An AI vision-based unmanned aerial vehicle intrusion detection and identification method

Abstract

The application discloses an unmanned aerial vehicle intrusion detection and identification method based on AI vision, relates to the field of AI vision technology, and comprises the following steps: collecting a video frame sequence of a monitoring camera device, and obtaining a to-be-processed frame sequence through preprocessing; performing pixel displacement estimation and calculating a motion energy graph to obtain a candidate image block sequence; constructing an improved SqueezeNet network to obtain an unmanned aerial vehicle candidate target set; generating a target trajectory set through layer-by-layer iterative minimum potential energy search; performing Grubbs test to output a purified trajectory set; constructing a continuous time position prediction curve to generate an intrusion event set; generating an evidence bag and outputting alarm information, and sending the alarm information to a higher-level platform. The unmanned aerial vehicle intrusion detection and identification are realized by combining the improved SqueezeNet network and the Grubbs test.

Description

Technical Field

[0001] This invention relates to the field of AI vision technology, and in particular to a method for drone intrusion detection and identification based on AI vision. Background Technology

[0002] In existing low-altitude security scenarios, drone intrusion detection typically relies on video surveillance, radar, radio detection, or multi-sensor fusion. Vision-based solutions offer advantages such as flexible deployment, scalable coverage, relatively low cost, and the ability to directly output target appearance and evidence information, making them widely used for low-altitude surveillance in key areas such as industrial parks, airports, power facilities, borders, and government agencies. However, vision-based solutions still face technical bottlenecks in practical applications. Long-range drones exhibit weak target features in images, characterized by small pixels, low contrast, and lack of detail. These features are easily superimposed on background interference from cloud textures, building edges, tree canopy movement, and light spot reflections. Furthermore, factors such as camera shake, gimbal rotation, focal length changes, video compression artifacts, low-light conditions at night, and haze can cause unstable changes in target appearance and motion characteristics. This leads to false positives, false negatives, and category confusion in complex scenarios using traditional motion detection, background modeling, simple feature-based classification, or general deep learning models, making it difficult to meet the engineering requirements for stable detection and accurate identification.

[0003] Current visual intrusion detection processes primarily rely on single-frame detection results for alarm judgment, lacking robust modeling for cross-frame consistency and effective strategies for identifying and eliminating abnormal observations. In complex backgrounds, birds, kites, reflective points, or noise interference targets may generate short-term high-confidence detections within a frame, triggering unnecessary alarms. Under occlusion, backlighting, or low-light conditions, missed keyframe detections can lead to trajectory breaks, target loss, and unreported intrusion events, affecting the integrity of the evidence chain. Regarding electronic fence boundary crossing judgment and risk assessment, some solutions only employ simple area entry judgments or fixed rules, failing to unify the modeling of continuous trajectory prediction, fence distance evolution, risk accumulation processes, and event generation logic. They also fail to organize keyframes, trajectories, judgment records, and anomaly elimination records into a structured evidence package, resulting in inexplicable alarms, incomplete evidence, and difficulty in verification.

[0004] Therefore, how to provide a drone intrusion detection and identification method based on AI vision is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an AI vision-based method for drone intrusion detection and identification. This invention comprehensively utilizes multi-source spatiotemporal information processing of video frames, an improved SqueezeNet network structure innovation, and Grubbs' test anomaly removal technology to form a complete process from video frame acquisition and preprocessing, sparse optical flow motion energy candidate region generation, target detection and identification, bidirectional spatiotemporal cost map trajectory generation, outlier observation purification of curvature flow anomaly circles and hypergraph energy spectrum boundaries, continuous time position prediction and intrusion event generation of electronic fence risk integral, evidence packet encapsulation, and alarm reporting. The improved SqueezeNet network is enhanced by connecting a spectral-spatial double-layer reconstruction network in series at the input end, connecting an accompanying ladder network to converge current in the middle feature layer, and setting a reverse trajectory repair branch at the output end, thereby improving the stable detection and identification capabilities under complex backgrounds and long-distance small target conditions. The Grubbs' test achieves purification of false detections and broken observations through curvature measurement iterative evolution and energy spectrum boundary, realizing a closed loop of low false alarms, low false negatives, low latency intrusion alarms and traceable evidence collection under the conditions of limited end-side computing power and network.

[0006] A drone intrusion detection and identification method based on AI vision according to an embodiment of the present invention includes: The video frame sequence output by the monitoring camera is acquired, and the video frame sequence is preprocessed to obtain the frame sequence to be processed; Pixel displacement estimation is performed on the frame sequence to be processed to generate a sparse optical flow vector field. Based on the optical flow vector field, the motion energy map is calculated, and an adaptive quadtree split-merge strategy is used to obtain a candidate image block sequence. An improved SqueezeNet network is constructed to extract features and detect and identify candidate image block sequences. Spectral domain transformation and alternating iterative reconstruction are performed based on the spectral-space double-overlap reconstruction network. Quasi-temporal evolution and adjoint gradient solution are performed through the adjoint ladder network convergence. Inverse trajectory repair is introduced to perform fracture completion, and a set of UAV candidate targets is obtained. A bidirectional spatiotemporal cost graph is established for the set of candidate UAV targets. Node pairing is completed on the bidirectional spatiotemporal cost graph through a layer-by-layer iterative minimum potential energy search to generate a set of target trajectories. The Grubbs test is performed on the target trajectory set, curvature measurement is calculated and curvature flow iterative evolution is performed based on the curvature flow anomaly circle, energy spectrum decomposition is performed through the hypergraph energy spectrum boundary, and the purified trajectory set is output. A continuous-time location prediction curve is constructed based on the set of purification trajectories. A risk gradient sequence is generated by calculating the distance point by point. The risk gradient sequence is then integrated to generate a set of intrusion events. Generate an evidence package from the set of intrusion events and output alarm information, then send the alarm information and evidence package to the upper platform.

[0007] Optionally, the video frame sequence includes a frame image matrix, frame timestamps, camera intrinsic parameters, camera extrinsic parameters, gimbal attitude parameters, and geolocation information.

[0008] Optionally, obtaining the sequence of frames to be processed includes: Acquire the video frame sequence output by the monitoring camera at continuous sampling time, write a frame timestamp for each frame, and read the camera intrinsic parameters, camera extrinsic parameters, PTZ attitude parameters and geolocation information corresponding to the current frame to establish a multi-source data alignment table indexed by the frame timestamp. Image preprocessing is performed on each frame in the video frame sequence. Image preprocessing includes normalizing the size of the frame image matrix, normalizing the pixel intensity, and denoising. Invalid frames are removed to generate a preprocessed frame sequence. The preprocessed frame sequence is written into the frame buffer in the order of frame timestamps and the frame sequence to be processed is output. At the same time, the binding relationship between the camera intrinsic parameters, camera extrinsic parameters, gimbal attitude parameters and geolocation information corresponding to each frame in the frame sequence to be processed is recorded in the frame buffer.

[0009] Optionally, obtaining the candidate image patch sequence includes: The frame image matrices of two adjacent frames in the frame sequence to be processed are grayscaled, a multi-resolution image pyramid is constructed, and the pixel displacement is solved at each resolution level under the conditions of constant brightness constraint and spatial smoothness constraint to obtain the sparse optical flow vector field. The sparse optical flow vector field is then back-projected to the original resolution coordinate system according to the pyramid level. The motion energy map is calculated based on the sparse optical flow vector field, and connected component generation is performed on the motion energy map to obtain a set of motion connected components. Adaptive quadtree splitting and merging are performed on the set of motion connected regions. Quadtree splitting is triggered when the statistical dispersion of motion energy in the current region exceeds a preset dispersion condition, and quadtree merging is triggered when the statistical difference of motion energy between adjacent sub-regions does not exceed a preset difference condition. The set of candidate motion regions is then output. For each candidate motion region in the candidate motion region set, calculate the bounding rectangle. After expanding the bounding rectangle in the width and height directions according to the expansion ratio, crop the corresponding image region to obtain the candidate image block sequence.

[0010] Optionally, obtaining the candidate target set for drones includes: An improved SqueezeNet network is constructed by connecting a spectral-space double-layer reconstruction network between the input of the original SqueezeNet network and the first feature extraction layer, connecting an accompanying ladder network convergence network between the original intermediate feature output layer and the detection and recognition output layer, and setting a reverse trajectory repair branch at the original detection and recognition output. The candidate image block sequence is input into the input end of the original SqueezeNet network in the order of frame timestamps. The output is a frame-level intermediate feature sequence. The frame-level intermediate feature sequence is input into the spectral-space double-overlay reconstruction network. The spectral-space double-overlay reconstruction network performs orthogonal basis transformation in the time dimension and spatial dimension respectively to obtain the spectral domain coefficient sequence. The spectral domain coefficient sequence is subjected to alternating iterative processing of energy clipping, inverse transformation reconstruction and minimum difference update in each round to output the background suppression feature sequence. The background suppression feature sequence is input into the adjoint ladder network convergence flow. The adjoint ladder network convergence flow represents each frame feature in the background suppression feature sequence as a two-dimensional discrete field. The two-dimensional discrete field is subjected to pseudo-time forward evolution according to the evolution step size to obtain the evolution feature field. An adjoint gradient field is established for the evolution feature field and an adjoint inverse solution is performed. The evolution feature field is iteratively updated according to the adjoint gradient field and the target focusing feature sequence is output. The target-focused feature sequence is input into the detection and recognition output layer of the original SqueezeNet network, which outputs a set of candidate detection results corresponding to the timestamps of each frame. The set of candidate detection results is input into the reverse trajectory repair branch. The reverse trajectory repair branch generates a set of trajectory segments based on the set of candidate detection results of adjacent frames, identifies the broken intervals in the set of trajectory segments, performs time reverse consistency verification on the trajectory segments on both sides of the broken interval to generate reverse completion results, writes the reverse completion results into the set of candidate detection results of the frame timestamps corresponding to the broken interval, and outputs a set of candidate targets for the UAV. The improved SqueezeNet network is trained by constructing a training sample set and inputting the candidate image patch sequence of the training sample set into the improved SqueezeNet network. The joint loss function is constructed based on the bounding box regression loss and the category classification loss. Backpropagation is used to calculate the gradient of the joint loss function and update the trainable parameters of the original SqueezeNet network part, the spectral-space double-overlap reconstruction network part and the accompanying ladder network convergence part in the improved SqueezeNet network until the training termination condition is met and the trained improved SqueezeNet network is obtained.

