An image comprehensive detection and recognition method and system based on DS evidence theory
Patent Information
- Application Number
- CN202311734170.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-12-15
- Publication Date
- 2026-10-23
- Estimated Expiration
- 2043-12-15
AI Technical Summary
[0005]针对现有技术的至少一个缺陷或改进需求，本发明提供了一种基于DS证据理论的图像综合检测识别方法及系统，用以至少解决现有检测算法和识别算法各自独立运行且不做融合，难以体系化解决视场内不同像素数目标检测识别的技术问题
[0042](1)本发明将基于先验知识的小尺度目标检测算法和基于深度学习的小尺度目标检测算法进行了融合，解决了传统单一的基于先验知识的小尺度目标检测算法虚警率偏高的技术问题。本发明将检测算法和识别算法进行了融合，解决了现有检测算法和识别算法各自独立运行且不做融合进而导致难以体系化解决视场内不同像素数目标检测识别的技术问题。
Smart Images
Figure CN117934999B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image detection and recognition technology, and more specifically, to an image comprehensive detection and recognition method and system based on DS evidence theory. Background Technology
[0002] Target tracking equipment plays a role in imaging and detecting the background and target within the field of view. Due to the influence of factors such as target type, size, distance, angle, and background environment, the number of pixels in the target image can range from a few pixels to hundreds of pixels or even fill the field of view, and the shapes are also very diverse.
[0003] Existing small-scale target detection algorithms based on prior knowledge often focus only on targets with a small number of pixels (1*1 to 20*20), and are sensitive to noise and complex scenes, easily generating false alarms. For targets with a larger number of pixels, they tend to extract only local details or not at all. In contrast, existing deep learning-based detection and recognition algorithms perform better in detecting and recognizing targets with a larger number of pixels (above 20*20), while having a lower detection rate and a lower false alarm rate for targets with a smaller number of pixels (1*1 to 20*20).
[0004] Existing single detection or recognition algorithms cannot simultaneously perform comprehensive detection and recognition of targets with different pixel distributions within the field of view, as well as more refined model identification. Summary of the Invention
[0005] To address at least one deficiency or improvement need in the existing technology, this invention provides an image comprehensive detection and recognition method and system based on DS evidence theory, which at least solves the technical problem that existing detection and recognition algorithms operate independently without fusion, making it difficult to systematically solve the detection and recognition of targets with different pixel counts within the field of view.
[0006] To achieve the above objectives, according to a first aspect of the present invention, an image comprehensive detection and recognition method based on DS evidence theory is provided, comprising:
[0007] Simultaneously perform detection and recognition on the same frame image in four dimensions: small-scale target detection based on prior knowledge, small-scale target detection based on deep learning, large-scale target recognition, and fine-grained target model recognition. Obtain the detection and recognition results of the coordinate position and model of the target to be identified in the image under each dimension.
[0008] The DS evidence theory is used to fuse the detection and identification results of the coordinate position and model of the target under various dimensions to obtain a comprehensive detection and identification result of the target.
[0009] Furthermore, the method of using DS evidence theory to fuse the detection and identification results of the coordinate position and model of the target under various dimensions to obtain a comprehensive detection and identification result of the target includes:
[0010] The detection and identification results of the coordinate position and model of the target to be identified in the four dimensions are fused into target boxes based on IOU to eliminate highly overlapping target boxes;
[0011] Based on eliminating highly overlapping target boxes, the detection and identification results of the coordinate position and model of the target to be identified under the four dimensions are fused with the identification type and confidence level based on DS evidence theory to obtain the configuration result of identification type and confidence level;
[0012] Based on the configuration results of the identification type and confidence level, small-scale targets are filtered and sorted to obtain comprehensive detection and identification results for the targets to be identified.
[0013] Furthermore, the algorithm logic for target box fusion based on IOU includes:
[0014] The intersection-union ratio is defined as the ratio of the size of the area where two bounding boxes intersect to the size of the target itself, with the larger value being the ratio of the two bounding boxes.
[0015] If the intersection-union ratio is greater than 0.6, the two bounding boxes are determined to belong to the same target; otherwise, the two bounding boxes do not belong to the same target.
