A UAV early warning system based on SCAD-SSD and its hardware acceleration method

Abstract

This invention belongs to the field of artificial intelligence and FPGA hardware accelerator design. It discloses a UAV early warning system based on SCAD-SSD and its hardware acceleration method. At the early warning system level, an infrared UAV dataset is constructed, an SCAD-SSD model is built, and an optimal detection model with infrared UAV target recognition capability is obtained through iterative training. At the hardware acceleration method level, a hardware accelerator consisting of multi-dimensional parallel memory access units, computing clusters, and a global controller is deployed on the programmable logic (PL) end of the FPGA. During forward inference, task scheduling is implemented by relying on the global controller, controlling the multi-dimensional parallel memory access units to perform data transfer, driving the general parameterized reconfigurable engine to realize the logical mapping of the optimal detection model through dynamic reconfiguration of hardware resources, and coordinating the computing cluster to switch between parallel working mode and collaborative mode, ultimately realizing on-chip pipelined inference for the optimal detection model.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and FPGA hardware accelerator design, and relates to a drone early warning system based on SCAD-SSD and its hardware acceleration method. Background Technology

[0002] In recent years, the widespread application of low-altitude drones has brought severe challenges to low-altitude airspace security. Building an efficient drone early warning system has become a key requirement in security monitoring and short-range air defense. Infrared target detection technology, with its advantages of all-weather operation and strong anti-interference capabilities, plays a crucial role in drone early warning, security monitoring, and short-range air defense. However, in long-range detection scenarios, due to the influence of imaging distance and the weak thermal radiation of the drone itself, infrared drone targets suffer severe energy attenuation and low signal-to-noise ratio, often appearing as "weak point sources" lacking texture and geometric features in images, making it difficult for the system to capture effective information. At the same time, clutter interference from complex backgrounds such as cloud edges and strong building reflections easily triggers high-frequency false alarms. This dual contradiction of weak signal and background interference necessitates that drone detection algorithms must balance detection performance and lightweight architecture while ensuring sensing capabilities.

[0003] At the algorithmic level, existing deep learning solutions still have limitations when dealing with infrared drone targets. On the one hand, while the stepwise downsampling architecture used by mainstream detectors can expand the receptive field, it easily leads to "feature annihilation" of weak drone signals at lower levels during deep transmission, resulting in the loss of information about small targets. On the other hand, while introducing global correlation or attention mechanisms can strengthen features, it inevitably generates huge computational loads and memory access overheads, making it difficult to meet the stringent real-time requirements of drone early warning systems. Although using lightweight backbone networks can reduce complexity, excessively high compression ratios often weaken the model's ability to capture weak signals. Therefore, how to prevent drone feature loss and suppress false alarms while maintaining a lightweight architecture is the core challenge in algorithm design.

[0004] At the hardware implementation level, existing hardware acceleration solutions mostly perform static parallel optimizations for regular matrix calculations, making it difficult to adapt to the requirements of irregular enhancement operators or non-continuous memory access logic in infrared detection algorithms. When handling heterogeneous computing loads, existing accelerators lack flexible task scheduling and resource reconfiguration capabilities, which can easily lead to idle computing resources, data diversion pauses, or inference pipeline cavitation, severely limiting the real-time processing performance of hardware accelerators in resource-constrained environments.

[0005] Therefore, exploring a lightweight algorithm that can effectively suppress the annihilation of weak target features and reduce background false alarms, and proposing an efficient hardware acceleration method to match it, and building a UAV early warning system that balances high detection rate and low false alarm rate, has become a key requirement for realizing real-time infrared UAV target detection. Summary of the Invention

[0006] The purpose of this invention is to propose an infrared target detection method and its hardware acceleration method based on SCAD-SSD for edge deployment, which aims to improve the detection accuracy of small infrared targets in complex backgrounds, suppress the annihilation of weak target features, reduce false alarms in the background, and solve the technical problems of low operator mapping efficiency and obvious memory access bottleneck in the pipelined inference process of hardware accelerators.

[0007] The technical solution of the present invention:

[0008] A drone early warning system based on SCAD-SSD includes the following steps:

[0009] Step 1: Construct a complete infrared UAV dataset, perform structured annotation on the infrared UAV dataset, and then divide it into training, validation, and test sets;

[0010] Step 1.1: Use an infrared camera to continuously track drone targets, collect infrared video sequences containing drone targets in different scenarios, extract the original infrared images from the infrared video sequences as positive samples, collect pure background images without drone targets as negative samples, and supplement isolated images containing drone targets in scenarios that are prone to false alarms as expanded positive samples to construct a complete infrared drone dataset.

[0011] Step 1.2: Use annotation tools to locate the bounding boxes of drone targets in each image of the infrared drone dataset, obtain the center coordinates and width and height information of the drone targets in the image, and normalize them; record the category label of the drone targets and the real target box position information, the real target box position information including the normalized center coordinates and width and height information, and generate the corresponding structured annotation file; divide the infrared drone dataset constructed in Step 1.1 into training set, validation set and test set; first, the original infrared images belonging to the same infrared video sequence are treated as a whole and assigned to a single set, and then the negative samples and expanded positive samples are assigned to different single sets, so that the divided training set, validation set and test set meet the preset ratio.

[0012] Step 2: Construct the SCAD-SSD model consisting of a feature extraction network and prediction branches; acquire images from the complete infrared UAV dataset constructed in Step 1 and perform spatial scale reconstruction, uniformly adjusting them to the network's preset size to obtain scale-reconstructed images; after mean and standard deviation correction, the scale-reconstructed images are sequentially passed through a feature extraction network with the SCB module as its core, producing 7 prediction layer feature maps, which are then branched into 7 corresponding prediction branches; the prediction branches perform feature correction for spatial saliency and local texture refinement through the ADR module, and the detection head produces position regression parameters and probability confidence scores; finally, combined with the absolute spatial reference benchmark, the infrared UAV target detection results are output through coordinate decoding and redundancy removal;

[0013] Step 2.1: Build the SCAD-SSD model, which consists of a feature extraction network and 7 prediction branches; the feature extraction network is composed of a feature extraction backbone and an extension layer cascaded sequentially; the prediction branches are composed of an ADR module and a detection head cascaded sequentially; the 7 prediction branches are coupled to the 7 outputs of the feature extraction network respectively.

[0014] Step 2.1.1: The feature extraction backbone consists of an initial sampling layer, 17 consecutive SCB modules, and a final dimension-upgrading layer connected in sequence; the initial sampling layer contains sequentially connected... Standard convolutional layers, batch normalized layers, and ReLU6 activation layers; 17 consecutive SCB modules, SCB 1-SCB 17; the final upscaling layer contains sequentially connected... The system consists of pointwise convolutional layers, batch normalization layers, and ReLU6 activation layers. The extended layer is composed of four consecutive SCB modules, defined as E-SCB 1, E-SCB 2, E-SCB 3, and E-SCB 4 in hierarchical order.

[0015] Each SCB module in the feature extraction backbone and extended layers is statically configured according to its layer level as one of the following: initial feature enhancement topology, downsampling topology, or standard feature enhancement topology. Specifically, SCB 1 in the feature extraction backbone is configured as the initial feature enhancement topology, with its input being an Identity unit performing identity mapping, its intermediate layers containing parallel main and side units, and its output being an element-wise multiplication unit and a projective convolutional layer. SCB 2, SCB 4, SCB 7, and SCB 14 in the feature extraction backbone and E-SCB 1 and E-SCB 2 in the extended layers are configured as a downsampling topology, consisting of sequentially cascaded channel expansion layers, main units, and projective convolutional layers. The remaining SCBs in the feature extraction backbone and E-SCB 3 and E-SCB 4 in the extended layers are configured as standard feature enhancement topologies, with their input being a channel expansion layer, and the structure of their intermediate and output layers being consistent with that of the initial feature enhancement topology.