[0011] Optionally, the generation of the target trajectory set includes: A bidirectional spatiotemporal cost graph is constructed based on the set of UAV candidate targets. The nodes of the bidirectional spatiotemporal cost graph are composed of UAV candidate targets corresponding to the timestamps of each frame. The edges of the bidirectional spatiotemporal cost graph include temporal edges and spatial edges. The edge weights of the bidirectional spatiotemporal cost graph are calculated and the total potential energy value is generated. The temporal edge cost is generated by the center point displacement, scale change and feature vector distance according to the combination rules. The spatial edge potential energy is generated by the overlap of the target boxes in the same frame according to the transformation rules. The total potential energy value is generated by the temporal edge cost and the spatial edge potential energy and corresponds to the inter-frame time difference. A layer-by-layer iterative minimum potential energy search is performed on the bidirectional spatiotemporal cost graph to complete node pairing. The layer-by-layer iterative minimum potential energy search includes selecting the node pair with the smallest total potential energy value as the pairing relationship according to the round, deleting the paired nodes and updating the total potential energy value of the remaining nodes until there are no more paired nodes. Based on the node pairing relationship, a target trajectory set is generated and the trajectory identifier and trajectory point sequence are output.

[0012] Optionally, the set of output purification trajectories includes: For each trajectory in the target trajectory set, construct a sliding window observation sequence and input the sliding window observation sequence into the Grubbs test; Based on the curvature flow anomaly loop, the curvature metric calculation and curvature flow iterative evolution are performed on the sliding window observation sequence. Adjacency relationships are constructed between adjacent observation vectors in the observation vector space and the discrete curvature value of each observation vector is calculated. The discrete curvature value is updated according to the iteration round and a curvature evolution sequence is generated. The observation vectors whose curvature values ​​in the curvature evolution sequence meet the mutation condition are marked as curvature outlier observations. The energy spectrum decomposition of the sliding window observation sequence is performed by the energy spectrum boundary of the hypergraph. The observation vectors in the sliding window observation sequence are used as hypergraph nodes. Hyperedges are constructed to form a higher-order correlated hypergraph based on cross-frame continuity and intra-frame mutual exclusion. The Laplacian matrix of the higher-order correlated hypergraph is calculated and eigenvalue decomposition is performed to obtain the energy spectrum eigenvector. The observation vectors are delineated according to the energy spectrum eigenvector. The observation vectors falling on the weakly connected boundary side are marked as energy spectrum outliers. Curvature outlier observations and energy spectrum outlier observations are merged into an outlier observation set. Based on the outlier observation set, the target trajectory set is updated by outlier removal or replacement, and a clean trajectory set is output.

[0013] Optionally, the generation of the intrusion event set includes: The trajectory point sequence of each purification trajectory in the purification trajectory set is processed to be time continuous according to the frame timestamp order, and a continuous time position prediction curve is generated based on the continuous time interpolation result. A risk gradient sequence is generated by calculating the point-by-point distance of the continuous time location prediction curve. The predicted location point is obtained at the sampling time of the continuous time location prediction curve. The shortest Euclidean distance from the predicted location point to the boundary of each area in the set of electronic fence areas is calculated and the minimum value is taken as the fence distance. At the same time, the shortest Euclidean distance from the predicted location point to each sensitive point in the set of sensitive points is calculated and the minimum value is taken as the point distance. The risk gradient values ​​at the corresponding sampling time are generated based on the fence distance and the point distance and are arranged into a risk gradient sequence in chronological order. The risk gradient sequence is integrated to generate a risk integral sequence. The risk gradient sequence is numerically integrated over the time interval of the sampling time to obtain the cumulative risk integral value. The intrusion status and risk level are determined based on the change process of the cumulative risk integral value. The time intervals in which the cumulative risk integral value of the same trajectory continuously meets the intrusion status condition are merged to generate an intrusion event set.

[0014] Optionally, generating an evidence package from the set of intrusion events and outputting alarm information includes: For each intrusion event in the intrusion event set, an evidence package is generated. The evidence package includes a set of keyframes corresponding to the time of the intrusion event, the target bounding box coordinates and category identifiers of each keyframe in the keyframe set, the trajectory point sequence of the cleanup trajectory set corresponding to the intrusion event, the electronic fence spatial relationship record corresponding to the intrusion event, and outlier observation records. The outlier observation records include the frame timestamps corresponding to the outlier observations, the outlier observation type identifiers, and the observation vectors of the outlier observations. Generate alarm information, which includes the event identifier, area level, risk level, occurrence time, duration, corresponding trajectory identifier, and evidence package index of the intrusion event. The evidence package index is the unique location identifier of the evidence package in the end-side storage. The alarm information and evidence package are encapsulated and sent to the upper platform. The alarm information and evidence package index are written into the alarm message body and the evidence package is written into the local storage according to the file structure. The alarm message body is sent to the upper platform through the network link. The evidence package or evidence package index is synchronized to the upper platform. At the same time, the sending status and retransmission records are recorded on the terminal side.

[0015] The beneficial effects of this invention are: This invention improves the stable detection and recognition capabilities of UAVs under complex backgrounds and long-distance small target conditions by integrating motion energy candidate region generation and edge detection and recognition of video frame sequences. Compared with existing vision solutions that rely on single-frame detection or direct inference from general models, this invention improves the SqueezeNet network by setting up a spectral-spatial double-overlap reconstruction network, an adjoint ladder network convergence, and a reverse trajectory repair branch. This enhances weak target features, suppresses background interference, and completes trajectory breaks caused by short-term missed detections. At the same time, by combining cross-frame correlation of bidirectional spatiotemporal cost maps with target trajectory generation, it can effectively reduce false positives and false negatives, thereby achieving detection and recognition effects with low false positives and low false negatives.

[0016] The closed-loop process of trajectory purification, fence determination, and evidence collection alarm proposed in this invention realizes the reliable generation of intrusion events and the output of traceable evidence, and is suitable for real-time deployment environments with limited edge computing power and network. This invention uses curvature flow anomaly circles and hypergraph energy spectrum boundaries from the Grubbs test to identify and remove outliers from the trajectory sliding window observation sequence, obtaining a purified trajectory set. A risk gradient sequence is generated based on continuous-time position prediction curves and electronic fence distance fields, and integrated to form an intrusion event set, ensuring that alarm determination and risk classification are consistent with trajectory evolution. By encapsulating keyframes, trajectories, fence spatial relationship records, and outlier observation records in an evidence package, the alarm information and evidence package are sent to the upper-level platform. Even under unstable network conditions, this achieves low-latency alarms, verifiable events, and a complete evidence chain, ensuring successful engineering implementation. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of an AI vision-based drone intrusion detection and recognition method proposed in this invention; Figure 2 This is a block diagram of the improved SqueezeNet network for a drone intrusion detection and recognition method based on AI vision proposed in this invention. Figure 3 This is a functional diagram of the Grubbs test in the drone intrusion detection and identification method based on AI vision proposed in this invention. Detailed Implementation

[0018] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0019] refer to Figure 1 , Figure 2 and Figure 3 A drone intrusion detection and identification method based on AI vision, comprising: The video frame sequence output by the monitoring camera is acquired, and the video frame sequence is preprocessed to obtain the frame sequence to be processed; Pixel displacement estimation is performed on the frame sequence to be processed to generate a sparse optical flow vector field. Based on the optical flow vector field, the motion energy map is calculated, and an adaptive quadtree split-merge strategy is used to obtain a candidate image block sequence. An improved SqueezeNet network is constructed to extract features and detect and identify candidate image block sequences. Spectral domain transformation and alternating iterative reconstruction are performed based on the spectral-space double-overlap reconstruction network. Quasi-temporal evolution and adjoint gradient solution are performed through the adjoint ladder network convergence. Inverse trajectory repair is introduced to perform fracture completion, and a set of UAV candidate targets is obtained. A bidirectional spatiotemporal cost graph is established for the set of candidate UAV targets. Node pairing is completed on the bidirectional spatiotemporal cost graph through a layer-by-layer iterative minimum potential energy search to generate a set of target trajectories. The Grubbs test is performed on the target trajectory set, curvature measurement is calculated and curvature flow iterative evolution is performed based on the curvature flow anomaly circle, energy spectrum decomposition is performed through the hypergraph energy spectrum boundary, and the purified trajectory set is output. A continuous-time location prediction curve is constructed based on the set of purification trajectories. A risk gradient sequence is generated by calculating the distance point by point. The risk gradient sequence is then integrated to generate a set of intrusion events. Generate an evidence package from the set of intrusion events and output alarm information, then send the alarm information and evidence package to the upper platform.

[0020] In this embodiment, the video frame sequence includes a frame image matrix, frame timestamps, camera intrinsic parameters, camera extrinsic parameters, gimbal attitude parameters, and geolocation information.

[0021] In this embodiment, obtaining the frame sequence to be processed includes: Acquire the video frame sequence output by the monitoring camera at continuous sampling time, write a frame timestamp for each frame, and read the camera intrinsic parameters, camera extrinsic parameters, PTZ attitude parameters and geolocation information corresponding to the current frame to establish a multi-source data alignment table indexed by the frame timestamp. Image preprocessing is performed on each frame in the video frame sequence. Image preprocessing includes normalizing the size of the frame image matrix, normalizing the pixel intensity, and denoising. Invalid frames are removed to generate a preprocessed frame sequence. The preprocessed frame sequence is written into the frame buffer in the order of frame timestamps and the frame sequence to be processed is output. At the same time, the binding relationship between the camera intrinsic parameters, camera extrinsic parameters, gimbal attitude parameters and geolocation information corresponding to each frame in the frame sequence to be processed is recorded in the frame buffer.