[0016] If two bounding boxes belong to the same target, the target's dimensions are the length and width of its bounding rectangle.
[0017] Furthermore, the algorithm logic for fusing the identification type and confidence level based on DS evidence theory includes:
[0018] If the target bounding box is composed of small-scale and large-scale target bounding boxes, then the recognition type and confidence score are taken as the large-scale recognition type and confidence score.
[0019] If the target bounding box is composed of small-scale target bounding boxes and fine-grained target bounding boxes, then the recognition type and confidence score are taken as those of the fine-grained recognition type and confidence score.
[0020] If the target bounding box consists of small-scale, large-scale, and fine-grained target bounding boxes, or contains only large-scale and fine-grained target bounding boxes, then based on the DS evidence theory, a basic probability allocation is performed, and then the confidence scores of the fused large-scale and fine-grained targets are calculated. The target with the higher confidence score is selected as the identification type and confidence score.
[0021] Furthermore, the algorithm logic for filtering and sorting small-scale targets includes:
[0022] Among all small-scale targets, prioritize outputting the detection and recognition results or fusion results of small-scale target detection based on deep learning;
[0023] In the detection and recognition results of small-scale target detection based only on prior knowledge, the top three targets are output in order of their features.
[0024] Furthermore, the algorithm logic for small-scale target detection based on prior knowledge includes:
[0025] Suppress background clutter to highlight the target;
[0026] Select a preset threshold to segment the target pixels;
[0027] The discrete target pixels are associated and labeled to obtain the target;
[0028] Targets with high confidence are selected based on their characteristics.
[0029] Furthermore, the deep learning-based small-scale target detection algorithm employs 3D convolutional kernels, convolutional long short-term memory units, and an attention mechanism to build an end-to-end network algorithm model based on the attention mechanism and convolutional long short-term memory units. This establishes the target correlation between different image frames in the infrared image sequence and filters out target-related information from the spatiotemporal information.
[0030] Furthermore, the algorithm for large-scale target recognition employs a neural network algorithm based on few-shot learning to perform coarse classification and recognition of target types;
[0031] The algorithm for large-scale target recognition is trained in two steps:
[0032] Training based on base classes;
[0033] Fine-tuning based on small samples of the base class and the new class.
[0034] Furthermore, the algorithm for fine-grained target model recognition extracts features from the input image through a classification network and classifies it. Based on the attention region information obtained by training the APN network on the extracted features, the attention region is then cropped and enlarged as the input of the second-layer scale network. This process is repeated three times to obtain the output results of the three scale networks.
[0035] According to a second aspect of the present invention, an image comprehensive detection and recognition system based on DS evidence theory is also provided, comprising:
[0036] The first processing unit is used to deploy a small-scale target detection algorithm based on prior knowledge and a small-scale target detection algorithm based on deep learning, so as to obtain the detection and recognition results of the coordinate position and model of the target to be identified in the input image in the corresponding dimension;
[0037] The second processing unit is used to deploy a large-scale target recognition algorithm to obtain the detection and recognition results of the coordinate position and model of the target to be identified in the input image in the corresponding dimension;
[0038] The third processing unit is used to deploy a fine-grained target model recognition algorithm to obtain the detection and recognition results of the coordinate position and model of the target to be identified in the input image in the corresponding dimension;
[0039] The fourth processing unit is used to fuse the detection and identification results of the coordinate position and model of the target under the various dimensions using the DS evidence theory to obtain a comprehensive detection and identification result of the target.
[0040] The four processing units all use multi-core heterogeneous SOC chips based on embedded application processors and programmable logic. These four SOC chips are deployed on the same board and communicate through the SERDES high-speed bus.
[0041] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:
[0042] (1) This invention integrates a small-scale target detection algorithm based on prior knowledge and a small-scale target detection algorithm based on deep learning, thus solving the technical problem of high false alarm rate in traditional single small-scale target detection algorithms based on prior knowledge. This invention integrates detection and recognition algorithms, thus solving the technical problem that existing detection and recognition algorithms operate independently without integration, making it difficult to systematically solve the detection and recognition of targets with different pixel counts within the field of view.