[0016] The channel expansion layer contains sequentially connected... Pointwise convolutional layers, batch normalized layers, and ReLU6 activation layers; main branch units contain sequentially connected... The layers consist of depthwise convolutional layers, batch normalization layers, and ReLU6 activation layers. The deep convolutional layer has a stride of 2 in the downsampling topology and a stride of 1 in the initial feature enhancement topology and the standard feature enhancement topology; the side branch units contain sequentially connected units with a stride of 1 and an expansion rate of 2. The system consists of depthwise dilated convolutional layers, batch normalization layers, and sigmoid activation layers; the projected convolutional layers are formed by sequentially concatenating... It consists of pointwise convolutional layers and batch normalization layers;

[0017] Step 2.1.2: The first two prediction branches out of the seven prediction branches are set at the outputs of SCB 6 and SCB 13 in the feature extraction backbone, respectively, and each is connected in series with an L2Norm operator; the third prediction branch is set at the output of the final upscaling layer; the fourth to seventh prediction branches are set at the outputs of E-SCB 1, E-SCB2, E-SCB 3 and E-SCB 4 in the extended layer, respectively.

[0018] The ADR module consists of a spatial saliency branch, a local texture refinement branch, a residual branch, and an end-fusion component. The spatial saliency branch is composed of channel pooling splicing units, standard convolutional layers, batch normalization layers, ReLU activation layers, standard convolutional layers, and sigmoid activation layers cascaded sequentially. The local texture refinement branch is composed of depth convolutional layers and batch normalization layers cascaded sequentially. The end-fusion component integrates a weight gain unit, an element-wise multiplication unit, a weight compensation unit, an element-wise addition unit, and a ReLU6 activation layer.

[0019] The detection head consists of a set of position regression convolutional layers and a set of confidence prediction convolutional layers. The position regression convolutional layer set contains 7 position regression convolutional layers, and the confidence prediction convolutional layer set contains 7 confidence prediction convolutional layers. The output of the ADR module in each of the 7 prediction branches is coupled to the position regression convolutional layer and the confidence prediction convolutional layer in the corresponding detection head.

[0020] Step 2.2: Input the images from the infrared UAV dataset obtained in Step 1 into the SCAD-SSD model constructed in Step 2.1. Obtain the corrected image by preprocessing the image through spatial scale reconstruction and mean standard deviation correction. Seven prediction layer feature maps are generated in the feature extraction network. The ADR module in the prediction branch performs feature fusion and correction. Then, the position regression and confidence prediction are performed by the detection head. Finally, the candidate box coordinate decoding and redundancy removal are completed by combining the absolute spatial reference benchmark to realize the inference prediction of infrared UAV targets.

[0021] Step 2.2.1: Acquire images from the infrared UAV dataset constructed in Step 1, construct piecewise cubic polynomial basis functions using the cubic spline interpolation algorithm, perform spatial scale reconstruction on the images, and uniformly adjust them to [the required scale]. The pixels are used to obtain a scale-reconstructed image; then, mean and standard deviation correction is performed on the scale-reconstructed image to obtain a corrected image, calculated using the following formula:

[0022]

[0023] in, These are the pixel coordinates of the image. To reconstruct the pixel values ​​of the corresponding coordinate points in the image at different scales, To correct the pixel values ​​at corresponding coordinate points in the image; and These are the pixel mean and pixel standard deviation of all pixels in the scaled reconstructed image, respectively.

[0024] Step 2.2.2: Input the corrected image into the initial sampling layer of the feature extraction backbone to perform preliminary spatial dimensionality reduction and output low-dimensional feature data; the low-dimensional feature data sequentially passes through all SCB modules in the feature extraction backbone to perform feature enhancement or spatial downsampling processing. After completing SCB 6 and SCB 13 processing, two shallow prediction layer feature maps are obtained; then, the feature dimension is increased and the semantic representation is enhanced through the end dimensionality enhancement layer of the feature extraction backbone to obtain the mid-level prediction layer feature map; then, it passes through the extended layers E-SCB 1-E-SCB 4 to construct a deep semantic space and capture global context information, obtaining four deep prediction layer feature maps, for a total of seven prediction layer feature maps.

[0025] When feature data passes through an SCB module configured with an initial feature enhancement topology or a standard feature enhancement topology, intermediate feature enhancement features are first obtained through Identity units or channel expansion layers. Subsequently, the intermediate features are enhanced. Parallel inputs are fed to the main branch unit and the side branch unit; local detail features are extracted by the main branch unit through a depthwise convolution with a stride of 1. Simultaneously, the side units capture long-range environmental cues and generate spatial attention weights through depthwise dilated convolutions. The calculation formula is as follows:

[0026]

[0027] in, This is an expanded convolutional layer with an expansion rate of 2 and a stride of 1; BN is batch normalization. Use the Sigmoid activation function;

[0028] Subsequently, local detail features Spatial attention weights Element-wise multiplication is performed, followed by channel fusion and dimensionality compression via a projective convolutional layer to generate output features. The calculation formula is as follows:

[0029]

[0030] in, This refers to pointwise convolution in a projective convolutional layer. This indicates element-wise multiplication;

[0031] When feature data is configured as an SCB module with a downsampled topology, it first performs dimensionality augmentation through a channel expansion layer to obtain intermediate downsampled features. Subsequently, the main branch unit utilizes a depthwise convolutional layer with a stride of 2 to... Spatial downsampling is performed, and finally, the output features are generated through a projection convolutional layer. The calculation formula is as follows:

[0032]

[0033] After the feature data has traversed all 17 SCB modules of the feature extraction backbone, it is input into the end dimension-up layer. The linear mapping and dimension expansion of the channel dimension are performed by pointwise convolution, and the high-dimensional feature data with strong semantic expression is output as the feature map of the middle prediction layer.

[0034] Simultaneously, the high-dimensional feature data sequentially enters the extension layer consisting of E-SCB 1, E-SCB 2, E-SCB 3, and E-SCB 4, obtaining four deep prediction layer feature maps. The high-dimensional feature data then passes through E-SCB 1 and E-SCB 2, configured with a downsampling topology, undergoing continuous spatial compression and receptive field expansion. This results in a progressively decreasing output resolution and a progressively increasing semantic abstraction level, ultimately yielding deep prediction layer features. Figure 1 Features of deep prediction layers Figure 2 Subsequently, deep prediction layer features Figure 2E-SCB 3 and E-SCB 4, configured with a standard feature enhancement topology, utilize long-range environmental cues provided by side units for spatial attention weighting, enhancing global contextual information and obtaining deep prediction layer features. Figure 3 Features of deep prediction layers Figure 4 ;

[0035] Thus far, the feature extraction network has produced a total of 7 prediction layer feature maps at different levels of the output.