[0022] In this embodiment, obtaining the candidate image patch sequence includes: The frame image matrices of two adjacent frames in the frame sequence to be processed are converted to grayscale, and a multi-resolution image pyramid is constructed. At each resolution level, pixel displacement is solved under constant brightness constraints and spatial smoothness constraints to obtain a sparse optical flow vector field. This sparse optical flow vector field is then back-projected onto the original resolution coordinate system according to the pyramid levels, where: To perform grayscale conversion and construct a multi-resolution image pyramid, the following steps are taken: For the frame image matrices of two adjacent frames, the red, green, and blue channel values ​​of each pixel are taken respectively, and the single-channel grayscale value is calculated by linear combination. The grayscale value calculation rule is 0.299 multiplied by the red channel value, plus 0.587 multiplied by the green channel value, plus 0.114 multiplied by the blue channel value, to obtain two grayscale images. Each grayscale image is used as the 0th layer. Before generating the next layer, the current layer grayscale image is first Gaussian smoothed. The Gaussian smoothing uses a Gaussian kernel with a standard deviation of 1.0 and a convolution kernel size of 5×5. The smoothed grayscale image is then downsampled by reducing its width and height to 1 / 2 of the original to obtain the next layer grayscale image. The Gaussian smoothing and 1 / 2 downsampling are repeated until a 4-layer pyramid is obtained. The pixel displacement is solved under constant brightness constraints and spatial smoothness constraints to obtain a sparse optical flow vector field. Specifically, starting from the third layer of the pyramid, the pixel displacement field is solved for adjacent grayscale images. The solution objective consists of two parts: the constant brightness constraint requires that the grayscale difference between the pixel position in the first frame and the corresponding position in the second frame after displacement be minimized; the spatial smoothness constraint requires that the displacement change of adjacent pixels be minimized. The two parts are summed to form the overall cost. The spatial smoothness constraint coefficient is set to 0.05. The pixel displacement is solved using successive linearization and iterative update methods. Each iteration is based on the current position... The displacement estimation performs a local linear approximation on the second frame grayscale image, updates the displacement component of each pixel, and sets the number of iterations per layer to 10. After completing the third layer, the layer displacement field is upsampled by the scale ratio and multiplied by 2 as the initial value of the second layer. The second layer, the first layer, and the 0th layer are iterated and updated in sequence to obtain the displacement field of the 0th layer. The brightness residual of each pixel position is calculated on the displacement field of the 0th layer. The brightness residual is the absolute difference between the grayscale of the first frame and the grayscale of the second frame at the corresponding displacement position. The displacement vectors of the pixel positions with brightness residuals greater than 15 are retained and the displacement vectors of the remaining pixel positions are discarded to obtain the sparse optical flow vector field. The motion energy map is calculated based on the sparse optical flow vector field. Connectivity generation is then performed on the motion energy map to obtain a set of motion connected components, where: The motion energy map is calculated based on the sparse optical flow vector field. Specifically, in the 0th layer coordinate system, all retained pixel positions in the sparse optical flow vector field are traversed. For each retained pixel position, the horizontal displacement component and the vertical displacement component are read, and the motion energy value of the pixel position is calculated. The motion energy value is calculated as the sum of the squares of the horizontal displacement component and the squares of the vertical displacement component. For pixel positions where the displacement vector is not retained, the motion energy value is set to 0. The motion energy values ​​of all pixel positions are written into the spatial coordinates of the frame image matrix to obtain the motion energy map. The motion energy map is then subjected to 3×3 neighborhood median filtering to remove isolated noise points. To generate a set of motion connected components, the motion energy map is binarized. Pixels with motion energy values ​​greater than 9 are marked as foreground pixels, and the remaining pixels are marked as background pixels. Connected components are marked on the binary map using the 8-neighborhood connectivity rule. The entire binary map is scanned, and region growing is performed on each unmarked foreground pixel. All foreground pixels connected to their 8 neighbors are grouped into the same connected component and assigned the same connected component number. The number of pixels in each connected component is counted, and connected components with fewer than 30 pixels are removed. The remaining connected components are used as the set of motion connected components. Adaptive quadtree splitting and merging is performed on the set of motion connected components. Quadtree splitting is triggered when the statistical dispersion of motion energy within the current region exceeds a preset dispersion condition, and quadtree merging is triggered when the statistical difference in motion energy between adjacent sub-regions does not exceed a preset difference condition. A set of candidate motion regions is output. Specifically, the adaptive quadtree splitting and merging of the set of motion connected components is performed as follows: For each motion connected component in the set of motion connected components, take the bounding rectangle as the root region of the quadtree. Read the set of motion energy values ​​of the corresponding pixels in the motion energy map within the root region, and calculate the statistical dispersion of the region. The statistical dispersion is defined as the standard deviation of the motion energy values ​​of the region. When the statistical dispersion is greater than 6, trigger the quadtree splitting. Divide the current region into 4 sub-regions in the horizontal and vertical directions according to the center point. Repeat the reading of motion energy values, calculation of standard deviation and determination of whether to continue splitting for each sub-region until the stopping condition is met. The stopping condition is that the statistical dispersion is not greater than 6 or the width or height of the sub-region is less than 16. After the split is completed, a merging decision is made for adjacent sub-regions. The merging decision is based on the difference in the mean motion energy of adjacent sub-regions. The statistical difference is defined as the absolute difference in the mean motion energy of two sub-regions. When the statistical difference is not greater than 3, the quadtree merging is triggered, the adjacent sub-regions are merged into the parent region, and the mean motion energy and standard deviation of the parent region are updated. The merging decision is repeated until there are no adjacent sub-regions that meet the merging conditions. The regions with no less than 50 foreground pixels in the final retained leaf regions are output as the candidate motion region set. For each candidate motion region in the candidate motion region set, calculate the bounding rectangle. Expand the bounding rectangle in both the width and height directions according to the expansion ratio, then crop the corresponding image region to obtain a candidate image patch sequence. Specifically, the candidate image patch sequence is obtained as follows: For each candidate motion region in the candidate motion region set, iterate through the foreground pixel positions within the region's coordinate range. Take the minimum x-coordinate, minimum y-coordinate, maximum x-coordinate, and maximum y-coordinate of all foreground pixels as the coordinates of the top-left and bottom-right corners of the bounding rectangle. Calculate the width and height of the bounding rectangle. Expand it symmetrically in both the width and height directions according to the expansion ratio, which is set to 1.25. That is, the expanded width is the original width multiplied by 1.25, and the expanded height is the original height multiplied by 1.25. Keep the center point of the bounding rectangle unchanged. Perform boundary clipping on the coordinates of the expanded bounding rectangle so that the x-coordinate and y-coordinate of the top-left corner are not less than 0, the x-coordinate of the bottom-right corner is not greater than the frame image width minus 1, and the y-coordinate of the bottom-right corner is not greater than the frame image height minus 1. Image blocks are obtained by cropping the corresponding frame image matrix according to the coordinates of the extended bounding rectangle. The image blocks are then filled into squares along their long sides and scaled to 224×224 to obtain candidate image blocks. Candidate motion regions with fewer than 400 pixels are not generated as candidate image blocks. All candidate image blocks generated in the order of frame timestamps are combined into a candidate image block sequence.

