Multi-view jigsaw type unmanned aerial vehicle video target identification method

Abstract

The invention discloses a multi-view jigsaw type unmanned aerial vehicle video target identification method, which comprises the following steps: step 1, carrying out progressive multi-attribute feature extraction on a current image frame in an unmanned aerial vehicle video to obtain a multi-scale depth fusion feature of the current image frame; step 2, constructing a single-view-angle attribute recognition model, and performing classification recognition based on the multi-scale deep fusion features extracted in the step 1 to obtain a single-view-angle attribute vector under the current view angle; 3, taking the next image frame in the unmanned aerial vehicle video as the current image frame, and repeating the steps 1 to 2; obtaining single-view-angle attribute vectors under different view angles; step 4, performing multi-view jigsaw feature fusion on the single-view attribute vectors under different views to obtain jigsaw image fusion features of the target; and step 5, according to the jigsaw image fusion features of the target, target model discrimination is carried out by using a weighted voting and Bayesian network joint decision-making method, and multi-view jigsaw unmanned aerial vehicle video target identification is completed.

Description

TECHNICAL FIELD

[0001] The present application relates to a video target recognition method, in particular to a multi-view puzzle type unmanned aerial vehicle video target recognition method. BACKGROUND

[0002] The information provided in this section is merely background information related to the present disclosure and can not necessarily be prior art.

[0003] During the process of high-altitude flight of the unmanned aerial vehicle, since the target is in a motion state, the key features under multi-view are easy to be shielded, and the to-be-recognized target category exists in several types, the categories are various, the features of the same category can change greatly, and the images also exist in multiple heights and multiple angles. Therefore, in the scene of target maneuvering, how to quickly find the key detailed features of the target and realize multi-granularity classification recognition is one of the difficulties.

[0004] It should be noted that the information disclosed in the above background section is only used to strengthen the understanding of the background of the present disclosure, and therefore can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0005] The present application aims to solve the technical problems of the prior art and provides a multi-view puzzle type unmanned aerial vehicle video target recognition method.

[0006] In order to solve the above technical problems, the present application discloses a multi-view puzzle type unmanned aerial vehicle video target recognition method, comprising the following steps:

[0007] Step 1: performing progressive multi-attribute feature extraction on the current image frame in the unmanned aerial vehicle video to obtain multi-scale deep fusion features of the current image frame;

[0008] Step 2: constructing a single-view attribute recognition model, performing classification recognition based on the multi-scale deep fusion features extracted in step 1 to obtain a single-view attribute vector under the current view;

[0009] Step 3: taking the next image frame in the unmanned aerial vehicle video as the current image frame, repeating steps 1 to 2; obtaining single-view attribute vectors under different views;

[0010] Step 4: performing multi-view puzzle type feature fusion on the single-view attribute vectors under different views to obtain puzzle type image fusion features of the target;

[0011] Step 5: using a weighted voting and Bayesian network joint decision method to discriminate the target type according to the puzzle type image fusion features of the target, and completing the multi-view puzzle type unmanned aerial vehicle video target recognition.

[0012] Further, the progressive multi-attribute feature extraction in step 1 includes:

[0013] Step 1-1, multi-scale analysis is performed using a 5-layer Gaussian pyramid to generate a 5-layer Gaussian pyramid;

[0014] Step 1-2, a feature pyramid network is constructed to extract feature maps from each layer of the 5-layer Gaussian pyramid;

[0015] Step 1-3, the final multi-scale deep fusion feature P is calculated fusion , as follows:

[0016]

[0017] where α l is the normalized weight coefficient of the feature map P l , and the calculation method is as follows:

[0018]

[0019] where exp represents the exponential function, H(P l ) represents the information entropy of the feature map P l , and the calculation method is as follows:

[0020]

[0021] where p l,c is the probability distribution of the pixel value of the l-th layer c-th channel feature.