[0036] Step 2.2.3: The 7 prediction layer feature maps are respectively branched from the 7 output ends of the feature extraction network into the ADR module in the corresponding prediction branch, and the three-way feature fusion logic of spatial saliency branch, local texture refinement branch and residual branch is executed;

[0037] The feature map of the prediction layer is first input to the spatial saliency branch, where it is aggregated and sampled by the response extrema and response mean through the channel pooling concatenation unit. After cascaded standard convolution, batch normalization, and activation processing, saliency weights representing the importance of spatial regions are generated. The calculation formula is as follows:

[0038]

[0039] in, Input the prediction layer feature map of the ADR module into the current prediction branch. and These are the max pooling and average pooling operations on the channel dimension, respectively. and All are standard convolution operations, with ReLU being the ReLU activation function;

[0040] Simultaneously, the predicted layer feature map enters the local texture refinement branch, where a deep convolutional layer extracts local high-frequency signals to form a texture feature stream representing the target edge. The residual branch performs an identity mapping to preserve the feature map of the prediction layer. semantic information;

[0041] Significance weight Spatial gain weights are obtained by performing offset processing on the weight gain unit in the end-fusion component. Spatial gain weights of spatial saliency branches and texture feature flow of local texture refinement branches Multiplying within the element-wise multiplication unit, and then using adaptive coefficients via the weight compensation unit. After weighted compensation, the prediction layer feature map is preserved along with the residual branches. Fusion is performed within the element-wise addition unit, and the corrected features are output after processing by the ReLU6 activation layer. The calculation formula is as follows:

[0042]

[0043] Step 2.2.4: By establishing a compacted prior box sequence as an absolute spatial reference, the location regression parameters and confidence scores are extracted simultaneously using the detection head to predict the target location and category;

[0044] At the prediction layers corresponding to the 7 prediction branches, a set of compacted prior box sequences are configured respectively; based on the target size distribution in the infrared UAV dataset in step 1, the basic size and number of prior boxes corresponding to each prediction layer are predetermined so that the prior boxes are spatially adapted to the geometric features of the targets in the infrared UAV dataset; the geometric center of each prior box is aligned with the pixel receptive field center of the corresponding prediction layer feature map to provide an absolute spatial reference benchmark for the results output by the detection head;

[0045] Corrected features of the ADR module output in the prediction branch The data is input to the detection head, and the location regression convolutional layer group predicts the relative offset of the center point and the logarithmic scaling ratio of the width and height for each prior box; at the same time, the confidence prediction convolutional layer group predicts the confidence score of the probability that each prior box contains a UAV target.

[0046] Finally, the relative offset of the center point output by the detection head and the logarithmic scaling ratio of width and height are applied to the absolute space reference through coordinate decoding logic to achieve the inverse transformation from relative regression parameters to absolute space coordinates of the image, thus constructing the original prediction candidate set. Furthermore, non-maximum suppression (NMS) is used to remove redundancy from the original prediction candidate set, retaining only the prediction box with the highest confidence score, and outputting the final infrared UAV target detection result.

[0047] Step 3: Use the training set divided in Step 1 to perform iterative training on the SCAD-SSD model constructed in Step 2; after initializing the hyperparameters and weights, input the training set into the SCAD-SSD model in batches for end-to-end training; drive the gradient update of the network weights through multi-task loss and adjust the hyperparameters to obtain the optimal detection model with infrared UAV target recognition capability.

[0048] Step 3.1: During the training startup phase, perform random initialization of the hyperparameters of the SCAD-SSD model; the hyperparameters include the initial learning rate, batch size, optimizer type, momentum factor, and weight decay coefficient.

[0049] Step 3.2: Input the images in the training set into the SCAD-SSD model in batches and output the infrared UAV target detection results. Compare the infrared UAV target detection results with the label values ​​determined by the structured annotation file and prior box matching. Calculate the loss value through the multi-task loss function and update the network weights using backpropagation.

[0050] Step 3.2.1: Input the images in the training set into the SCAD-SSD model in batches. After preprocessing by spatial scale reconstruction and mean-standard deviation correction, the images are processed by the SCB module in the feature extraction network for multi-scale feature extraction and target background environment perception. The extracted prediction layer feature maps are then input into 7 prediction branches. The ADR module in the prediction branches is used to mine salient features to suppress infrared background false alarm noise. Finally, the detection head outputs preliminary detection results containing probability confidence scores and location regression parameters.

[0051] Step 3.2.2: Using the absolute spatial reference benchmark generated in Step 2.2.4, the location information of the real target boxes recorded in the structured annotation file in Step 1.2 is mapped to the corresponding feature map space; by calculating the intersection-union ratio of the prior boxes and the real target boxes, the prior boxes that meet the preset threshold conditions are established as positive sample prior boxes and real offset labels are generated; a hard negative sample mining strategy is used to filter the background region to obtain negative sample prior boxes and generate corresponding category labels;

[0052] Step 3.2.3: Compare the position regression parameters and probability confidence scores output by the detector head in Step 3.2.1 with the true offset labels and category labels determined in Step 3.2.2; calculate the total loss value using the multi-task loss function. The weights of the SCAD-SSD model are updated layer by layer through error backpropagation; the formula for calculating the total loss is as follows:

[0053]

[0054] Where N is the number of positive prior boxes that meet the preset threshold condition. For confidence loss, To locate the loss; This is the balance coefficient;

[0055] Step 3.3: Monitor the changing trend of the validation set loss value and determine whether it has reached the global optimal convergence state; if it is satisfied, then stop training and obtain the optimal detection model with UAV target recognition capability; otherwise, repeat the training logic of step 3.2 by adjusting the hyperparameters of step 3.1.

[0056] A hardware acceleration method for a drone early warning system based on SCAD-SSD, comprising the following steps:

[0057] Step 1: Deploy a hardware accelerator for the UAV early warning system based on SCAD-SSD on the programmable logic (PL) end of the field-programmable gate array (FPGA), and coordinate with the processing system (PS) end to execute the forward inference of the optimal detection model obtained in Step 3. The hardware accelerator consists of a multi-dimensional parallel memory access unit, a computing cluster composed of two general-purpose parameterized reconfigurable engines UPRE 0 and UPRE 1, and a global controller. During the forward inference process, the global controller implements task scheduling, controls the multi-dimensional parallel memory access unit to perform data transfer, and coordinates the computing cluster to switch between different modes to complete the inference of the target recognition network.

[0058] Step 1.1: Construct a hardware accelerator for the UAV early warning system based on SCAD-SSD; among them, the multi-dimensional parallel memory access unit acts as the memory access hub, responsible for handling the data interaction between the system-side memory PS_DDR and the hardware accelerator; the general-purpose parameterized reconfigurable engine in the computing cluster acts as the core computing power unit, responsible for realizing the logical mapping of the initial sampling layer, SCB module, end-level dimensionality increase layer, ADR module and detection head in the SCAD-SSD model through the reconfiguration of hardware resources; the global controller acts as the command center, responsible for parsing control commands and coordinating the collaborative work of memory access and computing resources;

[0059] Step 1.1.1: The multidimensional parallel memory access unit integrates a DMA controller, a bit-width converter, a row cache module, and a double-buffered architecture. The DMA controller handles data transfer between the system-side memory PS_DDR and the cache array on the programmable logic PL (Programmable Gate Array) of the FPGA. The bit-width converter enables bidirectional conversion between 64-bit and 8-bit data. The row cache module temporarily stores the pixel data of the serially input correction image and reconstructs it into a feature map data sliding window that meets the local receptive field requirements of the convolutional kernel. The double-buffered architecture includes Buffer 0 and Buffer 1, employing a ping-pong operation mechanism to asynchronously prefetch the convolutional kernel weight parameters, thereby achieving parallel pipelined memory access and computation.

[0060] Step 1.1.2: The general-purpose parametric reconfigurable engine integrates a processing array, an interconnect control module, a nonlinear activation unit, a residual addition unit, and a weighting module. The processing array integrates a high-performance fixed-point multiplier and an addition tree logic, responsible for performing multiplication and addition operations on various convolutional layers in the optimal detection model. The interconnect control module dynamically adjusts the data flow according to the configuration signal, enabling the general-purpose parametric reconfigurable engine to adapt to the computational requirements of the SCB and ADR modules in the optimal detection model. The nonlinear activation unit performs activation mapping through an internally hardened lookup table (LUT) circuit. The residual addition unit is responsible for pixel fusion and overlay. The weighting module is responsible for requantizing and truncating the high-bit-width intermediate results produced by various convolutional layers, ensuring that the feature data or prediction results output to the system-side memory PS_DDR remain at an 8-bit fixed-point width.