[0023] In this embodiment, obtaining the candidate target set for UAVs includes: An improved SqueezeNet network is constructed by connecting a spectral-space double-layer reconstruction network between the input of the original SqueezeNet network and the first feature extraction layer, connecting an accompanying ladder network convergence network between the original intermediate feature output layer and the detection and recognition output layer, and setting a reverse trajectory repair branch at the original detection and recognition output. The candidate image patch sequence is input into the original SqueezeNet network in frame timestamp order, and the output is a frame-level intermediate feature sequence. This frame-level intermediate feature sequence is then input into a spectral-spatial double-overlay reconstruction network. The spectral-spatial double-overlay reconstruction network performs orthogonal basis transformations in both the time and spatial dimensions to obtain the spectral domain coefficient sequence. The spectral domain coefficient sequence is then subjected to alternating iterative processing of energy clipping, inverse transformation reconstruction, and minimum difference update in each round, outputting a background suppression feature sequence, where: The output frame-level intermediate feature sequence is as follows: The candidate image block sequence is input into the input end of the original SqueezeNet network in ascending order of frame timestamp. For each candidate image block, forward computation is performed. The forward computation passes through the first convolutional layer, pooling layer and multiple Fire modules of the original network to obtain intermediate feature maps. The preset intermediate feature output layer is set as the output end of the last Fire module in the Fire module group. The intermediate feature map at the output end is read and the number of channels, width and height are retained. When the same frame timestamp corresponds to multiple candidate image blocks, the multiple intermediate feature maps corresponding to the frame timestamp are concatenated according to the channel dimension and synthesized into a single frame-level intermediate feature map by the maximum value of the channel dimension. The frame-level intermediate feature maps corresponding to all frame timestamps are arranged in chronological order to form a frame-level intermediate feature sequence. The orthogonal basis transformation is performed to obtain the spectral domain coefficient sequence. Specifically, for each frame-level intermediate feature map in the frame-level intermediate feature sequence, an orthogonal basis transformation is performed in the time dimension. Using the feature maps corresponding to the timestamps of consecutive frames as the input sequence, a discrete cosine transform is performed on the time series of the same spatial location and the same channel to obtain the time spectral coefficients. An orthogonal basis transformation is performed in the spatial dimension. For each channel feature map of each frame, a discrete cosine transform is performed in the width and height directions to obtain the spatial spectral coefficients. The time spectral coefficients and spatial spectral coefficients are organized into a spectral domain coefficient sequence according to the channel and frequency index, and the channel number, time frequency number and spatial frequency number corresponding to each coefficient in the spectral domain coefficient sequence are recorded. The spectral coefficient sequence is processed by alternating iterative steps of energy pruning, inverse transform reconstruction, and minimum difference update in each round. Specifically, the number of iterations for the spectral coefficient sequence is set to 5 rounds. In each round, the energy value corresponding to each spectral coefficient is calculated. The energy value is calculated as the square of the absolute value of the spectral coefficient. The energy values ​​are sorted by channel and the cumulative energy ratio is calculated. The spectral coefficients with a cumulative energy ratio of less than 0.95 are retained, and the remaining spectral coefficients are set to 0 to complete the energy pruning. The spatial dimension inverse discrete cosine transform and the time dimension inverse discrete cosine transform are performed on the pruned spectral coefficient sequence to obtain the reconstructed feature sequence of the current round. The difference between the reconstructed feature sequence and the input feature sequence of the previous round is calculated at each element position. The sum of the squares of the differences is calculated as the reconstruction error. The retained spectral coefficients are updated with the goal of reducing the reconstruction error. The update rule is to adjust the spectral coefficient value at each spectral coefficient by a preset step size of 0.1 in the direction of reducing the reconstruction error. The updated spectral coefficient sequence is used as the input of the next round. After 5 rounds of iteration, the final reconstructed feature sequence is output as the background suppression feature sequence. The background suppression feature sequence is input into the adjoint ladder network convergence flow. The adjoint ladder network convergence flow represents each frame feature in the background suppression feature sequence as a two-dimensional discrete field. The two-dimensional discrete field is then subjected to quasi-temporal forward evolution according to the evolution step size to obtain the evolved feature field. An adjoint gradient field is established on the evolved feature field, and an adjoint inverse solution is performed. The evolved feature field is iteratively updated based on the adjoint gradient field, and the target focusing feature sequence is output, where: Each frame feature in the background suppression feature sequence is represented as a two-dimensional discrete field. The two-dimensional discrete field is then subjected to quasi-temporal forward evolution according to the evolution step size to obtain the evolved feature field. Specifically, for the feature map of each frame in the background suppression feature sequence, a two-dimensional grid coordinate is established according to the spatial width and height. The channel vector corresponding to each grid point is used as the field value of the grid point. The frame feature is represented as a two-dimensional discrete field. The two-dimensional discrete field is subjected to forward evolution according to the quasi-temporal step size, with the evolution step size set to 0.05 and the number of evolution iterations set to 6. In each iteration, the first-order difference gradient of the two-dimensional discrete field is calculated in the horizontal and vertical directions respectively. Then, the second-order difference is calculated to obtain the discrete Laplace term. The field value is updated based on the discrete Laplace term. The update method is to add the current field value to the evolution step size multiplied by the discrete Laplace term to obtain the evolved feature field. The evolutionary feature field is obtained by performing quasi-temporal forward evolution on the two-dimensional discrete field according to the evolution step size. An adjoint gradient field is then established on the evolutionary feature field and an adjoint backward solution is performed. Specifically, after completing one quasi-temporal forward evolution, a supervision error oriented towards the evolutionary feature field is constructed and a joint error value is generated. The joint error value is obtained by summing the target box position error and the category error. The partial derivative of the joint error value with respect to the evolutionary feature field is used as the initial value of the adjoint gradient field. The adjoint gradient field is then solved backward according to the reverse order of the quasi-temporal evolution. During the backward solution, the corresponding discrete Laplace operator is reconstructed at each quasi-temporal step. A backward recursive update is performed on the adjoint gradient field. The update method is to subtract the evolution step size multiplied by the result of the discrete Laplace operator acting on the current adjoint gradient field from the current adjoint gradient field to obtain the adjoint gradient field corresponding to each quasi-temporal step. The evolutionary feature field is iteratively updated based on the adjoint gradient field, and the target focusing feature sequence is output. Specifically, the adjoint gradient field of the initial pseudo-time step obtained by inverse solving is used as the update direction to perform iterative updates on the evolutionary feature field. The number of iterations is set to 4, and the update step size is set to 0.02. In each update, the channel vector of each grid point in the evolutionary feature field is subtracted by the update step size multiplied by the adjoint gradient vector of the corresponding grid point. The updated channel vector is pruned to make the value fall within -6 to 6. The updated evolutionary feature field is used as a new input to repeat the pseudo-time forward evolution and adjoint inverse solving once. After completing 4 updates, the final evolutionary feature field is output and the target focusing feature sequence is formed in the order of frame timestamps. The target-focused feature sequence is input into the detection and recognition output layer of the original SqueezeNet network, which outputs a set of candidate detection results corresponding to the timestamps of each frame. This set of candidate detection results is then input into the reverse trajectory repair branch. The reverse trajectory repair branch generates a set of trajectory segments based on the candidate detection result sets of adjacent frames, identifies break intervals in the trajectory segment set, performs reverse temporal consistency checks on the trajectory segments on both sides of the break interval to generate reverse completion results, and writes the reverse completion results into the candidate detection result set corresponding to the timestamp of the break interval. The result outputs a set of candidate targets for the UAV, where: The output is a set of candidate detection results corresponding to each frame timestamp. Specifically, the target focusing feature sequence is input into the detection and recognition output layer of the original SqueezeNet network in the order of frame timestamps. For each frame feature, forward computation is performed to obtain the frame's category score vector and target box regression vector. The category score vector is normalized to obtain the confidence of each category, and categories with a confidence greater than 0.4 are selected as frame candidate categories. The target box regression vector is decoded to obtain the coordinates of the bounding rectangle. The bounding rectangle coordinates include the x-coordinate of the upper left corner, the y-coordinate of the upper left corner, the x-coordinate of the lower right corner, and the y-coordinate of the lower right corner. When multiple candidate boxes are generated in the same frame, the intersection-union ratio (IUU) of each candidate box is calculated and overlap suppression is performed. For candidate boxes with an IUU greater than 0.5, only those with higher confidence are retained. The corresponding feature vector is extracted for each retained candidate box and bound to the candidate box to form a set of candidate detection results corresponding to the frame timestamp. A trajectory segment set is generated based on the candidate detection result set of adjacent frames. Specifically, candidate association pairs are established for the candidate detection result sets of two adjacent frame timestamps. The center point coordinates and width and height of each candidate box are calculated. The matching cost between the candidate box and the candidate box of the next frame is calculated. The matching cost is composed of the Euclidean distance of the center point, the rate of change of width and height, and the cosine distance of the feature vector in an additive manner. The upper limit of the Euclidean distance of the center point is set to 40 pixels. Candidate pairs exceeding 40 pixels do not participate in matching. Among the candidate pairs that meet the upper limit of the center point distance, the candidate pair with the smallest matching cost is selected as the association relationship. Each candidate box participates in at most one association relationship. The sequence of candidate boxes connected by association relationships in multiple consecutive frames is defined as a trajectory segment. A segment identifier is assigned to each trajectory segment to obtain a trajectory segment set. The system identifies broken sections in the trajectory segment set, performs reverse consistency checks on the trajectory segments on both sides of the broken section, and generates reverse completion results. Specifically: The trajectory segment set is sorted by frame timestamp. If the same segment identifier has missing frames between two adjacent frame timestamps and the number of missing frames is between 1 and 5, the range of missing frames is marked as a break interval. The end velocity vector and end scale change rate are calculated for the trajectory segments on both sides of the break interval. The end velocity vector is obtained by removing the center point of the last two frames by the time interval. The end scale change rate is obtained by dividing the width and height change of the last two frames by the width and height of the previous frame. The trajectory segments behind the break interval are pushed back in time. The center point position is compensated for in reverse displacement and the width and height are compensated for in reverse scale according to the time interval of each missing frame. The reverse-pushed prediction box sequence is obtained. For each predicted box sequence, a consistency index is calculated with the end state of the trajectory segment preceding the fracture interval. The consistency index is composed of three conditions that must be met simultaneously: the distance between the center point is no more than 30 pixels, the change rate of width and height is no more than 0.30, and the cosine distance of the feature vector is no more than 0.20. When the consistency index is met, the predicted box sequence is used as the reverse completion result. The completion result is assigned the segment identifier of the trajectory segment following the fracture interval. The reverse completion result is written into the candidate detection result set of the frame timestamp corresponding to the fracture interval, and the UAV candidate target set is output. The improved SqueezeNet network is trained by constructing a training sample set and inputting candidate image patch sequences from the training sample set into the improved SqueezeNet network. A joint loss function is constructed based on the bounding box regression loss and the category classification loss. Backpropagation is used to calculate the gradient of the joint loss function and update the trainable parameters of the original SqueezeNet network part, the spectral-space double-overlap reconstruction network part, and the accompanying ladder network convergence part in the improved SqueezeNet network until the training termination condition is met, resulting in the trained improved SqueezeNet network. Where: The training sample set is constructed as follows: raw video frame sequences containing different distances, lighting conditions, backgrounds, and weather conditions are collected from drone intrusion monitoring scenarios. Images are extracted frame by frame, and the bounding box coordinates and category identifier of the drone target are labeled for each frame. At the same time, background samples without drones are labeled. Candidate image block sequences are generated for the labeled frames. Each candidate image block is matched with the labeled box of the source frame. The matching rule is that when the intersection-union ratio (IU) of the bounding box and the labeled box of the candidate image block is greater than 0.5, the candidate image block is labeled as a positive sample and the corresponding target box coordinates and category identifier are written. When the IU is less than 0.20, the candidate image block is labeled as a negative sample and only the background category identifier is written. Candidate image blocks with an IU between 0.20 and 0.50 are not included in the training. Positive and negative samples are sampled at a ratio of 1:3 to form a training sample set. The training sample set is divided into a training subset and a validation subset at a ratio of 8:2. The joint loss function is constructed based on the target box regression loss and the category classification loss. Specifically, for each training sample, the network output is used to obtain the predicted coordinates of the target box and the category confidence. The target box regression loss is composed of the difference between the predicted target box coordinates and the labeled target box coordinates. The absolute differences of the top left x-coordinate, top left y-coordinate, bottom right x-coordinate, and bottom right y-coordinate are calculated and summed. The category classification loss is composed of the category confidence and the true category label. The negative logarithm of the confidence corresponding to the true category is taken as the single-sample classification loss. The summation is applied to all categories to obtain the multi-class classification loss. The target box regression loss and the category classification loss are combined to form the joint loss function. The target box regression loss coefficient is 0.5, and the category classification loss coefficient is 1.5. The batch joint loss value is obtained by averaging the joint loss of all samples in a batch. The training continues until the termination condition is met and the improved SqueezeNet network is obtained. Specifically, the improved SqueezeNet network is input into batches of training subsets and batch joint loss values ​​are calculated. Backpropagation is used to calculate the gradient of the batch joint loss values ​​and update the trainable parameters of the original SqueezeNet network, the spectral-space double-overlap reconstruction network, and the accompanying ladder network convergence network. The parameter update uses gradient descent with a learning rate of 0.001 and the gradient norm is pruned with an upper limit of 5. After each training epoch, the validation joint loss value and the average detection accuracy index are calculated on the validation subset. The training termination condition is triggered when the validation joint loss value does not decrease for 5 consecutive epochs or the average detection accuracy index increases by less than 0.005 for 5 consecutive epochs. At the same time, the upper limit of the training epochs is set to 120. Training stops when any termination condition is met and the current parameters are output as the improved SqueezeNet network that has been trained.

