on deep learning of big data technology according to claim 1, characterized in that: The extracted state feature information of the target includes position, speed and acceleration. Based on learning the motion features of small targets, the motion mode of small targets is changed to obtain the generated echo data, thereby supporting the extraction of the state feature information of the target from the echo data after signal processing.
5. The intelligent target tracking method based on deep learning of big data technology according to claim 1, characterized in that: The data preprocessing includes reading, reorganization and normalization, and provides support for long short-term memory network simulation training.
6. The intelligent target tracking method based on big data technology deep learning as claimed in claim 1, characterized in that: The long short-term memory network maintains a continuous optimization state after construction is completed, and the optimization categories include changing the number of neurons, changing the number of network layers and using different regularization methods.
7. The intelligent target tracking method based on deep learning of big data technology according to claim 1, characterized in that: The data preprocessing and the long short-term memory network are combined to generate a data set, and the data set generation process involves data set segmentation.
8. An intelligent target tracking method based on deep learning of big data technology as described in claims 1 to 7, characterized in that: include: Step 1: Extract the position, velocity and acceleration state feature information; Step 2: Perform data preprocessing on the extracted state features; Step 3: Generate a data set and split it according to requirements; Step 4: Build and optimize the long short-term memory network; Step 5: Output estimated parameters.
9. The intelligent target tracking method based on deep learning of big data technology as claimed in claim 8, characterized in that: The step of extracting the state feature information includes: S1: Report data during information processing; S2: The data parsing module parses the data packets reported by the signal center into radar status parameters and point trace information according to the protocol; S3: The point trace screening module retains the points whose sum difference ratio amplitude modulus, azimuth ratio amplitude and signal-to-noise ratio meet the requirements, whose speed and distance are within the range of several pre-processing parameters, and removes the points that do not meet the processing requirements; S4: The dimension conversion module converts the point trace azimuth angle measurement amplitude into the point trace azimuth angle, converts the point trace pitch and difference ratio amplitude into the point trace pitch angle, and converts the point trace amplitude and noise amplitude into the point trace signal-to-noise ratio; S5: Convert the point data in polar coordinates into point data in rectangular coordinates through a coordinate conversion module.