[0061] Step 1.1.3: The global controller drives the two general parameterized reconfigurable engines to switch between parallel working mode and collaborative working mode according to the network topology logic; at the same time, the global controller is responsible for configuring the memory access address and transmission length of the DMA controller, and coordinating the data synchronization status of the double buffer and the line buffer to realize on-chip pipelined inference of the optimal detection model.

[0062] Step 1.2: After parsing the instructions from the processing system PS, the global controller coordinates the multi-dimensional parallel memory access units and the computing cluster to perform forward inference. During the forward inference process, the global controller drives the computing cluster to switch between parallel working mode and collaborative mode according to the data dependency and distribution characteristics of the network topology logic, and completes the computation tasks of the feature extraction backbone, extension layer and prediction branch. Finally, the prediction results produced by the computing cluster are written back to the system memory PS_DDR, and the processing system PS completes the post-processing of the results and coordinate output.

[0063] Step 1.2.1: The global controller parses the inference start instruction from the processing system PS and begins to accelerate inference. First, the multidimensional parallel memory access unit asynchronously prefetches network weights and reads the pixels of the image to be detected. The computation cluster is configured to work in parallel mode, with UPRE 0 and UPRE 1 processing the initial sampling layer, SCB 1-6, SCB 8-13, SCB 15-17, and the final up-dimensional layer in the feature extraction network in sequence. At the same time, the multidimensional parallel memory access unit continuously performs asynchronous data prefetching on the weights according to the network level index to compensate for the memory access latency of the system-side memory PS_DDR.

[0064] Step 1.2.2: When the pipeline advances to the prediction branch of the feature extraction backbone, UPRE 0 and UPRE 1 jointly complete the calculation of SCB 6, SCB13 or the final dimension-upgrading layer, producing the corresponding shallow prediction layer feature map or mid-level prediction layer feature map, resulting in two data flows: one flows into the corresponding next-level SCB 7, SCB 14 or E-SCB 1 in the extended layer, and the other flows out to the ADR module in the corresponding prediction branch; based on this data splitting characteristic, the global controller drives the computation cluster to switch to a collaborative mode; UPRE 0 is responsible for continuing to execute the calculation of the SCB module in the feature extraction network, and UPRE 1 is responsible for executing the ADR module and detection head operation of the corresponding prediction branch;

[0065] Depending on the network depth, the computation cluster collaboration mode exhibits two waiting states: First, for the shallow prediction layer feature maps produced by SCB 6 and SCB 13, UPRE 1 enters the idle state after completing the prediction branch operation to wait for UPRE 0 to complete the subsequent SCB 7 and SCB 14 calculations in the feature extraction backbone; Second, for the mid-level prediction feature maps produced by the terminal up-dimensional layer, the computational amount required for E-SCB 1, which is handled by UPRE 0, is less than the computational amount required for the prediction branch handled by UPRE 1, so UPRE 0 enters the idle state in advance to wait for UPRE 1.

[0066] Step 1.2.3: When the inference pipeline delves into the extended layer stage, the global controller drives the computation clusters to switch back completely and maintain parallel operation, sequentially executing computation tasks. UPRE 0 and UPRE 1 work together to sequentially complete the ADR module and detection head operation corresponding to E-SCB 1, E-SCB 2 operation, E-SCB 2's ADR module and detection head operation, E-SCB 3 operation, E-SCB 3's ADR module and detection head operation, E-SCB 4 operation, and E-SCB 4's ADR module and detection head operation. The global controller, through high-frequency task switching and dual-engine joint operation, quickly digests the dense prediction computation load of the extended layer, ensuring that the feature maps of the seven prediction layers are generated sequentially.

[0067] Step 1.2.4: The prediction results output by each detection head are generated by the computing cluster and written back to the system-side memory PS_DDR; the processing system PS end combines the preset prior box to complete the coordinate decoding and non-maximum suppression (NMS) processing of the prediction results, and finally outputs the position coordinates and category confidence of the infrared weak target.

[0068] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0069] This invention addresses the challenges of weak target signals and complex environmental backgrounds in long-range infrared detection by proposing a UAV early warning system based on SCAD-SSD and its hardware acceleration method that balances high detection rate and low false alarm rate.

[0070] For UAV early warning systems based on SCAD-SSD, this invention improves network input resolution by introducing spatial scale reconstruction to increase sampling redundancy for weak point source signals, and establishes a compact prior box sequence as an absolute spatial reference benchmark to adapt to the geometric features of the target. These strategies effectively mitigate signal loss and annihilation of target features during network downsampling and deep propagation, significantly improving the sensitivity for capturing infrared UAV targets. Simultaneously, a feature extraction network is constructed with the SCB module as its core, and an ADR module is cascaded in the prediction branch, forming a closed-loop mechanism of "feature enhancement-background suppression." Specifically, the SCB module enhances the target response by utilizing long-range environmental cues captured by its internal side units, achieving feature enhancement; the ADR module actively filters out high-frequency background clutter by mining spatially salient features, achieving background suppression. Finally, the detection head outputs position regression parameters and probability confidence scores. This system achieves a precise balance between detection performance and computational power consumption while maintaining a lightweight architecture.

[0071] This invention addresses a hardware acceleration method for UAV early warning systems based on SCAD-SSD. It deploys a hardware accelerator on the FPGA's programmable logic (PL) side, utilizing a bit-width converter in a multi-dimensional parallel memory access unit to achieve bidirectional data conversion, physically aligning bus bandwidth with computational requirements. A dual-buffered architecture with a ping-pong operation mechanism asynchronously prefetches convolution kernel weight parameters, coupled with a row cache module, effectively hiding data transfer latency and reducing memory access frequency. Furthermore, a global controller implements task scheduling, driving the computing cluster to switch between parallel and collaborative modes, enabling on-chip pipelined inference for the optimal detection model. This acceleration method effectively solves the problems of data diversion and resource imbalance at the prediction branch exit point, significantly reduces inference pipeline cavitation, and greatly improves the real-time detection efficiency of forward inference, demonstrating significant engineering practical value. Attached Figure Description

[0072] Figure 1 This is a schematic diagram of the overall architecture of the SCAD-SSD model of the present invention;

[0073] Figure 2 This is a schematic diagram of the SCB module of the present invention;

[0074] Figure 3 This is a schematic diagram illustrating the structural principle of the ADR module of the present invention;

[0075] Figure 4 This is a diagram showing the overall hardware architecture of the hardware-software co-acceleration system of the present invention.

[0076] Figure 5 This is a hardware logic architecture diagram of the multidimensional parallel memory access unit of the present invention;

[0077] Figure 6 This is a logical architecture diagram of the Universal Parametric Reconfigurable Engine UPRE of the present invention;

[0078] Figure 7 This is a pipeline timing diagram in an embodiment of the present invention. Detailed Implementation

[0079] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings and technical solutions.

[0080] Example

[0081] A UAV early warning system and hardware acceleration method based on SCAD-SSD are proposed. At the early warning system level, a complete infrared UAV dataset is constructed, and an SCAD-SSD model with an SCB module as the core and an ADR module for feature correction is built. The optimal detection model with infrared UAV target recognition capability is obtained through iterative training. At the hardware acceleration level, a hardware accelerator consisting of a computing cluster composed of multi-dimensional parallel memory access units, a general-purpose parameterized reconfigurable engine, and a global controller is deployed on the programmable logic (PL) side of the FPGA. During forward inference, task scheduling is implemented by the global controller, and the multi-dimensional parallel memory access units are controlled to perform efficient data transfer. The general-purpose parameterized reconfigurable engine is driven to realize the logical mapping of the initial sampling layer, SCB module, end-level dimensionality increase layer, ADR module, and detection head in the optimal detection model through dynamic reconfiguration of hardware resources. The computing cluster switches between parallel working mode and division of labor mode, and finally realizes on-chip pipelined inference for the optimal detection model.