[0022] Further, the 5-layer Gaussian pyramid generated in step 1-1 is as follows:

[0023] In the 5-layer Gaussian pyramid, the l-th layer pyramid image G l (x,y) is generated from the (l-1)-th layer image G l-1 by Gaussian filtering and 2 times downsampling, and is represented as follows:

[0024]

[0025] where (x,y) is the pixel coordinate of the l-th layer image, G l-1 (2x+i,2y+j) represents the pixel value of the corresponding position in the (l-1)-th layer image, where 2x+i and 2y+j are the coordinates of the previous layer image, i.e. the key information of the even rows or columns is retained by 2 times downsampling, and W(i,j) is the Gaussian kernel weight matrix, represented as follows:

[0026]

[0027] In the generated 5-layer Gaussian pyramid, G0 is the original image.

[0028] Further, the feature maps are extracted from each layer image of the 5-layer Gaussian pyramid in step 1-2, specifically as follows:

[0029] 1 / 4 resolution feature C2 is extracted from image G0, 1 / 8 resolution feature C3 is extracted from image G1, 1 / 16 resolution feature C4 is extracted from image G2, and 1 / 32 resolution feature C5 is extracted from image G3;

[0030] Feature fusion is performed from top to bottom, as shown below:

[0031] P5 = Conv1x1(C5)

[0032] P4 = Upsample(P5) + Conv1x1(C4)

[0033] P3 = Upsample(P4) + Conv1x1(C3)

[0034] P2 = Upsample(P3) + Conv1x1(C2)

[0035] Where P2, P3, P4 and P5 represent the fusion results of each layer feature, Upsample is a bilinear interpolation up-sampling operation, Conv1x1 represents a 1*1 convolution operation for adjusting the channel number of the feature map.

[0036] Further, the classification recognition in step 2 includes:

[0037] Step 2-1, the attribute set of the current image frame is divided into M levels;

[0038] Step 2-2, for the mth level attribute, a special classifier Cm is designed m , inputting the multi-scale deep fusion feature P fusion , outputting the confidence s m of the level attribute, as shown below:

[0039] s m = C m (P fusion ; θ m )

[0040] Where θ m is the parameter of the mth level special classifier, m = 1, 2, …, M;

[0041] Step 2-3, the mth component V m of the single-view attribute vector V of the current image frame is calculated, as shown below:

[0042] V m = γ m * s m

[0043] wherein, γ m is a hierarchical attribute weight, satisfying:

[0044]

[0045] Step 2-4, obtaining the attribute vector set V of the current image frame, i.e. the single-view attribute vector under the current view, is expressed as follows:

[0046] V = {V1, V2,..., V M}.

[0047] Further, the multi-view puzzle type feature fusion described in step 4 includes:

[0048] Step 4-1, assuming that the single-view attribute vector of the kth view is V k , expressed as follows:

[0049] V k = [V k,1 , V k,2 ,..., V k,M ]

[0050] Step 4-2, mapping the single-view attribute vector V k of the kth view to a unified coordinate system through the alignment matrix A i , expressed as follows:

[0051] V' k = A i * V k

[0052] wherein, the alignment matrix A i is a diagonal matrix, and the diagonal elements represent the space scaling coefficients;

[0053] Step 4-3, assuming that the weight has a positive correlation with the attribute, expressed as follows:

[0054]

[0055] wherein, w k,m represents the normalized weight of the mth component of the attribute vector under the kth view, and V k,m represents the mth component of the single-view attribute vector of the kth view;

[0056] Step 4-4, performing puzzle type fusion.

[0057] Further, the puzzle type fusion described in step 4-4 is specifically as follows:

[0058]

[0059] where F k represents the kth component of the kth view of the complete attribute vector F after puzzle fusion;

[0060] The puzzle image fusion feature of the target is represented as follows:

[0061] F = [F1, F2,..., F K ]

[0062] where K is the number of views.

[0063] Further, the target model discrimination in step 5 includes:

[0064] Step 5-1, calculating the weighted prediction probability P(d) of each model d of the target;

[0065] Step 5-2, determining the preliminary discrimination result according to the weighted prediction probability P(d);

[0066]

[0067] that is, selecting the model d corresponding to the maximum probability P(d) as the preliminary discrimination result;

[0068] Step 5-3, using the Bayesian network for posterior verification.