[0043] (2) This invention refines the two levels of large-scale target recognition and fine-grained target model recognition, and constructs a recognition type and confidence fusion algorithm based on DS evidence theory. It can simultaneously realize coarse type recognition of targets with fewer details and fine model recognition of targets with more details, and solves the technical problems of target detection and recognition with different pixel numbers and fine-grained model recognition.
[0044] (3) This invention deploys a series of algorithms on four interconnected SOC chips, solving the technical problem of applying the algorithm framework to actual devices and achieving the technical effect of integrated detection and identification. Attached Figure Description
[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0046] Figure 1 A core flowchart of an image comprehensive detection and recognition method based on DS evidence theory provided in this embodiment of the invention;
[0047] Figure 2 A schematic diagram of the framework of the image comprehensive detection and recognition algorithm based on DS evidence theory provided in the embodiments of the present invention;
[0048] Figure 3 A flowchart illustrating a small-scale target detection algorithm based on prior knowledge provided in an embodiment of the present invention;
[0049] Figure 4 This is a schematic diagram of a small-scale target detection network structure based on deep learning provided in an embodiment of the present invention;
[0050] Figure 5 This is a schematic diagram of a large-scale target recognition network structure provided in an embodiment of the present invention;
[0051] Figure 6 This is a schematic diagram of the target fine-grained model identification network structure provided in an embodiment of the present invention;
[0052] Figure 7 A flowchart illustrating the target comprehensive detection and recognition algorithm provided in an embodiment of the present invention;
[0053] Figure 8 This is a schematic diagram of target box fusion based on IOU provided in an embodiment of the present invention;
[0054] Figure 9 A flowchart illustrating the identification type and confidence fusion algorithm based on DS evidence theory provided in this embodiment of the invention;
[0055] Figure 10 A flowchart illustrating the small-scale target screening and sorting algorithm provided in an embodiment of the present invention;
[0056] Figure 11 This is a schematic diagram of the circuit structure of an image comprehensive detection and recognition system based on DS evidence theory deployed using a domestically produced SOC, as provided in an embodiment of the present invention. Detailed Implementation
[0057] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.
[0058] The terms "first," "second," or "third," etc., used in the specification, claims, or accompanying drawings of this application are used to distinguish different objects and not to describe a particular order. Furthermore, the terms "comprising" or "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such processes, methods, products, or apparatus.
[0059] To systematically address the high false alarm rate caused by existing small-scale target detection algorithms based solely on prior knowledge, and to solve the technical problems of target detection and recognition with different pixel counts and fine-grained model identification, as well as the technical challenges of deploying these algorithms in practical devices, this application proposes an image comprehensive detection and recognition algorithm based on DS evidence theory and a corresponding 4SOC deployment scheme. This technology first employs a multi-dimensional detection and recognition method, simultaneously performing four different detection and recognition methods on the same frame: small-scale target detection based on prior knowledge, small-scale target detection based on deep learning, large-scale target recognition, and fine-grained model identification. This comprehensively acquires the coordinate position and model information of the target to be identified from different dimensions. Then, based on DS evidence theory, the detection and recognition results from different dimensions are fused to obtain a comprehensive detection and recognition result. Figure 1 Finally, the algorithm is deployed based on the 4SOC processing board, enabling the effective application of this algorithm framework to actual devices.
[0060] refer to Figure 2 The image comprehensive detection and recognition algorithm framework based on DS evidence theory includes a small-scale target detection algorithm based on prior knowledge, a small-scale target detection algorithm based on deep learning, a large-scale target recognition algorithm, a fine-grained target model recognition algorithm, and a comprehensive target detection and recognition algorithm. The first four algorithms detect and recognize the coordinate position and model of the target in the image from different dimensions. Then, the detection and recognition results are sent to the comprehensive target detection and recognition algorithm for fusion of target position, recognition type, and confidence level, etc., to obtain the comprehensive detection and recognition result of the target in the image.