[0024] In this embodiment, generating the target trajectory set includes: A bidirectional spatiotemporal cost graph is constructed based on the set of UAV candidate targets. The nodes of the bidirectional spatiotemporal cost graph consist of UAV candidate targets corresponding to the timestamps of each frame. The edges of the bidirectional spatiotemporal cost graph include temporal edges and spatial edges. The construction of the bidirectional spatiotemporal cost graph based on the set of UAV candidate targets is as follows: Based on the candidate target set of UAVs, a temporal index is established in ascending order of frame timestamps. For each candidate target within each frame timestamp, a node is generated, and node attributes are written for the node. The node attributes include the coordinates of the target box center point, the target box width, the target box height, the confidence score, and the feature vector. A temporal edge is established between adjacent frame timestamps for cross-frame node pairs. The connection rules for the temporal edge are that the Euclidean distance between the center points of the previous frame node and the next frame node is no greater than 60 pixels and the relative change in the target box width and height is no greater than 0.50. A temporal edge is established for node pairs that meet the connection rules, and the corresponding previous frame node identifier and next frame node identifier are recorded. Within the same frame timestamp, spatial edges are established for intra-frame node pairs. The connection rule for spatial edges is that spatial edges are established for any node pair in the same frame, and the intersection-union ratio of the target boxes is recorded. The intersection-union ratio is calculated by dividing the intersection area of ​​the two target boxes by the union area. All nodes, temporal edges, and spatial edges are organized into a bidirectional spatiotemporal cost graph according to the node identifier and frame timestamp. Edge weights are calculated for the bidirectional spatiotemporal cost graph, and a total potential energy value is generated. The temporal edge cost is generated by combining the center point displacement, scale change, and feature vector distance according to a combination rule. The spatial edge potential energy is generated by the overlap of the target boxes within the same frame according to a transformation rule. The total potential energy value is generated from the temporal edge cost and the spatial edge potential energy and corresponds to the inter-frame temporal difference, where: The temporal edge cost is generated by combining the center point displacement, scale change, and feature vector distance according to a combination rule; the spatial edge potential energy is generated by the overlap of the target boxes in the same frame according to a transformation rule. Specifically: For each time-directed edge, read the center point coordinates of the target bounding boxes of the previous frame node and the next frame node, calculate the center point displacement, which is the Euclidean distance of the difference between the two center point coordinates. Read the width and height of the target bounding boxes of the two nodes, and calculate the relative change in width and relative change in height respectively. The relative change in width is the absolute value of the difference between the two widths divided by the width of the previous frame, and the relative change in height is the absolute value of the difference between the two heights divided by the height of the previous frame. The scale change is the average of the relative change in width and relative change in height. Read the feature vectors of the two nodes and calculate the feature vector distance, which is 1 minus the cosine similarity of the two feature vectors. Divide the center point displacement by 60 to obtain the displacement normalization value, divide the scale change by 0.50 to obtain the scale normalization value, and define the temporal edge cost as the displacement normalization value plus the scale normalization value plus the feature vector distance. For each spatial edge, read the target box coordinates of the two nodes in the same frame, calculate the target box overlap, which is the intersection-union ratio, and define the spatial edge potential energy as the square of the target box overlap. The total potential value is generated by the temporal edge cost and the spatial edge potential energy and corresponds to the inter-frame time difference. Specifically, for each temporal edge, the corresponding previous frame timestamp and next frame timestamp are read, the inter-frame time difference is calculated, which is the next frame timestamp minus the previous frame timestamp, the temporal edge cost is multiplied by the inter-frame time difference to obtain the temporal weighted cost, the spatial edge potential energy introduced in the same frame is considered for each candidate pairing relationship, the maximum value of the spatial edge potential energy related to the pairing relationship is taken as the mutual exclusion potential energy, the total potential energy value is defined as the temporal weighted cost plus the mutual exclusion potential energy, the total potential energy value of the candidate pairing relationship is obtained, and the total potential energy value is bound and stored with the previous frame timestamp and next frame timestamp corresponding to the candidate pairing relationship. A layer-by-layer iterative minimum potential energy search is performed on the bidirectional spatiotemporal cost graph to complete node pairing. The layer-by-layer iterative minimum potential energy search includes selecting the node pair with the smallest total potential energy value as the pairing relationship according to the round, deleting the paired nodes and updating the total potential energy value of the remaining nodes until there are no more paired nodes. Based on the node pairing relationship, a target trajectory set is generated and the trajectory identifier and trajectory point sequence are output. Specifically, the layer-by-layer iterative minimum potential energy search is performed on the bidirectional spatiotemporal cost graph to complete node pairing. In the bidirectional spatiotemporal cost graph, two adjacent frame timestamps are used as a layer. All time-directed edges and corresponding total potential values ​​in each layer are processed sequentially from early to late. The set of unpaired nodes is initialized as the set of nodes of the previous frame and the set of nodes of the next frame in the current layer. In each round, the node pair with the smallest total potential value is selected from all time-directed edges connecting unpaired nodes in the current layer as a pairing relationship. The node pair is written into the pairing table. The previous frame node and the next frame node of the node pair are deleted from the set of unpaired nodes. At the same time, the remaining time-directed edges connected to these two nodes are deleted. The mutual exclusion potential is recalculated for the remaining time-directed edges and the total potential value is updated. The mutual exclusion potential is updated by recalculating the potential of the spatial edges connected to the nodes at both ends of the time-directed edge in the same frame timestamp and taking the maximum value. This maximum value is added to the time-weighted cost to obtain the new total potential value. Repeatedly execute the rounds of selecting the minimum total potential energy value, deleting paired nodes, and updating the total potential energy value until there are no time-directed edges connecting unpaired nodes in the current layer. After completing all layers, the continuous pairing relationships in the cross-layer pairing table are concatenated in time order to form a trajectory point sequence. A unique trajectory identifier is assigned to each concatenated sequence, and the target trajectory set is output. The target trajectory set contains the trajectory identifier and the corresponding trajectory point sequence.