[0069] Further, the calculation of the weighted prediction probability P(d) of each model d of the target in step 5-1 is represented as follows:

[0070]

[0071] where K is the number of views, w k is the weight of the kth view, P k (d) is the prediction probability of the kth view for the model d.

[0072] Further, the posterior verification using the Bayesian network in step 5-3 is as follows:

[0073] Step 5-3-1, obtaining the prior probability P(d) of each model according to historical data or domain knowledge;

[0074] Step 5-3-2, calculating the likelihood probability P(w|d), which represents the probability of the combination of each view feature w appearing under the condition that the model is d;

[0075] Step 5-3-3: Calculate the posterior probability P(d|w) using Bayes' theorem, as follows:

[0076]

[0077] Where P(d) is the prior probability and P(w) is the evidence probability;

[0078] Step 5-3-4: Set a threshold τ. When P(d|w)>τ, confirm the discrimination result.

[0079] Beneficial effects:

[0080] 1. The multi-view jigsaw puzzle-style UAV video target fine-grained recognition method proposed in this invention constructs a multi-scale target attribute recognition model to support target attribute recognition at different distances.

[0081] 2. This invention employs a target classification technology based on distance metrics to achieve continuous model discrimination of target image sequences, thereby improving the accuracy of fine-grained target recognition. Attached Figure Description

[0082] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0083] Figure 1 This is a schematic diagram of the overall process of the method of the present invention. Detailed Implementation

[0084] The overall idea of ​​this invention is as follows: Addressing the shortcomings of existing UAV video target recognition capabilities, this invention provides a multi-view mosaic-style UAV video target recognition method. Employing techniques such as model building, data fusion, and attribute extraction, it proposes a multi-view mosaic-style fine-grained UAV video target recognition method. A multi-scale target attribute recognition model is constructed to support target attribute recognition at different distances. A distance-based target classification technique is used to achieve continuous model discrimination of target image sequences, thereby improving the accuracy of fine-grained target recognition.

[0085] like Figure 1 As shown, the technical solution of the present invention is as follows: an attribute-based multi-view jigsaw puzzle-style UAV video target recognition method, which mainly solves the problem of low target recognition due to the wide variety of target equipment types, similar model structures and texture features. It mainly includes four steps: progressive multi-attribute feature extraction and analysis, single-view attribute recognition, multi-view jigsaw puzzle-style feature fusion, and target model discrimination.

[0086] Step one, progressive multi-attribute feature extraction. A three-level attribute analysis framework of "bottom vision - middle structure - high-level semantics" is adopted to gradually excavate the distinguishable features of the target. The focus is on extracting fine-grained attributes such as texture roughness, component relative position, and functional module layout, breaking through the bottleneck of traditional recognition methods based on global features in similar type discrimination.

[0087] By constructing a hybrid multi-scale feature system containing a Gaussian pyramid and a feature pyramid network (FPN). By capturing the natural scale changes of the image through the Gaussian pyramid, combined with the semantic enhancement mechanism of the FPN, the deep fusion of multi-scale features is realized. Specifically as follows:

[0088] Firstly, a 5-layer Gaussian pyramid is used to realize multi-scale analysis. The image G l (x,y) of the l-th layer is generated by the Gaussian filtering and 2 times down-sampling of the image G l-1 of the (l-1)-th layer:

[0089]

[0090] where (x,y) is the pixel coordinate of the l-th layer image. G l-1 (2x+i,2y+j) represents the pixel value of the corresponding position in the (l-1)-th layer image, where 2x+i and 2y+j are the coordinates of the previous layer image (the key information of even rows / columns is preserved by 2 times down-sampling). The Gaussian kernel weight matrix W(i,j) is:

[0091]

[0092] Generate a 5-layer pyramid. Among them, G0 is the original image, with a resolution of 1920*1080, covering close-range small targets, G1 resolution 960*540, covering medium-sized targets, G2 resolution 480*270, covering large equipment, G3 resolution 240*135, covering cluster targets, G4 resolution 120*67, covering scene-level targets.