[0061] The small-scale target detection algorithm based on prior knowledge employs a morphological filtering-based detection algorithm, capable of detecting targets in images with few pixels and lacking clear contours. Its process is as follows: Figure 3 As shown. Background suppression effectively reduces background clutter interference and highlights the target. Threshold segmentation selects a certain threshold to segment the target pixels. Target clustering associates and labels discrete target pixels to obtain the target. Target filtering further filters out high-confidence targets based on target features. The final output is the full-image multi-target detection result.
[0062] like Figure 4 As shown, the small-scale target detection algorithm based on deep learning uses 3D convolutional kernels, convolutional long short-term memory units, and attention mechanisms to build an end-to-end network algorithm model based on attention mechanisms and convolutional long short-term memory units. It establishes the target correlation between different image frames in the infrared image sequence and filters out important target-related information from a large amount of spatiotemporal information, thereby achieving effective detection of weak infrared targets.
[0063] The large-scale target recognition algorithm employs a neural network algorithm based on few-shot learning to perform coarse classification and recognition of target types. Training is performed in two steps: training based on the base class (with a large amount of labeled data), and fine-tuning based on few samples using both the base class and the new class (with only a small amount of labeled data). The network structure is as follows: Figure 5 As shown, the overall structure is quite complex, mainly consisting of the following three modules: feature extractor, reweighting module, and prediction layer.
[0064] Fine-grained model identification algorithms can achieve more precise identification of target models. For example... Figure 6 As shown, the target fine-grained model recognition algorithm extracts features from the input image through a classification network and classifies it. Based on the extracted features, the APN network is trained to obtain attention region information. Then, the attention region is cropped and enlarged as the input of the second-layer scale network. This process is repeated three times to obtain the output results of the three scale networks. Thus, recognition with granularity from coarse to fine is obtained in a progressive manner.
[0065] This application specifically and creatively designs a comprehensive target detection and recognition algorithm, the processing flow of which is as follows: Figure 7As shown, the algorithm first receives small-scale target detection results (including those based on prior knowledge and those based on deep learning), large-scale target recognition results, and fine-grained target model recognition results. Next, it performs IOU-based bounding box fusion to eliminate highly overlapping bounding boxes, thus solving the technical problem of multiple bounding boxes being identified for the same target. Then, it fuses the recognition type and confidence level based on DS evidence theory to obtain a comprehensive target recognition result. Finally, it filters and sorts the small-scale targets and outputs the final result.
[0066] IOU-based bounding box fusion algorithms, such as Figure 8 As shown. The Intersection over Union (IOU) of a target is calculated by dividing the size of the area where the target bounding boxes intersect by the size of the target itself (i.e., the Intersection over Union ratio). The maximum value is taken for both target boxes. If the IOU is greater than 0.6, the two target boxes are considered to belong to the same target; otherwise, they do not belong to the same target. If the two target boxes belong to the same target, the target's dimensions are the length and width of its bounding rectangle.
[0067] Based on the DS evidence theory, the identification type and confidence fusion algorithm is as follows: Figure 9 As shown, since a target may contain multiple bounding boxes, classification analysis is required. If the bounding box consists of small-scale bounding boxes (including those based on prior knowledge and those based on deep learning) and large-scale bounding boxes, then the recognition type and confidence score are those of the large-scale bounding box. If the bounding box consists of small-scale bounding boxes and fine-grained bounding boxes, then the recognition type and confidence score are those of the fine-grained bounding box. If the bounding box consists of small-scale bounding boxes, large-scale bounding boxes, and fine-grained bounding boxes, or contains only large-scale bounding boxes and fine-grained bounding boxes, then based on the DS evidence theory, a basic probability allocation is performed, and then the confidence scores of the fused large-scale and fine-grained targets are calculated. The recognition type and confidence score of the target with the higher confidence score are then selected.
[0068] Because deep learning-based small-scale target detection algorithms have a lower false alarm rate, the results are more reliable. Therefore, as... Figure 10 As shown, among all small-scale targets, the deep learning-based small-scale target detection results or fusion results are prioritized for output. Finally, among the small-scale target detection results that are only detected based on prior knowledge, the top three targets ranked according to their features are output.