[0025] In this embodiment, the set of output purification trajectories includes: For each trajectory in the target trajectory set, construct a sliding window observation sequence and input the sliding window observation sequence into the Grubbs test; Based on curvature flow anomaly cycles, curvature metric calculation and curvature flow iterative evolution are performed on the sliding window observation sequence. Adjacency relationships are constructed between adjacent observation vectors in the observation vector space, and the discrete curvature value of each observation vector is calculated. The discrete curvature values ​​are updated according to the iteration cycle, generating a curvature evolution sequence. Observation vectors whose curvature values ​​satisfy the abrupt change condition in the curvature evolution sequence are marked as curvature outlier observations, where: The curvature metric calculation and curvature flow iterative evolution are performed as follows: For each observation vector in the sliding window observation sequence, component normalization is first performed. Adjacent observation vectors are connected pairwise in chronological order to construct an adjacency relationship. The Euclidean norm of the observation difference vector is calculated as the side length for each pair of adjacent observation vectors. For the observation vector in the middle position, the discrete curvature value is calculated. The discrete curvature value is calculated as the inverse of the cosine of the angle between the difference vector between the observation vector and the previous observation vector and the difference vector between the observation vector and the next observation vector. The discrete curvature values ​​of the observation vectors at the beginning and end of the sequence are set to 0. The discrete curvature values ​​are arranged into an initial curvature sequence according to the order of the observation vectors. The curvature flow iterative evolution is performed. The number of iterations of the curvature flow iterative evolution is set to 8, and the iteration step size is set to 0.10. In each round, a one-dimensional discrete Laplace update is performed on the curvature sequence. The update method is to add the current curvature value to the iteration step size multiplied by the sum of the previous and next adjacent curvature values ​​minus 2 times the current curvature value to obtain the updated curvature sequence and generate the curvature evolution sequence. Observation vectors whose curvature values ​​in the curvature evolution sequence meet the mutation condition are marked as curvature outliers. Specifically, the curvature sequence after the last round of updates is taken as the stable curvature sequence. The mean and standard deviation of the stable curvature sequence are calculated. The standardized deviation is calculated for the stable curvature value corresponding to each observation vector. The standardized deviation is the absolute value of the stable curvature value minus the mean, divided by the standard deviation. Observation vectors with a standardized deviation greater than 3.0 are marked as curvature outliers. The frame timestamp and trajectory identifier corresponding to the observation vector are written into the curvature outlier observation list, and the curvature outlier observation set is output. Energy spectrum decomposition is performed on the sliding window observation sequence using hypergraph energy spectrum boundaries. The observation vectors in the sliding window observation sequence are used as hypergraph nodes. Hyperedges are constructed based on cross-frame continuity and intra-frame mutual exclusion relationships to form a higher-order correlated hypergraph. The Laplacian matrix of the higher-order correlated hypergraph is calculated and eigenvalue decomposition is performed to obtain energy spectrum eigenvectors. Observation vectors are then demarcated based on these eigenvectors, and observation vectors falling on the weakly connected boundary are marked as energy spectrum outliers. Specifically, the energy spectrum decomposition of the sliding window observation sequence is performed as follows: Each observation vector in the sliding window observation sequence is assigned a hypergraph node, and the corresponding frame timestamp, trajectory identifier, center point coordinates, scale, confidence score, and feature vector distance are recorded for each node. When constructing a high-order relational hypergraph, cross-frame hyperedges are generated for cross-frame continuity relationships. Cross-frame hyperedges consist of 3 nodes with the same trajectory identifier within 3 consecutive frame timestamps, and the Euclidean distance between the center points of two adjacent frames is required to be no greater than 40 pixels and the relative scale change is no greater than 0.30. Same-frame hyperedges are generated for same-frame mutual exclusion relationships. Same-frame hyperedges consist of any 2 nodes within the same frame timestamp, and the intersection-union ratio of the two target boxes is required to be greater than 0.30. Hyperedge weights are assigned to each hyperedge. The weight of a cross-frame hyperedge is the average of the confidence scores of the 3 nodes, and the weight of a same-frame hyperedge is the intersection-union ratio of the target boxes of the two nodes. A hypergraph correlation matrix is ​​constructed based on the association between nodes and hyperedges. A Laplacian matrix is ​​generated from the hypergraph correlation matrix, hyperedge weights, and node degree matrix. The node degree is the sum of the weights of the hyperedges associated with the node. Eigenvalue decomposition is performed on the Laplacian matrix to obtain spectral eigenvectors arranged in ascending order of eigenvalues. The spectral eigenvector corresponding to the second smallest eigenvalue is taken as the boundary vector. The boundary vector is divided into two sides according to the sign of the corresponding components of each node. Nodes with components less than 0 are divided into one side, and nodes with components greater than or equal to 0 are divided into the other side. The sum of the hyperedge weights inside both sides is calculated. The side with the smaller sum of the internal hyperedge weights is defined as the weakly connected boundary side. The observation vectors falling into the weakly connected boundary side are marked as spectral outliers and the spectral outlier observation set is output. Curvature outlier observations and energy spectrum outlier observations are merged into an outlier observation set. Based on this outlier observation set, the target trajectory set is updated by outlier removal or replacement, resulting in a cleaned trajectory set. Specifically, the outlier removal or replacement update of the target trajectory set based on the outlier observation set is performed as follows: Curvature outlier observations and energy spectrum outlier observations are deduplicated and merged according to frame timestamps and trajectory identifiers to obtain an outlier observation set. For each outlier observation record, a corresponding trajectory identifier, frame timestamp, and observation vector are assigned. The target trajectory set is traversed one by one, and trajectory points with the same trajectory identifier and frame timestamp as those in the outlier observation set are identified as outlier trajectory points. When the proportion of outlier trajectory points of the same trajectory in the sliding window to the total number of trajectory points in the window is greater than 0.30, outlier removal and update are performed on the trajectory. Outlier removal and update involves deleting outlier trajectory points from the trajectory point sequence and recording the deletion position as a missing frame. When the proportion of outlier trajectory points is no greater than 0.30, outlier replacement update is performed on the trajectory. Outlier replacement update involves finding the adjacent non-outlier trajectory points before and after each outlier trajectory point, and using linear interpolation to generate the center point coordinates and scale of the replacement trajectory point. The linear interpolation uses the frame timestamp as the independent variable and is calculated according to the time interval ratio. The confidence of the replacement trajectory point is set to the minimum of the confidence of the adjacent non-outlier trajectory points before and after it. The trajectory length, average speed and direction angle of the trajectory that has been removed or replaced are recalculated and the trajectory state parameters are updated. The updated trajectory set is output as the cleaned trajectory set.

[0026] In this embodiment, generating the intrusion event set includes: The trajectory point sequence of each purification trajectory in the purification trajectory set is processed to be time-continuous according to the frame timestamp order. Based on the continuous time interpolation result, a continuous time position prediction curve is generated. Specifically, the generation of the continuous time position prediction curve is as follows: For each purification trajectory in the purification trajectory set, the trajectory point sequence is arranged in ascending order of frame timestamp. The x-coordinate of the center point, y-coordinate of the center point, width of the target box, and height of the target box are extracted for each trajectory point. Four sets of time series data are formed with the frame timestamp as the independent variable. Cubic spline functions are established for the x-coordinate and y-coordinate of the center point respectively. A cubic polynomial is constructed between two adjacent frame timestamps, such that the cubic polynomial takes the coordinate values ​​of the sampling points at both ends of the interval. At the same time, the function value is continuous, the first derivative is continuous, and the second derivative is continuous at the connection point of adjacent intervals. Natural boundary conditions are used at the beginning and end of the sequence to make the second derivative 0. The coefficients of the cubic polynomial in all intervals are solved and the spline function is obtained. Linear interpolation functions are established for the target box width and height respectively. A linear function is constructed between two adjacent frame timestamps, such that the linear function takes the width or height value of the sampling point at each end of the interval. The slope and intercept of the linear function are calculated from the two end values ​​and the time interval to obtain the piecewise linear interpolation function. The continuous time sampling step is set to 0.05 seconds. Sampling time is generated according to the sampling step in each adjacent frame timestamp interval. The x-coordinate and y-coordinate of the prediction center point are calculated by the spline function. The prediction width and prediction height are calculated by the linear interpolation function to obtain the prediction position and scale sequence arranged in time order. The sequence is then connected to form a continuous time position prediction curve. A risk gradient sequence is generated by calculating the point-by-point distance of the continuous time location prediction curve. The predicted location point is obtained at the sampling time of the continuous time location prediction curve. The shortest Euclidean distance from the predicted location point to the boundary of each area in the set of electronic fence areas is calculated and the minimum value is taken as the fence distance. At the same time, the shortest Euclidean distance from the predicted location point to each sensitive point in the set of sensitive points is calculated and the minimum value is taken as the point distance. The risk gradient values ​​at the corresponding sampling time are generated based on the fence distance and the point distance and are arranged into a risk gradient sequence in chronological order. The risk gradient sequence is integrated to generate a risk integral sequence. The risk gradient sequence is then numerically integrated over time intervals to obtain a cumulative risk integral value. The intrusion state and risk level are determined based on the change in the cumulative risk integral value. Finally, time intervals where the cumulative risk integral value of the same trajectory continuously satisfies the intrusion state condition are merged to generate an intrusion event set, where: The risk gradient sequence is integrated to generate a risk integral sequence. The risk gradient sequence is then numerically integrated according to the time interval of the sampling time to obtain the cumulative risk integral value. Specifically, the risk gradient sequence corresponding to each purification trajectory is sorted in ascending order of sampling time. The initial cumulative risk integral value is set to 0. The time interval between two adjacent sampling times is calculated. The risk gradient sequence is cumulatively integrated using trapezoidal numerical integration. For each pair of adjacent sampling points, the average of the previous and next risk gradient values ​​is taken and multiplied by the corresponding time interval to obtain the risk increment of the time interval. The risk increment is accumulated to the cumulative risk integral value of the previous sampling time to obtain the cumulative risk integral value of the current sampling time. The cumulative risk integral values ​​of each sampling time are arranged in chronological order to form a risk integral sequence. The intrusion status and risk level are determined based on the change process of the cumulative risk score. The time intervals in which the cumulative risk score corresponding to the same trajectory continuously meets the intrusion status conditions are merged to generate an intrusion event set, specifically: The risk score sequence is used to determine the status. The intrusion status and risk level are determined by the cumulative risk score value and the time change rate. The time change rate is the difference between the cumulative risk score values ​​at two adjacent sampling times divided by the time interval. The intrusion status is set to three levels: warning, alert, and prohibited. When the cumulative risk score value is greater than or equal to 80 and the time change rate is greater than or equal to 10, it is determined to be a warning status. When the cumulative risk score value is greater than or equal to 200 and the time change rate is greater than or equal to 20, it is determined to be an alert status. When the cumulative risk score value is greater than or equal to 400 and the time change rate is greater than or equal to 30, it is determined to be a prohibited status. If the conditions are not met, it is determined to be a non-intrusion status. Risk levels correspond to intrusion states: early warning state corresponds to low risk level, alert state corresponds to medium risk level, and prohibited state corresponds to high risk level. The intrusion state sequence is scanned sequentially for the same trajectory. The time period when the intrusion state is continuously in early warning, alert, or prohibited state is taken as candidate intrusion interval. Candidate intrusion intervals with an interval of less than or equal to 0.5 seconds between two adjacent candidate intrusion intervals are merged. An intrusion event is generated for each merged candidate intrusion interval. The occurrence time of the intrusion event is taken as the start sampling time of the interval, the duration is taken as the end sampling time of the interval minus the start sampling time, the area level is taken as the area level corresponding to the highest intrusion state in the interval, and the risk level is taken as the highest risk level in the interval. An intrusion event set is generated, and a corresponding trajectory identifier is written for each intrusion event.

[0027] In this embodiment, the step of generating an evidence package from the intrusion event set and outputting alarm information includes: For each intrusion event in the intrusion event set, an evidence package is generated. The evidence package includes a set of keyframes corresponding to the time of the intrusion event, the target bounding box coordinates and category identifiers of each keyframe in the keyframe set, the trajectory point sequence of the cleanup trajectory set corresponding to the intrusion event, the electronic fence spatial relationship record corresponding to the intrusion event, and outlier observation records. The outlier observation records include the frame timestamps corresponding to the outlier observations, the outlier observation type identifiers, and the observation vectors of the outlier observations. Generate alarm information, which includes the event identifier, area level, risk level, occurrence time, duration, corresponding trajectory identifier, and evidence package index of the intrusion event. The evidence package index is the unique location identifier of the evidence package in the end-side storage. The alarm information and evidence package are encapsulated and sent to the upper platform. The alarm information and evidence package index are written into the alarm message body and the evidence package is written into the local storage according to the file structure. The alarm message body is sent to the upper platform through the network link. The evidence package or evidence package index is synchronized to the upper platform. At the same time, the sending status and retransmission records are recorded on the terminal side.