[0082] A drone early warning system based on SCAD-SSD, the specific steps of which are as follows:

[0083] Step 1: Continuously track UAV targets using an infrared sensor, collect infrared video sequences in different scenarios such as deep space, complex cloud layers, and near-ground objects, extract the original infrared images from the infrared video sequences containing UAV targets as positive samples; collect pure background images without UAV targets as negative samples, and supplement them with isolated images containing UAV targets that are prone to false alarms as expanded positive samples to construct a complete infrared UAV dataset.

[0084] Subsequently, the bounding rectangles of the drone targets in each image of the infrared drone dataset are located using annotation tools. The center coordinates and width and height information of the drone targets in the images are obtained and normalized, which is called the true target bounding box position information. The category labels of the drone targets and the true target bounding box position information are recorded to generate a structured annotation file.

[0085] Finally, the constructed infrared drone dataset is divided: first, the original infrared images belonging to the same infrared video sequence are treated as a whole and assigned to a single set; then, the negative samples and expanded positive samples are assigned to different sets so that the divided training set, validation set and test set meet the preset ratio of 8:1:1.

[0086] Step 2: First, acquire images from the complete infrared drone dataset, then adjust the images to... The network's preset size. In this embodiment, since the original image size is... To ensure that the image content maintains its aspect ratio during reconstruction and avoids geometric distortion, a scaling factor of 1 is set for both the height and width directions. The 512 rows of effective height pixels are centered, and the blank areas at the top and bottom (64 rows each) are uniformly filled with grayscale using a pixel value of 128, resulting in a scaled reconstructed image. Although the scaling factor in this embodiment is 1, this system defines a cubic spline kernel function... This ensures the algorithm's compatibility with heterogeneous resolution inputs. The formula for calculating a piecewise cubic polynomial is as follows:

[0087]

[0088] Among the control parameters , This is the normalized relative distance between the pixel to be interpolated and its neighboring sampled pixels.

[0089] Subsequently, mean and standard deviation correction is performed on the scale-reconstructed image: pixel values Normalize to the [0,1] interval and set the mean. Standard deviation and according to the formula Calculate the pixel values ​​of the corrected image. .

[0090] The overall architecture of the constructed SCAD-SSD model is as follows: Figure 1 As shown, the SCAD-SSD model consists of a feature extraction network and seven prediction branches. The feature extraction network is composed of a feature extraction backbone and extended layers cascaded sequentially: the feature extraction backbone starts at an output channel of 32 and a step size of... The network consists of an initial sampling layer of 2, followed by 17 consecutive SCB modules (divided into 7 feature extraction stages, Stage 1 to Stage 7), and finally a final upscaling layer with an output channel of 1280. The parameter configurations for each feature extraction stage are shown in Table 1. The expansion layer consists of 4 consecutive SCB modules (defined as E-SCB 1 to E-SCB 4 in hierarchical order). The prediction branches are composed of an ADR module and a detection head cascaded sequentially, and the 7 prediction branches are coupled to the 7 output ends of the feature extraction network.

[0091] Table 1. Parameter configuration table for each feature extraction stage

[0092]

[0093] like Figure 2 The diagram shows the structure of the SCB module of this invention. Each SCB module is statically configured as an initial feature enhancement topology, a downsampling topology, or a standard feature enhancement topology according to its layer.

[0094] Taking the SCB module configured as a standard feature enhancement topology as an example, observe... Figure 2 Data flow: Feature data first flows into the channel expansion layer, where dimensionality augmentation is performed to obtain intermediate features for feature enhancement. Subsequently, the data stream branches, inputting in parallel to the upper main branch unit and the lower side branch unit: the main branch unit utilizes a step size of 1. Deep convolutional layers extract local detail features The side support unit utilizes Deeply dilated convolutional layers (diffraction rate of 2) capture long-range environmental cues, and batch normalization is concatenated with sigmoid activation layers to generate spatial attention weights. The calculation formula is as follows:

[0095]

[0096] Local detail features output by the main branch unit Spatial attention weights of the side unit output Element-wise multiplication is performed by a projective convolutional layer to perform channel fusion and dimensionality compression, generating output features. The calculation formula is as follows:

[0097]

[0098] Furthermore, when the SCB module is configured as a downsampling topology, the feature data is processed through a channel expansion layer to obtain downsampling intermediate features. Subsequently, only the main branch unit performs space compression. The stride of the deep convolutional layer is switched to 2, and the computation of the lower side branches is disabled. Finally, the output features are generated by the projection convolutional layer. The calculation formula is as follows:

[0099]

[0100] The extended layer in the feature extraction network consists of four consecutive E-SCB 1, E-SCB 2, E-SCB 3, and E-SCB 4. In terms of static topology configuration, E-SCB 1 and E-SCB 2 are configured as a downsampling topology to perform continuous spatial compression and receptive field expansion; E-SCB 3 and E-SCB 4 are configured as a standard feature enhancement topology to utilize long-range environmental cues provided by the side units for spatial attention weighting, thereby enhancing global contextual information.

[0101] During forward inference, the corrected image is first input to the initial sampling layer of the feature extraction backbone to perform preliminary spatial dimensionality reduction, outputting low-dimensional feature data. This low-dimensional feature data sequentially passes through all SCB modules in the feature extraction backbone for feature enhancement or spatial downsampling. During its flow through the SCB modules, after processing by SCB 6 and SCB 13, two shallow prediction layer feature maps are first extracted. Subsequently, after the feature data has traversed all 17 SCB modules of the feature extraction backbone, it is input to the final dimensionality-enhancing layer, where pointwise convolution is used to perform linear mapping and dimensionality expansion of the channel dimensions, outputting high-dimensional feature data with strong semantic expressive power, which is then used as the feature map for the middle prediction layer.

[0102] Simultaneously, the high-dimensional feature data sequentially enters the extended layer consisting of E-SCB 1 to E-SCB 4, and the deep prediction layer features are extracted from the outputs of E-SCB 1 and E-SCB 2. Figure 1 Features of deep prediction layers Figure 2 Features of the deep prediction layer are derived from the outputs of E-SCB 3 and E-SCB 4. Figure 3 Features of deep prediction layers Figure 4 It constructs a deep semantic space and captures global contextual information.

[0103] Thus far, the feature extraction network has produced a total of 7 prediction layer feature maps at different levels, which are then branched into the corresponding 7 prediction branches.

[0104] like Figure 3 The diagram shows the structural principle of the ADR module of this invention. For the shallow prediction layer feature maps derived from SCB 6 and SCB 13, an L2Norm operator is cascaded for dimensional normalization before entering the ADR module, and a learnable scaling factor with an initial value of 20 is introduced. Features entering the ADR module... Simultaneously, traffic is split into three branch paths:

[0105] In spatially salient branches, features The response extrema and mean are aggregated and sampled by the channel pooling concatenation unit, then passed through a cascaded standard convolutional layer, batch normalization layer, ReLU activation layer, and another standard convolutional layer. Finally, the saliency weights are output by the sigmoid activation layer. The calculation formula is as follows:

[0106]

[0107] In the local texture refinement branch, a deep convolutional layer is used to extract local high-frequency signals, which are then batch normalized and ReLU activated to form a texture feature stream representing the target edge. Simultaneously, the residual branches perform identity mappings to preserve features. The original semantic information of the graph.

[0108] In the end-fusion component, saliency weights Spatial gain weights are obtained by performing offset processing on the weighted gain unit. It is multiplied with the texture feature flow of the local texture refinement branch in the element-wise multiplication unit, and then processed by the weight compensation unit using an adaptive coefficient with an initial value of 0.1. After weighted compensation, the residual branches are retained. Fusion is performed within the element-wise addition unit, and the final output is the corrected feature after processing by the ReLU6 activation layer. The calculation formula is as follows:

[0109]

[0110] A compacted prior box sequence was established as an absolute spatial reference. Based on the size distribution of targets in the infrared UAV dataset, a set of compacted prior box sequences was configured at the prediction layers corresponding to the seven prediction branches, aligning the geometric center of each prior box with the center of the pixel receptive field of the corresponding prediction layer feature map. The specific compacted prior box sequence configuration is shown in Table 2.