[0093] Secondly, the feature pyramid network (FPN) is constructed. On the basis of the Gaussian pyramid, through the fusion of features from bottom to top and from top to bottom, the semantic information of different scale features is enhanced, solving the problem of low resolution of high-level features and weak semantic of bottom-level features. ResNet is used as the backbone network to extract feature maps from each layer of the Gaussian pyramid. C2 (1 / 4 resolution) is extracted from G0, C3 (1 / 8 resolution) is extracted from G1, C4 (1 / 16 resolution) is extracted from G2, and C5 (1 / 32 resolution) is extracted from G3.

[0094] The feature fusion formula from top to bottom FPN is:

[0095] P5 = Conv1x1(C5)

[0096] P4 = Upsample (P5) + Conv1x1 (C4)

[0097] P3 = Upsample (P4) + Conv1x1 (C3)

[0098] P2 = Upsample (P3) + Conv1x1 (C2)

[0099] Where Upsample is a bilinear interpolation up-sampling operation, used to increase the resolution of the feature map, in this way, the semantic information of the high layer is passed to the low layer, while the detailed information of the low layer is preserved.

[0100] Conv1x1 represents a 1*1 convolution operation, used to adjust the number of channels of the feature map, and the number of channels is unified to 256 dimensions.

[0101] Then the multi-scale coverage of the Gaussian pyramid is combined with the semantic enhancement of the FPN to form the final multi-scale deep fusion feature P fusion :

[0102]

[0103] Where p l,c is the pixel value probability distribution of the c-th channel feature of the l-th layer.

[0104] The fused feature contains the global semantic information of the image, which plays an important role in target classification and recognition.

[0105] Step two, single-view attribute recognition. Use the multi-scale deep fusion feature P fusion formed in step one to construct a multi-modal attribute classifier, design a differentiated recognition model for different levels of attribute features, and generate a single-view attribute vector by attribute confidence weighted fusion, which provides quantitative attribute basis for distinguishing similar models (such as Volkswagen Passat and Jetta).

[0106] The differentiated hierarchical attribute recognition model first divides the attribute set into M levels (m = 1, 2,..., M, such as bottom / middle / high)

[0107] For the m-th level attribute, a special classifier C m is designed (the model used in this application is Rest101), which inputs the fusion feature P fusion and outputs the confidence of this level attribute.

[0108] s m = C m (P fusion ; θ m )

[0109] where θ m are parameters of the mth hierarchical classifier, s m ∈ [0, 1] represents the predicted confidence of the mth hierarchical attribute.

[0110] The mth component of the single-view attribute vector V is:

[0111] V m = γ m * s m

[0112] where γ m is the hierarchical attribute weight, satisfying

[0113]

[0114] The final attribute vector is the collection of all components:

[0115] V = {V1, V2,..., V M ,}

[0116] Through the above steps, the single-view bottom, middle and high-level attribute vector information is formed.

[0117] Step three, multi-view puzzle feature fusion. Through geometric alignment and attribute association, multi-view information is "puzzle" fused. The complementary attributes under different views (such as the radar layout of the front view and the engine nozzle number of the side view) are mainly fused to construct a complete target attribute description and solve the recognition ambiguity caused by incomplete single-view information.

[0118] Suppose the attribute feature vector of the kth view is V k = [V k,1 , V k,2 ,..., V k,M ]

[0119] First, the view k feature is mapped to a unified coordinate system through the alignment matrix A i :

[0120] V' k = A i * V k

[0121] A i is a diagonal matrix, and the diagonal elements represent the space scaling coefficients.