[0028] Example 1: To verify the feasibility of this invention in practice, it was applied to a continuous low-altitude security monitoring cycle. The system received video frame sequence data from monitoring cameras, totaling 2160 frames, with a resolution of 1920×1080 and a frame rate of 25 frames per second. The data covered environments with strong light, low illumination, and supplemental lighting. The background included swaying tree canopies, metallic reflections, light sweeping, and compression artifacts. Manual annotation revealed 16 real drone intrusion events and 12 instances of distant small targets, with target bounding boxes all less than 20 pixels in width and height, the smallest being 12×9 pixels. Non-drone moving objects included birds, floating objects, and reflective points, accounting for 31.5% of the interference segments. During operation at the endpoint, network bandwidth intermittently decreased, with a minimum uplink bandwidth of 0.9 Mbps. The system primarily used local closed-loop transmission and retransmitted evidence packets after bandwidth recovery.

[0029] After the data enters the process, video frames are normalized to 0 to 1 and denoised using a 5×5 median filter. Invalid frames with a mean brightness less than 0.03 or greater than 0.97 are removed, at a rate of 0.6%. After preprocessing, the average contrast ratio increases from 0.18 to 0.23, and the number of edge noise points decreases from 412 per frame to 265. Sparse optical flow is calculated by constructing a 4-layer pyramid. Each layer is first smoothed with a standard deviation of 1.0 and a 5×5 Gaussian smoothing scheme, then downsampled by half its width and height. The displacement calculation uses a combination of constant brightness and spatial smoothing at a smoothing coefficient of 0.05, iterating 10 times per layer, backtracking from layer 3 to layer 0, and retaining pixels with a brightness residual greater than 15 as sparse optical flow points. Statistics show that approximately 820 to 1450 optical flow points are retained per frame for strongly interfering segments, and approximately 210 to 460 for segments with stable backgrounds.

[0030] Motion energy maps are calculated based on sparse optical flow, where energy is the sum of squares of horizontal and vertical displacement components. Locations not retained are set to 0, and a 3×3 median filter is applied. Pixels with motion energy values ​​greater than 9 are marked as foreground. Connected regions are generated based on 8-neighborhoods, and those with fewer than 30 pixels are removed, resulting in a set of motion connected regions. On average, there are 1.9 connected regions per frame across the entire dataset, with a maximum of 7. Adaptive quadtree splitting and merging is performed on the motion connected regions: regions with a motion energy standard deviation greater than 6 are split, with a minimum sub-region of 16×16; adjacent sub-regions with a mean difference of no more than 3 are merged; leaf regions with at least 50 foreground pixels are retained as candidate motion regions. After this step, there are an average of 2.6 candidate regions per frame, with a maximum of 6, achieving a target coverage of 0.97 for real UAVs. The bounding rectangle of each candidate region is expanded by 1.25 and cropped. Regions with fewer than 400 pixels are not generated as image patches; the remaining image patches are filled with squares and scaled to 224×224 to form a candidate image patch sequence.

[0031] In the detection and recognition stage, candidate image block sequences are input into the improved SqueezeNet network according to frame timestamps. The network first outputs frame-level intermediate feature sequences, and then synthesizes multiple feature blocks from the same frame by concatenating them by channel and taking the maximum value of each channel dimension. The spectral-spatial double-overlay reconstruction network performs orthogonal basis transformations in the temporal and spatial dimensions on the frame-level intermediate features, iterating for 5 rounds. In each round, spectral domain coefficients with a cumulative energy percentage within 0.95 are retained and inversely reconstructed, updated with a step size of 0.1, outputting background suppression features. The mean channel variance in strongly interfering segments decreases from 0.42 to 0.28, while the mean channel response in the UAV-labeled region increases from 0.31 to 0.47. Accompanied by ladder network convergence, the network evolves forward 6 times with an evolution step size of 0.05, and then backwards using supervision errors, updating the target focusing features 4 times with a step size of 0.02. After focusing, the peak response of the UAV region is improved by approximately 29%, and the background response by approximately 6%. The detection and recognition output threshold is 0.40, and the overlap suppression cross-union ratio threshold is 0.50. The reverse trajectory repair generates trajectory segments with a center point distance of up to 40 pixels, and missing frames 1 to 5 are identified as broken intervals. Consistency checks are performed based on a center point distance of no more than 30 pixels, a scale change rate of no more than 0.30, and a cosine distance between feature vectors of no more than 0.20. This repair was triggered 9 out of 16 events, completing an average of 2.4 frames, increasing the proportion of continuous trajectory frames from 0.78% to 0.92%.

[0032] In the trajectory generation and purification stage, a bidirectional spatiotemporal cost graph is established based on UAV candidate targets. The temporal edge cost and spatial edge potential energy are calculated, and a layer-by-layer iterative minimum potential energy search is performed to generate a set of target trajectories. Subsequently, a sliding window observation sequence is constructed for each trajectory, and an improved Grubbs test is performed: the curvature flow anomaly circle normalizes the mean and standard deviation of the observation vector within the window, calculates the discrete curvature, and generates a curvature evolution sequence through 8 iterations with an iteration step size of 0.10. Curvature outliers with a normalized deviation greater than 3.0 are marked as curvature outliers. The hypergraph energy spectrum boundary is defined by three consecutive frames forming a cross-frame hyperedge with a center point distance of no more than 40 pixels and a scale change of no more than 0.30. The same-frame hyperedge is defined by an intersection-union ratio greater than 0.30. After eigenvalue decomposition of the Laplacian matrix, energy spectrum outliers are marked by weakly connected sides. After merging the two types of outliers, trajectories with an outlier ratio greater than 0.30 are removed; otherwise, linear interpolation is performed for replacement. Statistics show that there were 1,840 curvature outliers and 1,265 energy spectrum outliers. After merging and deduplication, there were 2,410 outliers. After purification, the number of short pseudo-trajectory triggers caused by birds and reflections decreased from 4.1 times per hour to 0.7 times per hour.

[0033] During the intrusion event generation phase, the purification trajectory is processed to be time-continuous. A cubic spline function is established for the x and y coordinates of the center point, and a piecewise linear interpolation function is established for the width and height. A continuous time-location prediction curve is generated with a continuous sampling step of 0.05 seconds. The shortest Euclidean distance to the electronic fence boundary and the sensitive point is calculated point by point to obtain the risk gradient sequence. A risk integral sequence is generated by accumulating the risk gradient using a trapezoidal integral, with the integral value limited to 0 to 1000. The intrusion status is determined jointly based on the cumulative risk integral value and the rate of change: a value greater than or equal to 80 and a rate of change greater than or equal to 10 indicates a warning; a value greater than or equal to 200 and a rate of change greater than or equal to 20 indicates a alert; and a value greater than or equal to 400 and a rate of change greater than or equal to 30 indicates a prohibited entry. Events with an interval of no more than 0.5 seconds between adjacent candidate intervals are merged. Event records were generated for all 16 real intrusion events, including 10 prohibited entries, 4 alert entries, and 2 warning entries; the average alarm delay was 0.46 seconds, with a maximum of 0.71 seconds. On average, the evidence package contains 12 keyframes, 112 trajectory sampling points, and 8 outlier records. When bandwidth decreases, an alarm message body and index are reported first, and the evidence package is retransmitted after the link is restored, with a retransmission success rate of 1.00%.

[0034] To enhance the persuasiveness of the comparison, this embodiment provides training samples and comparative experiments. Training samples are drawn from candidate image patches; samples with an intersection-over-union (IoU) ratio greater than 0.50 are considered positive, those with an IoU less than 0.20 are considered negative, and those between 0.20 and 0.50 are not included in the training. The positive-to-negative sampling ratio is 1:3. There are 4800 training samples, including 1200 positive samples and 3600 negative samples, and 1200 validation samples. There are 1500 test samples, with small target positive samples accounting for 41.8%. Traditional methods use the unmodified SqueezeNet detection and recognition process combined with conventional trajectory association and simple anomaly removal, with consistent training samples and classification. The results for the same batch of test samples and the same full video data are as follows: small target detection rate: traditional method 0.74, this invention 0.91; overall false negative rate: traditional method 0.18, this invention 0.06; false alarms per hour: traditional method 4.2, this invention 0.7.

[0035] As can be seen from the embodiments, the present invention has quantifiable data changes in trajectory generation, outlier removal, fence event generation and evidence collection closed loop, and has advantages over traditional methods in terms of long-distance small target detection, false alarm suppression, trajectory continuity, alarm delay and evidence collection integrity, verifying the engineering feasibility and effectiveness under complex background and edge-side constraints.

[0036] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A drone intrusion detection and identification method based on AI vision, characterized in that, include: The video frame sequence output by the monitoring camera is acquired, and the video frame sequence is preprocessed to obtain the frame sequence to be processed; Pixel displacement estimation is performed on the frame sequence to be processed to generate a sparse optical flow vector field. Based on the optical flow vector field, the motion energy map is calculated, and an adaptive quadtree split-merge strategy is used to obtain a candidate image block sequence. An improved SqueezeNet network is constructed to extract features and detect and identify candidate image block sequences. Spectral domain transformation and alternating iterative reconstruction are performed based on the spectral-space double-overlap reconstruction network. Quasi-temporal evolution and adjoint gradient solution are performed through the adjoint ladder network convergence. Inverse trajectory repair is introduced to perform fracture completion, and a set of UAV candidate targets is obtained. A bidirectional spatiotemporal cost graph is established for the set of candidate UAV targets. Node pairing is completed on the bidirectional spatiotemporal cost graph through a layer-by-layer iterative minimum potential energy search to generate a set of target trajectories. The Grubbs test is performed on the target trajectory set, curvature measurement is calculated and curvature flow iterative evolution is performed based on the curvature flow anomaly circle, energy spectrum decomposition is performed through the hypergraph energy spectrum boundary, and the purified trajectory set is output. A continuous-time location prediction curve is constructed based on the set of purification trajectories. A risk gradient sequence is generated by calculating the distance point by point. The risk gradient sequence is then integrated to generate a set of intrusion events. Generate an evidence package from the set of intrusion events and output alarm information, then send the alarm information and evidence package to the upper platform.