[0111] Table 2 Compacted Prior Frame Sequence Configuration

[0112]

[0113] Corrected features of the ADR module output in the prediction branch The data is input to the detection head, where a group of position regression convolutional layers and a group of confidence prediction convolutional layers simultaneously predict the relative offset of the center point, the logarithmic scaling ratio of width and height, and the probability confidence score of containing a UAV target for each prior box. Finally, the relative regression parameters output from the detection head are inversely transformed to an absolute spatial reference using coordinate decoding logic to construct the original prediction candidate set. Redundancy is further eliminated using non-maximum suppression (NMS), and the infrared UAV target detection result is output.

[0114] Step 3: During the training startup phase, the hyperparameters of the SCAD-SSD model are first randomly initialized. Based on the feature distribution of infrared targets, this embodiment sets the initial learning rate to 10. -3 A momentum-driven stochastic gradient descent (SGD) optimizer was employed, with the momentum factor set to 0.937 and the weight decay coefficient set to [value missing]. It also configures a cosine annealing strategy to dynamically adjust the learning rate during the training cycle.

[0115] Subsequently, the images from the training set divided in step 1 are input into the SCAD-SSD model in batches. After spatial scale reconstruction and mean-standard deviation correction preprocessing, the images are processed by a feature extraction network for multi-scale feature extraction. The ADR module in the prediction branch is used to mine salient features to suppress background clutter. Finally, the detection head outputs preliminary detection results containing probability confidence scores and location regression parameters.

[0116] In the error calculation and sample matching stage, the absolute spatial reference base generated in step 2 is used to map the real target box location information recorded in the structured annotation file to the corresponding feature map space. By calculating the intersection-over-union (IoU) ratio between the prior boxes and the real target boxes, the prior boxes that meet the preset threshold conditions are established as positive sample prior boxes, and real offset labels are generated. To alleviate the problem of uneven distribution of positive and negative samples, this embodiment adopts a hard negative sample mining strategy to perform strict screening of the background region, controlling the ratio of negative to positive samples to 3:1, obtaining negative sample prior boxes and generating corresponding category labels.

[0117] The position regression parameters and probability confidence scores output by the detection head are compared with the determined true offset labels and category labels, respectively. The total loss value is calculated using a multi-task loss function, and forward propagation and backpropagation are performed to update the weights of the SCAD-SSD model. The formula for calculating the total loss value is as follows:

[0118]

[0119] in, The positioning loss is calculated using the Smooth L1 loss function, and the balance coefficient is... Set to 1.0.

[0120] The changing trend of the validation set loss value is continuously monitored to determine whether it has reached a good global optimal convergence state. If this condition is met, training is stopped; otherwise, the above-mentioned network weight update logic based on multi-task loss is repeatedly executed by adjusting hyperparameters such as the initial learning rate, batch size, optimizer type, momentum factor, or weight decay coefficient as appropriate. Through continuous iterative optimization, the optimal detection model with infrared UAV target recognition capability is finally obtained.

[0121] A hardware acceleration method for a drone early warning system based on SCAD-SSD is as follows:

[0122] A hardware accelerator is deployed at the programmable logic (PL) end of the field-programmable gate array (FPGA) to perform forward inference of the optimal detection model obtained in step 3, in conjunction with the processing system's (PS) end. The overall architecture diagram of the hardware accelerator is shown below. Figure 4 As shown, this hardware accelerator consists of a multidimensional parallel memory access unit, a computing cluster composed of two general-purpose parameterized reconfigurable engines, UPRE 0 and UPRE 1, and a global controller. During forward inference, the global controller performs task scheduling, controls the multidimensional parallel memory access unit to perform data transfer, and coordinates the computing cluster to switch between different modes to complete the inference of the target recognition network.

[0123] The hardware logic architecture of the multidimensional parallel memory access unit is as follows: Figure 5 As shown, this multidimensional parallel memory access unit acts as a memory access hub, responsible for handling data interaction between the system-side memory PS_DDR and the hardware accelerator. During data transfer, its internal bit-width converter enables bidirectional conversion between 64-bit and 8-bit data; the line cache module temporarily stores the pixel data of the serially input corrected image and reconstructs it into a feature map data sliding window that meets the local receptive field requirements of the convolutional kernel; the dual-buffered architecture includes Buffer 0 and Buffer 1, and uses a ping-pong operation mechanism to asynchronously prefetch the fixed-point convolutional kernel weight parameters to achieve parallel pipelined memory access and computation.

[0124] The logical architecture of the general parameterized reconfigurable engine UPRE is as follows: Figure 6As shown, a computational cluster composed of two general-purpose parameterized reconfigurable engines performs efficient computations under the coordination of multi-dimensional parallel memory access units. Each engine achieves logical mapping of the initial sampling layer, SCB module, end-level upscaling layer, ADR module, and detection head in the optimal detection model through the dynamic combination of processing arrays and interconnect control modules. In specific implementation: the processing array is responsible for performing the core convolution multiplication and addition operations; the interconnect control module synchronously adjusts the data flow direction dynamically according to the configuration signal and performs topology reconstruction to adapt to specific computational requirements; at the same time, the nonlinear activation unit integrated inside the engine carries activation mappings such as Sigmoid and ReLU6 through the internally hardened lookup table (LUT) circuit, and supports the switching of activation paths and pixel fusion and superposition of parallel branches in conjunction with the residual addition unit; finally, the weighting module performs requantization and bit width truncation on the high bit width intermediate results to ensure that the data output to the system-side memory PS_DDR maintains an 8-bit fixed-point bit width.

[0125] The global controller, acting as the command center for the hardware accelerator, is responsible for parsing control commands from the processing system's PS (Power Supply). Combined with... Figure 7 The pipeline timing diagram shown illustrates that after the global controller parses the inference start instruction, it coordinates the multi-dimensional parallel memory access units and the computing cluster to collaboratively execute forward inference. The specific dynamic scheduling logic is as follows:

[0126] During the pipeline startup phase, the multidimensional parallel memory access unit enables asynchronous data prefetching, loading the pixels of the image to be detected and the network weights to mask the memory access latency of accessing the system-side memory PS_DDR. Simultaneously, the global controller configures the computation clusters to operate in parallel, driving the combined computing power of the dual engines UPRE 0 and UPRE 1 to sequentially execute the branchless computational tasks in the feature extraction backbone, i.e., jointly completing the operations of the initial sampling layer, SCB 1-6, SCB 8-13, SCB 15-17, and the final dimension-upgrading layer.

[0127] As time progresses, when feature data flows through the outgoing end with branch topology (i.e., SCB 6, SCB 13, or the final upscaling layer is completed), the resulting prediction layer feature map generates two data flows. To address this data splitting characteristic, the global controller drives the computing cluster to switch to a collaborative mode to decouple computing power: UPRE 0 is scheduled to handle the computation of the next-level SCB module in the feature extraction backbone (such as SCB 7, SCB 14, or E-SCB 1), while UPRE 1 is scheduled to handle the newly introduced prediction branch task, which includes the corresponding level's ADR module and the detection head calculation.

[0128] Due to the difference in computational load under the collaborative task mode, the two engines in the computing cluster exhibit two waiting states: For example, when processing shallow prediction layer feature maps, the computational load of the prediction branch is relatively small, and UPRE 1 enters an idle state after completing the branch operation to wait for UPRE 0 to complete its computation. Conversely, if the computational power required by SCB 7, SCB 14, or E-SCB 1 in the feature extraction backbone is lower than that of the prediction branch (e.g., when flowing to E-SCB 1), then UPRE 0 enters an idle state earlier to wait for UPRE 1. Once both engines have completed their respective computational tasks, the global controller switches the idle engine back to parallel operation mode.