[0122] Then the weight of attribute m and view k is positively related to the attribute determination:

[0123]

[0124] Finally, the puzzle fusion is performed

[0125] The kth component of the complete attribute vector F after fusion is:

[0126]

[0127] The overall representation is:

[0128] F = [F1, F2,..., F M ]

[0129] Through the above steps, a complete puzzle image fusion feature is formed, and a complete target attribute feature description is formed.

[0130] Step four, target model discrimination.

[0131] A joint decision mechanism of "weighted voting + Bayesian network" is adopted to realize accurate mapping from the attribute vector to the model category. The discrimination ability of the differentiated attributes between similar models is highlighted, such as the subtle difference in tail light layout between the Volkswagen Magotan and the Passat.

[0132] The weighted voting decision is to determine the target model according to the prediction results of each view angle on different target models, combined with the weights of each view angle.

[0133] For each view angle k, the puzzle feature F k obtained in step three is used, and the special classifier C m in step two outputs the prediction probability P m (d) of different target models d. According to the weights w m of each view angle, the prediction probabilities of different models are weighted and summed, and the model with the maximum weighted sum probability is selected as the preliminary discrimination result.

[0134] The weighted prediction probability P(d) of each model d is calculated:

[0135]

[0136] Where M is the number of view angles, w m is the weight of the mth view angle, and P m (d) is the prediction probability of the mth view angle for model d. The preliminary discrimination result

[0137]

[0138] That is, the model d corresponding to the maximum P(d) is selected as the preliminary discrimination result.

[0139] Bayesian network posterior check. The Bayesian network uses prior knowledge and information from each view angle to perform posterior check on the preliminary result of weighted voting, further improving the accuracy of discrimination.

[0140] The prior probability P(d) is determined, and the prior probability of each model appearing can be obtained according to historical data or domain knowledge.

[0141] The likelihood probability P(w|d) is calculated, which represents the probability of each view feature combination w appearing under the condition that the model is d.

[0142] The posterior probability P(d|w) is calculated according to the Bayes formula.

[0143] A threshold τ is set, when P(d|w)>τ, the discrimination result is confirmed; otherwise, the multi-view re-inspection mechanism is triggered.

[0144] Bayes formula:

[0145] Where P(d|w) is the posterior probability of the target model being d given the view feature combination w, P(w|d) is the likelihood probability; P(d) is the prior probability; P(w) is the evidence probability, which can be calculated by the total probability formula:

[0146]

[0147] Embodiment:

[0148] Taking the unmanned aerial vehicle (UAV) to identify the Volkswagen Magotan and Passat in urban roads as an example, the specific application of the method in civilian scenarios is detailed. Both car models belong to Volkswagen B-class cars, and their appearances are highly similar, so they need to be accurately identified through subtle attribute differences. The scene is set as follows: the UAV flies at a height of 100 meters at a speed of 30 km / h along the urban trunk road, takes a video of the morning rush hour traffic, and needs to distinguish between the mixed Magotan (80 vehicles) and Passat (60 vehicles). The key differences between the two car models are as follows: the Magotan has a "Magotan" logo on the tail and the inner side of the tail light is a square light strip, while the Passat has a "Passat" logo on the tail and the inner side of the tail light is a round light strip.

[0149] The main steps include:

[0150] Step 1: Perform progressive multi-attribute feature extraction.

[0151] Step 1-1 generates a Gaussian pyramid. The original image (G0, 1920×1080) clearly captures the details of the vehicle's rear markings and lighting, such as the edges of the "Magotan" lettering and the outline of the square light strip. G1 (960×540) preserves the vehicle's proportions and overall rear structure, suitable for extracting attributes such as "rear length-to-width ratio" and "taillight spacing." G2 (480×270) covers the overall vehicle shape and is used to identify mid-level attributes such as "vehicle length" and "window-to-body ratio." G3 (240×135) identifies the vehicle's position and direction of travel within the lane, helping to eliminate interference from non-target areas. G4 (120×67) analyzes the impact of road environment (such as main roads / minor roads) on vehicle type distribution and optimizes recognition priority.

[0152] Steps 1-2 involve bottom-up feature extraction from the FPN.