[0069] In another embodiment, the domestically produced SOC deployment scheme for the image comprehensive detection and recognition algorithm framework based on DS evidence theory consists of four SOCs. These four SOCs are deployed on the same single board and communicate via the SERDES high-speed bus. A multi-core heterogeneous SOC chip based on an embedded application processor and programmable logic is selected as the basic processing unit for algorithm deployment. Through the efficient computing power and heterogeneous parallel programming architecture of the SOC chip, software-programmable high-real-time signal processing is achieved, providing hardware support for the hardware-software separation of the device. The software adopts an embedded real-time operating system, which facilitates the deployment of deep learning algorithms.
[0070] like Figure 11 As shown, among the four SOCs, SOC1 deploys a small-scale target detection algorithm based on prior knowledge and a small-scale target detection algorithm based on deep learning. SOC2 deploys a large-scale target recognition algorithm. SOC3 deploys a fine-grained target model recognition algorithm. The main SOC deploys a photoelectric target integrated recognition algorithm (i.e., a comprehensive target detection and recognition algorithm). Images are input to these four SOC boards and simultaneously transmitted to SOC1, SOC2, and SOC3 via a high-speed bus interface. After detection and recognition by these three SOCs, the results are transmitted to the main SOC via the high-speed bus. The main SOC then performs comprehensive recognition processing before outputting the results.
[0071] In general, the beneficial effects of this invention compared with the prior art are mainly reflected in the following aspects:
[0072] (1) This invention integrates the traditional small-scale target detection algorithm based on prior knowledge and the small-scale target detection algorithm based on deep learning, thereby reducing the technical problem of high false alarm rate of the traditional small-scale target detection algorithm based solely on prior knowledge.
[0073] (2) The present invention solves the technical problem that existing detection algorithms and recognition algorithms operate independently without fusion, making it difficult to systematically solve the detection and recognition of targets with different pixel counts within the field of view.
[0074] (3) This invention refines the two levels of large-scale target recognition and fine-grained target model recognition, and constructs a recognition type and confidence fusion algorithm based on DS evidence theory, which can simultaneously realize coarse type recognition of targets with fewer details and fine model recognition of targets with more details.
[0075] (4) This invention deploys a series of algorithms on four interconnected SOC chips, solving the technical problem of applying the algorithm framework to actual devices and achieving the technical effect of integrated detection and identification.
[0076] Those skilled in the art will understand that the features described in the various embodiments and / or claims of this disclosure can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this disclosure. In particular, the features described in the various embodiments and / or claims of this disclosure can be combined and / or combined in various ways without departing from the spirit and teachings of this disclosure, and all such combinations and / or combinations fall within the scope of this disclosure.
[0077] Although this disclosure has been shown and described with reference to specific exemplary embodiments thereof, those skilled in the art will understand that various changes in form and detail may be made to this disclosure without departing from the spirit and scope of the disclosure as defined by the appended claims and their equivalents. Therefore, the scope of this disclosure should not be limited to the above embodiments, but should be defined not only by the appended claims, but also by their equivalents.
Claims
1. An image comprehensive detection and recognition method based on DS evidence theory, characterized in that, include: Simultaneously perform detection and recognition on the same frame image in four dimensions: small-scale target detection based on prior knowledge, small-scale target detection based on deep learning, large-scale target recognition, and fine-grained target model recognition. Obtain the detection and recognition results of the coordinate position and model of the target to be identified in the image under each dimension. The DS evidence theory is used to fuse the detection and identification results of the coordinate position and model of the target under various dimensions to obtain a comprehensive detection and identification result of the target. The method of using DS evidence theory to fuse the detection and identification results of the coordinate position and model of the target under various dimensions to obtain a comprehensive detection and identification result of the target includes: The detection and identification results of the coordinate position and model of the target to be identified in the four dimensions are fused into target boxes based on IOU to eliminate highly overlapping target boxes; Based on eliminating highly overlapping target boxes, the detection and identification results of the coordinate position and model of the target to be identified under the four dimensions are fused with the identification type and confidence level based on DS evidence theory to obtain the configuration result of identification type and confidence level; Based on the configuration results of the identification type and confidence level, small-scale targets are filtered and sorted to obtain comprehensive detection and identification results for the targets to be identified. The algorithm logic for fusing identification type and confidence level based on DS evidence theory includes: If the target bounding box is composed of small-scale and large-scale target bounding boxes, then the recognition type and confidence score are taken as the large-scale recognition type and confidence score. If the target bounding box is composed of small-scale target bounding boxes and fine-grained target bounding boxes, then the recognition type and confidence score are taken as those of the fine-grained recognition type and confidence score. If the target bounding box consists of small-scale, large-scale, and fine-grained target bounding boxes, or contains only large-scale and fine-grained target bounding boxes, then based on the DS evidence theory, a basic probability allocation is performed, and then the confidence scores of the fused large-scale and fine-grained targets are calculated. The target with the higher confidence score is selected as the identification type and confidence score.