[0129] When the inference pipeline enters the extended layer stage, the global controller drives the computation cluster to switch back to parallel operation mode to handle the intensive prediction computation load of this stage. In this mode, UPRE 0 and UPRE 1 work together at high frequency to sequentially complete the ADR module and detection head operations corresponding to E-SCB 1, E-SCB 2, E-SCB 2, E-SCB 3, E-SCB 3, E-SCB 4, and E-SCB 4, ensuring the efficient generation and processing of the feature maps of the seven prediction layers.

[0130] After all computation tasks are completed, the prediction results produced by each detection head are written back to the system-side memory PS_DDR via the bus by the computation cluster. The processing system PS performs coordinate decoding and non-maximum suppression (NMS) processing in conjunction with preset prior box parameters. By calculating the intersection-over-union (IoU) ratio of overlapping boxes, redundant boxes with low scores are suppressed to filter background false alarms and lock onto a single target. When restoring the predicted target's position in the original image, the processing system PS subtracts the 64-pixel offset introduced by spatial scale reconstruction in step 2, mapping the coordinates at the enhanced resolution back to the original low-resolution image space, thereby obtaining the target's precise pixel coordinates and completing the closed-loop inference acceleration of the entire algorithm chain.

Claims

1. A drone early warning system based on SCAD-SSD, characterized in that, Includes the following steps: Step 1: Construct a complete infrared UAV dataset, perform structured annotation on the infrared UAV dataset, and then divide it into training, validation, and test sets; Step 2: Construct the SCAD-SSD model consisting of a feature extraction network and prediction branches; acquire images from the complete infrared UAV dataset constructed in Step 1 and perform spatial scale reconstruction, uniformly adjusting them to the network's preset size to obtain scale-reconstructed images; after mean and standard deviation correction, the scale-reconstructed images are sequentially passed through a feature extraction network with the SCB module as its core, producing 7 prediction layer feature maps, which are then branched into 7 corresponding prediction branches; the prediction branches perform feature correction for spatial saliency and local texture refinement through the ADR module, and the detection head produces position regression parameters and probability confidence scores; finally, combined with the absolute spatial reference benchmark, the infrared UAV target detection results are output through coordinate decoding and redundancy removal; Step 2.1: Build the SCAD-SSD model, which consists of a feature extraction network and 7 prediction branches; the feature extraction network is composed of a feature extraction backbone and an extension layer cascaded sequentially; the prediction branches are composed of an ADR module and a detection head cascaded sequentially; the 7 prediction branches are coupled to the 7 outputs of the feature extraction network respectively. Step 2.2: Input the images from the infrared UAV dataset obtained in Step 1 into the SCAD-SSD model constructed in Step 2.

1. Obtain the corrected image by preprocessing the image through spatial scale reconstruction and mean standard deviation correction. Seven prediction layer feature maps are generated in the feature extraction network. The ADR module in the prediction branch performs feature fusion and correction. Then, the position regression and confidence prediction are performed by the detection head. Finally, the candidate box coordinate decoding and redundancy removal are completed by combining the absolute spatial reference benchmark to realize the inference prediction of infrared UAV targets. Step 3: Use the training set divided in Step 1 to perform iterative training on the SCAD-SSD model constructed in Step 2; after initializing the hyperparameters and weights, input the training set into the SCAD-SSD model in batches for end-to-end training; drive the gradient update of the network weights through multi-task loss and adjust the hyperparameters to obtain the optimal detection model with infrared UAV target recognition capability.

2. The UAV early warning system based on SCAD-SSD according to claim 1, characterized in that, The specific implementation process of step 1 is as follows: Step 1.1: Use an infrared camera to continuously track drone targets, collect infrared video sequences containing drone targets in different scenarios, extract the original infrared images from the infrared video sequences as positive samples, collect pure background images without drone targets as negative samples, and supplement isolated images containing drone targets in scenarios that are prone to false alarms as expanded positive samples to construct a complete infrared drone dataset. Step 1.2: Use the annotation tool to locate the bounding rectangles of the drone targets in each image of the infrared drone dataset, obtain the center coordinates and width and height information of the drone targets in the image and normalize them; record the category label of the drone targets and the real target box position information, the real target box position information including the normalized center coordinates and width and height information, and generate the corresponding structured annotation file. The infrared UAV dataset constructed in step 1.1 is divided into training set, validation set and test set. First, the original infrared images belonging to the same infrared video sequence are classified into a single set as a whole. Then, the negative samples and expanded positive samples are assigned to different single sets so that the divided training set, validation set and test set meet the preset ratio.

3. The UAV early warning system based on SCAD-SSD according to claim 2, characterized in that, The specific implementation process of step 2.1 is as follows: Step 2.1.1: The feature extraction backbone consists of an initial sampling layer, 17 consecutive SCB modules, and a final dimension-upgrading layer connected in sequence; the initial sampling layer contains sequentially connected... Standard convolutional layers, batch normalized layers, and ReLU6 activation layers; 17 consecutive SCB modules, SCB 1-SCB 17; the final upscaling layer contains sequentially connected... The system consists of pointwise convolutional layers, batch normalization layers, and ReLU6 activation layers; the extended layer is composed of four consecutive SCB modules, defined as E-SCB 1, E-SCB 2, E-SCB 3, and E-SCB 4 in hierarchical order. Each SCB module in the feature extraction backbone and extended layers is statically configured according to its layer level as one of the following: initial feature enhancement topology, downsampling topology, or standard feature enhancement topology. Specifically, SCB 1 in the feature extraction backbone is configured as the initial feature enhancement topology, with its input being an Identity unit performing an identity mapping, its intermediate layers containing parallel main and side units, and its output being an element-wise multiplication unit and a projective convolutional layer. SCB 2, SCB 4, SCB 7, and SCB 14 in the feature extraction backbone and E-SCB 1 and E-SCB 2 in the extended layers are configured as a downsampling topology, consisting of sequentially cascaded channel expansion layers, main units, and projective convolutional layers. The remaining SCBs in the feature extraction backbone and E-SCB 3 and E-SCB 4 in the extended layers are configured as standard feature enhancement topologies, with their input being a channel expansion layer, and the structure of their intermediate and output layers being consistent with that of the initial feature enhancement topology. The channel expansion layer contains sequentially connected... Pointwise convolutional layers, batch normalized layers, and ReLU6 activation layers; main branch units contain sequentially connected... The layers consist of depthwise convolutional layers, batch normalization layers, and ReLU6 activation layers. The deep convolutional layer has a stride of 2 in the downsampling topology and a stride of 1 in the initial feature enhancement topology and the standard feature enhancement topology; the side branch units contain sequentially connected units with a stride of 1 and an expansion rate of 2. The system consists of depthwise dilated convolutional layers, batch normalization layers, and sigmoid activation layers; the projected convolutional layers are formed by sequentially concatenating... It consists of pointwise convolutional layers and batch normalization layers; Step 2.1.2: The first two prediction branches out of the seven prediction branches are set at the outputs of SCB6 and SCB13 in the feature extraction backbone, respectively, and each is connected in series with an L2Norm operator; the third prediction branch is set at the output of the final upscaling layer; the fourth to seventh prediction branches are set at the outputs of E-SCB1, E-SCB2, E-SCB3 and E-SCB4 in the extended layer, respectively. The ADR module consists of a spatial saliency branch, a local texture refinement branch, a residual branch, and an end-fusion component. The spatial saliency branch is composed of channel pooling splicing units, standard convolutional layers, batch normalization layers, ReLU activation layers, standard convolutional layers, and sigmoid activation layers cascaded sequentially. The local texture refinement branch is composed of depth convolutional layers and batch normalization layers cascaded sequentially. The end-fusion component integrates a weight gain unit, an element-wise multiplication unit, a weight compensation unit, an element-wise addition unit, and a ReLU6 activation layer. The detection head consists of a set of position regression convolutional layers and a set of confidence prediction convolutional layers. The position regression convolutional layer set contains 7 position regression convolutional layers, and the confidence prediction convolutional layer set contains 7 confidence prediction convolutional layers. The output of the ADR module in each of the 7 prediction branches is coupled to the position regression convolutional layer and the confidence prediction convolutional layer in the corresponding detection head.