[0153] C2 (480×270): Edge features from G0, highlighting the right-angled edges of the Magotan's square light strip and the linear features of the lettering. C3 (240×135): Texture features from G1, distinguishing the font texture of the rear badge (the difference between the letters "Magotan" and "Passat" in Magotan). C4 (120×67): Structural features from G2, capturing the relative position of the taillights and license plate frame (the Magotan's taillights are closer to the license plate frame). C5 (60×33): Semantic features from G3, relating to major categories such as "family sedan" and "B-segment car".

[0154] Steps 1-3 perform top-down fusion.

[0155] P5 = Conv1A1(C5): Compressed to 256 channels, retaining the global semantic information of "car".

[0156] P4 = Upsample(P5) + Conv1A1(C4): Integrates the structural features of "taillight-license plate relative position" with global semantics.

[0157] P3 = Upsample(P4) + Conv1A1(C3): Enhances the recognizability of the texture features of the "tail logo font". P2 = Upsample(P3) + Conv1A1(C2): Highlights fine-grained edge features such as "light strip shape" and "letter edges".

[0158] Steps 1-4: Dynamic Weight Fusion

[0159] Calculate the information entropy of each layer: H(P2) = 2.1 (high resolution with rich details), H(P5) = 0.7 (low resolution with strong semantic generalization).

[0160] Weights: α2 = 0.5, α3 = 0.3, α4 = 0.15, α5 = 0.05 (emphasizing fine-grained features).

[0161] Fusion feature: Ffusion=0.5P2+0.3P3+0.15P4+0.05P5, focus on keeping the differentiation of tail logo and lamp group.

[0162] Step 2: Single-view attribute recognition. Extract texture attribute: Magotan letter texture matching degree = 0.92 (Passat = 0.21); extract edge attribute: Magotan tail lamp inside square lamp belt edge corner point number = 4 (Passat circular lamp belt = 0); extract structure attribute: Magotan "tail lamp distance from license plate frame = 8 cm" (Passat = 12 cm); extract semantic attribute: "tail has Magotan logo" confidence = 0.91 (Passat = 0.07).

[0163] Step 3: Multi-view puzzle feature fusion.

[0164] Step 3-1: Multi-view data acquisition. View 1 (directly behind the tail): clearly shoot the tail logo and overall lamp group layout. View 2 (45° right side of the tail): show the side shape of the tail lamp and the connection relationship with the body line. View 3 (30° left side of the tail): present the relative position of the license plate and the left tail lamp.

[0165] Step 3-2: Geometric alignment and attribute fusion. Homography matrix: H1 (view 1→reference system) corrects the tail deformation caused by shooting angle, H2 (view 2→reference system) aligns the right lamp group features. Attribute conflict resolution: "letter number = 6 (Magotan)" in view 1, "letter number = 5 (partially obscured)" in view 3, through similarity weighting p^ = 5.7. Puzzle feature map: arrange the attributes of 3 views according to spatial position as 64x64 heat map, where: tail central area: "Magotan logo" confidence = 0.93. Left tail lamp area: "square lamp belt" confidence = 0.89.

[0166] Step 4: Target model discrimination. Template matching: cosine similarity with Magotan template = 0.91 (Passat = 0.35). Through CNN classification: attribute map input ResNet-50, output Magotan probability = 0.92. Through joint decision: p = 0.3A|0.91+0.7A|0.92 = 0.917 (> 0.85 threshold). Finally, it is identified as Volkswagen Magotan, with a confidence of 0.92, which is 28% higher than traditional methods.

[0167] In a specific implementation, the present application provides a computer storage medium and a corresponding data processing unit, wherein the computer storage medium is capable of storing a computer program, the computer program is executable to run the invention content of the multi-view puzzle type unmanned aerial vehicle video target recognition method and part or all steps in each embodiment when executed by the data processing unit. The storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), a random access memory (RAM), or the like.