2. The image comprehensive detection and recognition method based on DS evidence theory as described in claim 1, characterized in that, The algorithm logic for IOU-based bounding box fusion includes: The intersection-union ratio is defined as the ratio of the size of the area where two bounding boxes intersect to the size of the target itself, with the larger value being the ratio of the two bounding boxes. If the intersection-union ratio is greater than 0.6, the two bounding boxes are determined to belong to the same target; otherwise, the two bounding boxes do not belong to the same target. If two bounding boxes belong to the same target, the target's dimensions are the length and width of its bounding rectangle.
3. The image comprehensive detection and recognition method based on DS evidence theory as described in claim 1, characterized in that, The algorithm logic for filtering and sorting small-scale targets includes: Among all small-scale targets, prioritize outputting the detection and recognition results or fusion results of small-scale target detection based on deep learning; In the detection and recognition results of small-scale target detection based only on prior knowledge, the top three targets are output in order of their features.
4. The image comprehensive detection and recognition method based on DS evidence theory as described in claim 1, characterized in that, The algorithm logic for small-scale target detection based on prior knowledge includes: Suppress background clutter to highlight the target; Select a preset threshold to segment the target pixels; Associative labeling is performed on discrete target pixels; Targets with high confidence are selected based on their characteristics.
5. The image comprehensive detection and recognition method based on DS evidence theory as described in claim 1, characterized in that, The deep learning-based small-scale target detection algorithm employs 3D convolutional kernels, convolutional long short-term memory units, and an attention mechanism to build an end-to-end network algorithm model based on the attention mechanism and convolutional long short-term memory units. It establishes the target correlation between different image frames in the infrared image sequence and filters out target-related information from the spatiotemporal information.
6. The image comprehensive detection and recognition method based on DS evidence theory as described in claim 1, characterized in that, The algorithm for large-scale target recognition uses a neural network algorithm based on few-shot learning to perform coarse classification and recognition of target types; The algorithm for large-scale target recognition is trained in two steps: Training based on base classes; Fine-tuning based on small samples of the base class and the new class.
7. The image comprehensive detection and recognition method based on DS evidence theory as described in claim 1, characterized in that, The algorithm for fine-grained target model recognition extracts features from the input image through a classification network and classifies it. Based on the extracted features, the APN network is trained to obtain attention region information. The attention region is then cropped and enlarged as the input to the second-layer scale network. This process is repeated three times to obtain the output results of the three scale networks.
8. An image comprehensive detection and recognition system based on DS evidence theory, using the method described in any one of claims 1-7, characterized in that, include: The first processing unit is used to deploy a small-scale target detection algorithm based on prior knowledge and a small-scale target detection algorithm based on deep learning, so as to obtain the detection and recognition results of the coordinate position and model of the target to be identified in the input image in the corresponding dimension; The second processing unit is used to deploy a large-scale target recognition algorithm to obtain the detection and recognition results of the coordinate position and model of the target to be identified in the input image in the corresponding dimension; The third processing unit is used to deploy a fine-grained target model recognition algorithm to obtain the detection and recognition results of the coordinate position and model of the target to be identified in the input image in the corresponding dimension; The fourth processing unit is used to fuse the detection and identification results of the coordinate position and model of the target under the various dimensions using the DS evidence theory to obtain a comprehensive detection and identification result of the target. The four processing units all use multi-core heterogeneous SOC chips based on embedded application processors and programmable logic. These four SOC chips are deployed on the same board and communicate through the SERDES high-speed bus.