2. The method for drone intrusion detection and identification based on AI vision according to claim 1, characterized in that, The video frame sequence includes a frame image matrix, frame timestamps, camera intrinsic parameters, camera extrinsic parameters, gimbal attitude parameters, and geolocation information.

3. The method for drone intrusion detection and identification based on AI vision according to claim 1, characterized in that, The process of obtaining the frame sequence to be processed includes: Acquire the video frame sequence output by the monitoring camera at continuous sampling time, write a frame timestamp for each frame, and read the camera intrinsic parameters, camera extrinsic parameters, PTZ attitude parameters and geolocation information corresponding to the current frame to establish a multi-source data alignment table indexed by the frame timestamp. Image preprocessing is performed on each frame in the video frame sequence. Image preprocessing includes normalizing the size of the frame image matrix, normalizing the pixel intensity, and denoising. Invalid frames are removed to generate a preprocessed frame sequence. The preprocessed frame sequence is written into the frame buffer in the order of frame timestamps and the frame sequence to be processed is output. At the same time, the binding relationship between the camera intrinsic parameters, camera extrinsic parameters, gimbal attitude parameters and geolocation information corresponding to each frame in the frame sequence to be processed is recorded in the frame buffer.

4. The method for drone intrusion detection and identification based on AI vision according to claim 1, characterized in that, The process of obtaining the candidate image patch sequence includes: The frame image matrices of two adjacent frames in the frame sequence to be processed are grayscaled, a multi-resolution image pyramid is constructed, and the pixel displacement is solved at each resolution level under the conditions of constant brightness constraint and spatial smoothness constraint to obtain the sparse optical flow vector field. The sparse optical flow vector field is then back-projected to the original resolution coordinate system according to the pyramid level. The motion energy map is calculated based on the sparse optical flow vector field, and connected component generation is performed on the motion energy map to obtain a set of motion connected components. Adaptive quadtree splitting and merging are performed on the set of motion connected regions. Quadtree splitting is triggered when the statistical dispersion of motion energy in the current region exceeds a preset dispersion condition, and quadtree merging is triggered when the statistical difference of motion energy between adjacent sub-regions does not exceed a preset difference condition. The set of candidate motion regions is then output. For each candidate motion region in the candidate motion region set, calculate the bounding rectangle. After expanding the bounding rectangle in the width and height directions according to the expansion ratio, crop the corresponding image region to obtain the candidate image block sequence.

5. The method for drone intrusion detection and identification based on AI vision according to claim 1, characterized in that, The obtained set of candidate targets for unmanned aerial vehicles includes: An improved SqueezeNet network is constructed by connecting a spectral-space double-layer reconstruction network between the input of the original SqueezeNet network and the first feature extraction layer, connecting an accompanying ladder network convergence network between the original intermediate feature output layer and the detection and recognition output layer, and setting a reverse trajectory repair branch at the original detection and recognition output. The candidate image block sequence is input into the input end of the original SqueezeNet network in the order of frame timestamps. The output is a frame-level intermediate feature sequence. The frame-level intermediate feature sequence is input into the spectral-space double-overlay reconstruction network. The spectral-space double-overlay reconstruction network performs orthogonal basis transformation in the time dimension and spatial dimension respectively to obtain the spectral domain coefficient sequence. The spectral domain coefficient sequence is subjected to alternating iterative processing of energy clipping, inverse transformation reconstruction and minimum difference update in each round to output the background suppression feature sequence. The background suppression feature sequence is input into the adjoint ladder network convergence flow. The adjoint ladder network convergence flow represents each frame feature in the background suppression feature sequence as a two-dimensional discrete field. The two-dimensional discrete field is subjected to pseudo-time forward evolution according to the evolution step size to obtain the evolution feature field. An adjoint gradient field is established for the evolution feature field and an adjoint inverse solution is performed. The evolution feature field is iteratively updated according to the adjoint gradient field and the target focusing feature sequence is output. The target-focused feature sequence is input into the detection and recognition output layer of the original SqueezeNet network, which outputs a set of candidate detection results corresponding to the timestamps of each frame. The set of candidate detection results is input into the reverse trajectory repair branch. The reverse trajectory repair branch generates a set of trajectory segments based on the set of candidate detection results of adjacent frames, identifies the broken intervals in the set of trajectory segments, performs time reverse consistency verification on the trajectory segments on both sides of the broken interval to generate reverse completion results, writes the reverse completion results into the set of candidate detection results of the frame timestamps corresponding to the broken interval, and outputs a set of candidate targets for the UAV. The improved SqueezeNet network is trained by constructing a training sample set and inputting the candidate image patch sequence of the training sample set into the improved SqueezeNet network. The joint loss function is constructed based on the bounding box regression loss and the category classification loss. Backpropagation is used to calculate the gradient of the joint loss function and update the trainable parameters of the original SqueezeNet network part, the spectral-space double-overlap reconstruction network part and the accompanying ladder network convergence part in the improved SqueezeNet network until the training termination condition is met and the trained improved SqueezeNet network is obtained.

6. The method for drone intrusion detection and identification based on AI vision according to claim 1, characterized in that, The generated target trajectory set includes: A bidirectional spatiotemporal cost graph is constructed based on the set of UAV candidate targets. The nodes of the bidirectional spatiotemporal cost graph are composed of UAV candidate targets corresponding to the timestamps of each frame. The edges of the bidirectional spatiotemporal cost graph include temporal edges and spatial edges. The edge weights of the bidirectional spatiotemporal cost graph are calculated and the total potential energy value is generated. The temporal edge cost is generated by the center point displacement, scale change and feature vector distance according to the combination rules. The spatial edge potential energy is generated by the overlap of the target boxes in the same frame according to the transformation rules. The total potential energy value is generated by the temporal edge cost and the spatial edge potential energy and corresponds to the inter-frame time difference. A layer-by-layer iterative minimum potential energy search is performed on the bidirectional spatiotemporal cost graph to complete node pairing. The layer-by-layer iterative minimum potential energy search includes selecting the node pair with the smallest total potential energy value as the pairing relationship according to the round, deleting the paired nodes and updating the total potential energy value of the remaining nodes until there are no more paired nodes. Based on the node pairing relationship, a target trajectory set is generated and the trajectory identifier and trajectory point sequence are output.

7. The method for drone intrusion detection and identification based on AI vision according to claim 1, characterized in that, The set of output purification trajectories includes: For each trajectory in the target trajectory set, construct a sliding window observation sequence and input the sliding window observation sequence into the Grubbs test; Based on the curvature flow anomaly loop, the curvature metric calculation and curvature flow iterative evolution are performed on the sliding window observation sequence. Adjacency relationships are constructed between adjacent observation vectors in the observation vector space and the discrete curvature value of each observation vector is calculated. The discrete curvature value is updated according to the iteration round and a curvature evolution sequence is generated. The observation vectors whose curvature values ​​in the curvature evolution sequence meet the mutation condition are marked as curvature outlier observations. The energy spectrum decomposition of the sliding window observation sequence is performed by the energy spectrum boundary of the hypergraph. The observation vectors in the sliding window observation sequence are used as hypergraph nodes. Hyperedges are constructed to form a higher-order correlated hypergraph based on cross-frame continuity and intra-frame mutual exclusion. The Laplacian matrix of the higher-order correlated hypergraph is calculated and eigenvalue decomposition is performed to obtain the energy spectrum eigenvector. The observation vectors are delineated according to the energy spectrum eigenvector. The observation vectors falling on the weakly connected boundary side are marked as energy spectrum outliers. Curvature outlier observations and energy spectrum outlier observations are merged into an outlier observation set. Based on the outlier observation set, the target trajectory set is updated by outlier removal or replacement, and a clean trajectory set is output.

8. The method for drone intrusion detection and identification based on AI vision according to claim 1, characterized in that, The generated intrusion event set includes: The trajectory point sequence of each purification trajectory in the purification trajectory set is processed to be time continuous according to the frame timestamp order, and a continuous time position prediction curve is generated based on the continuous time interpolation result. A risk gradient sequence is generated by calculating the point-by-point distance of the continuous time location prediction curve. The predicted location point is obtained at the sampling time of the continuous time location prediction curve. The shortest Euclidean distance from the predicted location point to the boundary of each area in the set of electronic fence areas is calculated and the minimum value is taken as the fence distance. At the same time, the shortest Euclidean distance from the predicted location point to each sensitive point in the set of sensitive points is calculated and the minimum value is taken as the point distance. The risk gradient values ​​at the corresponding sampling time are generated based on the fence distance and the point distance and are arranged into a risk gradient sequence in chronological order. The risk gradient sequence is integrated to generate a risk integral sequence. The risk gradient sequence is numerically integrated over the time interval of the sampling time to obtain the cumulative risk integral value. The intrusion status and risk level are determined based on the change process of the cumulative risk integral value. The time intervals in which the cumulative risk integral value of the same trajectory continuously meets the intrusion status condition are merged to generate an intrusion event set.

9. The method for drone intrusion detection and identification based on AI vision according to claim 1, characterized in that, The process of generating an evidence package from the set of intrusion events and outputting alarm information includes: For each intrusion event in the intrusion event set, an evidence package is generated. The evidence package includes a set of keyframes corresponding to the time of the intrusion event, the target bounding box coordinates and category identifiers of each keyframe in the keyframe set, the trajectory point sequence of the cleanup trajectory set corresponding to the intrusion event, the electronic fence spatial relationship record corresponding to the intrusion event, and outlier observation records. The outlier observation records include the frame timestamps corresponding to the outlier observations, the outlier observation type identifiers, and the observation vectors of the outlier observations. Generate alarm information, which includes the event identifier, area level, risk level, occurrence time, duration, corresponding trajectory identifier, and evidence package index of the intrusion event. The evidence package index is the unique location identifier of the evidence package in the end-side storage. The alarm information and evidence package are encapsulated and sent to the upper platform. The alarm information and evidence package index are written into the alarm message body and the evidence package is written into the local storage according to the file structure. The alarm message body is sent to the upper platform through the network link. The evidence package or evidence package index is synchronized to the upper platform. At the same time, the sending status and retransmission records are recorded on the terminal side.

Images