[0168] Those skilled in the art can clearly understand that the technical solutions in the embodiments of the present application can be realized by means of a computer program and its corresponding general hardware platform. Based on such understanding, the technical solutions in the embodiments of the present application can be embodied in the form of a computer program, i.e., a software product, which can be stored in a storage medium and includes a plurality of instructions for causing a device (which can be a personal computer, a server, a single-chip microcomputer, an MCU, or a network device) containing a data processing unit to execute the method described in each embodiment or some parts of the embodiments of the present application.

[0169] The present application provides a multi-view puzzle type unmanned aerial vehicle video target recognition method, and there are many methods and approaches to realize the technical solutions. The above description is only the preferred embodiment of the present application. It should be pointed out that for ordinary skilled in the art, without departing from the principles of the present application, some improvements and refinements can be made, which should be considered as the protection scope of the present application. The components not explicitly described in the embodiments can be realized by using existing technologies.

Claims

1. A multi-view puzzle drone video target recognition method, characterized in that, The method comprises the following steps: Step 1: Progressive multi-attribute feature extraction is performed on the current image frame in the unmanned aerial vehicle video to obtain multi-scale deep fusion features of the current image frame; Step 2: A single-view attribute recognition model is constructed, and classification recognition is performed based on the multi-scale deep fusion features extracted in step 1 to obtain a single-view attribute vector under the current view angle; Step 3: The next image frame in the unmanned aerial vehicle video is taken as the current image frame, and steps 1 to 2 are repeatedly executed; Single-view attribute vectors under different view angles are obtained; Step 4: The single-view attribute vectors under different view angles are subjected to multi-view jigsaw feature fusion to obtain jigsaw image fusion features of the target; Step 5: Target model discrimination is performed according to the jigsaw image fusion features of the target by using a weighted voting and Bayesian network joint decision method to complete multi-view jigsaw unmanned aerial vehicle video target recognition. 2.The multi-view puzzle unmanned aerial vehicle video target recognition method of claim 1, wherein, The progressive multi-attribute feature extraction in step 1 comprises: Step 1-1: Multi-scale analysis is performed by using a 5-layer Gaussian pyramid to generate a 5-layer Gaussian pyramid; Step 1-2: A feature pyramid network is constructed to extract feature maps from the images of each layer of the 5-layer Gaussian pyramid; Step 1-3, the final multi-scale depth fusion feature P is calculated fusion is represented as follows: wherein α l is a normalized weight coefficient of the feature map P l , and the calculation method is as follows: where exp denotes an exponential function, H(P l ) denotes the information entropy of the feature map P l , and the calculation method is as follows: where p l,c is the probability distribution of pixel values for the cth channel of the 1st layer.

3. The multi-view puzzle unmanned aerial vehicle video target recognition method according to claim 2, characterized in that, The generation of the 5-layer Gaussian pyramid in step 1-1 is specifically as follows: In the 5-level Gaussian pyramid, the l-th level pyramid image G l (x, y) is generated from the l-1-th level image G l-1 Gaussian filtered and down-sampled by a factor of 2, and is represented as follows: where (x, y) is the pixel coordinate of the l-th layer image, G l-1 (2x+i,2y+j) represents the pixel value of the corresponding position in the l-1-th layer image, where 2x+i and 2y+j are the coordinates of the previous layer image, that is, the key information of the even rows or columns is reserved by 2 times down-sampling, and W(i,j) is a Gaussian kernel weight matrix, which is represented as follows: In the generated 5-layer Gaussian pyramid, G0 is the original image.

4. The multi-view puzzle unmanned aerial vehicle video target recognition method according to claim 3, characterized in that, The extraction of the feature maps from the images of each layer of the 5-layer Gaussian pyramid in step 1-2 is specifically as follows: 1 / 4 resolution feature C2 is extracted from the image G0, 1 / 8 resolution feature C3 is extracted from the image G1, 1 / 16 resolution feature C4 is extracted from the image G2, and 1 / 32 resolution feature C5 is extracted from the image G3; Feature fusion is performed from top to bottom and is represented as follows: P5 = Conv1x1(C5) P4 = Upsample(P5) + Conv1x1(C4) P3 = Upsample(P4) + Conv1x1(C3) P2 = Upsample(P3) + Conv1x1(C2) Wherein, P2, P3, P4 and P5 represent the feature fusion results of each layer, Upsample is a bilinear interpolation up-sampling operation, and Conv1x1 represents a 1*1 convolution operation for adjusting the channel number of the feature map.

5. The multi-view puzzle unmanned aerial vehicle video target recognition method according to claim 4, characterized in that, The classification recognition in step 2 comprises: Step 2-1: The attribute set of the current image frame is divided into M levels; Step 2-2, design a special classifier C for the m-th level attribute m , input multi-scale deep fusion features P fusion , output the confidence s of the level attribute m , is expressed as follows: s m = C m (P fusion ; θ m ) where θ m are parameters of the mth level specialized classifier, m = 1, 2, …, M; Step 2-3, compute the mth component V of the single view attribute vector V of the current image frame m is represented as follows: V m = γ m * s m where γ m is a hierarchical attribute weight satisfying: Step 2-4: The attribute vector set V of the current image frame, i.e., the single-view attribute vector under the current view angle, is obtained and is represented as follows: V = {V1, V2,..., V M ,}.

6. The multi-view puzzle unmanned aerial vehicle video target recognition method according to claim 5, characterized in that, The multi-view jigsaw feature fusion in step 4 comprises: Step 4-1, let the single view attribute vector of the kth view be V k , is expressed as follows: V k = [V k,1 , V k,2 ,..., V k,M ] Step 4-2, aligning the matrix A i The monoscopic attribute vector V k mapped to the uniform coordinate system, is represented as follows: V' k = A i *V k where the alignment matrix A i is a diagonal matrix, the diagonal elements representing spatial scaling coefficients; Step 4-3: The positive correlation between the weight and the attribute is set and is represented as follows: wherein w k,m represents the normalized weight of the mth component of the attribute vector under the kth view angle, V k,m represents the mth component of the single-view attribute vector of the kth view angle; Step 4-4: Jigsaw fusion is performed.

7. The multi-view puzzle unmanned aerial vehicle video target recognition method according to claim 6, characterized in that, The jigsaw fusion in step 4-4 is specifically as follows: where F k denotes the component of the kth view of the complete attribute vector F after puzzle fusion. The jigsaw image fusion features of the target are represented as follows: F = [F1, F2,..., F K ] Wherein, K is the number of view angles.

8. The multi-view puzzle unmanned aerial vehicle video target recognition method according to claim 7, characterized in that, The target model discrimination in step 5 comprises: Step 5-1: The weighted prediction probability P(d) of each model d of the target is calculated; Step 5-2, determining a preliminary decision result according to the weighted prediction probability P(d) is represented as follows: That is, the model d corresponding to the maximum probability P(d) is selected as the preliminary discrimination result; Step 5-3: Posterior verification is performed by using a Bayesian network.

9. The multi-view puzzle unmanned aerial vehicle video target recognition method according to claim 8, characterized in that, The weighted prediction probability P(d) of each model d of the target calculated in step 5-1 is expressed as follows: where K is the number of views, w k is the weight of the kth view, P k (d) is the predicted probability of the kth view for model d.

10. The multi-view puzzle drone video target recognition method of claim 9, wherein, The posteriori verification using the Bayesian network in step 5-3 is specifically as follows: In step 5-3-1, the prior probability P(d) of each model appearing is obtained according to historical data or domain knowledge; In step 5-3-2, the likelihood probability P(w|d) is calculated, which represents the probability of each view feature combination w appearing under the condition that the model is d; In step 5-3-3, the posteriori probability P(d|w) is calculated according to the Bayesian formula, which is expressed as follows: Wherein, P(d) is the prior probability, and P(w) is the evidence probability; In step 5-3-4, a threshold τ is set, and when P(d|w)>τ, the discrimination result is confirmed.

Images