4. The UAV early warning system based on SCAD-SSD according to claim 3, characterized in that, The specific implementation process of step 2.2 is as follows: Step 2.2.1: Acquire images from the infrared UAV dataset constructed in Step 1, construct piecewise cubic polynomial basis functions using the cubic spline interpolation algorithm, perform spatial scale reconstruction on the images, and uniformly adjust them to [the required scale]. The pixels are used to obtain a scale-reconstructed image; then, mean and standard deviation correction is performed on the scale-reconstructed image to obtain a corrected image, calculated using the following formula: in, These are the pixel coordinates of the image. To reconstruct the pixel values ​​of the corresponding coordinate points in the image at different scales, To correct the pixel values ​​at corresponding coordinate points in the image; and These are the pixel mean and pixel standard deviation of all pixels in the scaled reconstructed image, respectively. Step 2.2.2: Input the corrected image into the initial sampling layer of the feature extraction backbone to perform preliminary spatial dimensionality reduction and output low-dimensional feature data; the low-dimensional feature data sequentially passes through all SCB modules in the feature extraction backbone to perform feature enhancement or spatial downsampling processing. After completing SCB 6 and SCB 13 processing, two shallow prediction layer feature maps are obtained; then, the feature dimension is increased and the semantic representation is enhanced through the end dimensionality enhancement layer of the feature extraction backbone to obtain the mid-level prediction layer feature map; then, it passes through the extended layers E-SCB 1-E-SCB 4 to construct a deep semantic space and capture global context information, obtaining four deep prediction layer feature maps, for a total of seven prediction layer feature maps. When feature data passes through an SCB module configured with an initial feature enhancement topology or a standard feature enhancement topology, intermediate feature enhancement features are first obtained through Identity units or channel expansion layers. Subsequently, the intermediate features are enhanced. Parallel inputs are fed to the main branch unit and the side branch unit; local detail features are extracted by the main branch unit through a depthwise convolution with a stride of 1. Simultaneously, the side units capture long-range environmental cues and generate spatial attention weights through depthwise dilated convolutions. The calculation formula is as follows: in, This is an expanded convolutional layer with an expansion rate of 2 and a stride of 1; BN is batch normalization. Use the Sigmoid activation function; Subsequently, local detail features Spatial attention weights Element-wise multiplication is performed, followed by channel fusion and dimensionality compression via a projective convolutional layer to generate output features. The calculation formula is as follows: in, This refers to pointwise convolution in a projective convolutional layer. This indicates element-wise multiplication; When feature data is configured as an SCB module with a downsampled topology, it first performs dimensionality augmentation through a channel expansion layer to obtain intermediate downsampled features. Subsequently, the main branch unit utilizes a depthwise convolutional layer with a stride of 2 to... Spatial downsampling is performed, and finally, the output features are generated through a projection convolutional layer. The calculation formula is as follows: After the feature data has traversed all 17 SCB modules of the feature extraction backbone, it is input into the end dimension-up layer. The linear mapping and dimension expansion of the channel dimension are performed by pointwise convolution, and the high-dimensional feature data with strong semantic expression is output as the feature map of the middle prediction layer. Simultaneously, the high-dimensional feature data sequentially enters the extension layer composed of E-SCB 1, E-SCB 2, E-SCB 3, and E-SCB 4 to obtain four deep prediction layer feature maps. The high-dimensional feature data passes through E-SCB 1 and E-SCB 2 configured with a downsampling topology, performing continuous spatial compression and receptive field expansion, resulting in a progressively decreasing output resolution and progressively increasing semantic abstraction, thus obtaining deep prediction layer feature map 1 and deep prediction layer feature map 2. Subsequently, deep prediction layer feature map 2 passes through E-SCB 3 and E-SCB 4 configured with a standard feature enhancement topology, utilizing long-range environmental cues provided by the side units for spatial attention weighting to enhance global contextual information, thus obtaining deep prediction layer feature map 3 and deep prediction layer feature map 4. Thus far, the feature extraction network has produced a total of 7 prediction layer feature maps at different levels of the output. Step 2.2.3: The 7 prediction layer feature maps are respectively branched from the 7 output ends of the feature extraction network into the ADR module in the corresponding prediction branch, and the three-way feature fusion logic of spatial saliency branch, local texture refinement branch and residual branch is executed; The feature map of the prediction layer is first input to the spatial saliency branch, where it is aggregated and sampled by the response extrema and response mean through the channel pooling concatenation unit. After cascaded standard convolution, batch normalization, and activation processing, saliency weights representing the importance of spatial regions are generated. The calculation formula is as follows: in, Input the prediction layer feature map of the ADR module into the current prediction branch. and These are the max pooling and average pooling operations on the channel dimension, respectively. and All are standard convolution operations, with ReLU being the ReLU activation function; Simultaneously, the predicted layer feature map enters the local texture refinement branch, where a deep convolutional layer extracts local high-frequency signals to form a texture feature stream representing the target edge. The residual branch performs an identity mapping to preserve the feature map of the prediction layer. semantic information; Significance weight Spatial gain weights are obtained by performing offset processing on the weight gain unit in the end-fusion component. Spatial gain weights of spatial saliency branches and texture feature flow of local texture refinement branches Multiplying within the element-wise multiplication unit, and then using adaptive coefficients via the weight compensation unit. After weighted compensation, the prediction layer feature map is preserved along with the residual branches. Fusion is performed within the element-wise addition unit, and the corrected features are output after processing by the ReLU6 activation layer. The calculation formula is as follows: Step 2.2.4: By establishing a compacted prior box sequence as an absolute spatial reference, the location regression parameters and confidence scores are extracted simultaneously using the detection head to predict the target location and category; At the prediction layers corresponding to the 7 prediction branches, a set of compacted prior box sequences are configured respectively; based on the target size distribution in the infrared UAV dataset in step 1, the basic size and number of prior boxes corresponding to each prediction layer are predetermined so that the prior boxes are spatially adapted to the geometric features of the targets in the infrared UAV dataset; the geometric center of each prior box is aligned with the pixel receptive field center of the corresponding prediction layer feature map to provide an absolute spatial reference benchmark for the results output by the detection head; Corrected features of the ADR module output in the prediction branch The data is input to the detection head, and the location regression convolutional layer group predicts the relative offset of the center point and the logarithmic scaling ratio of the width and height for each prior box; at the same time, the confidence prediction convolutional layer group predicts the confidence score of the probability that each prior box contains a UAV target. Finally, the relative offset of the center point output by the detection head and the logarithmic scaling ratio of width and height are applied to the absolute space reference through coordinate decoding logic to achieve the inverse transformation from relative regression parameters to absolute space coordinates of the image, thus constructing the original prediction candidate set. Furthermore, non-maximum suppression (NMS) is used to remove redundancy from the original prediction candidate set, retaining only the prediction box with the highest confidence score, and outputting the final infrared UAV target detection result.

5. The UAV early warning system based on SCAD-SSD according to claim 4, characterized in that, The specific implementation process of step 3 is as follows: Step 3.1: During the training startup phase, perform random initialization of the hyperparameters of the SCAD-SSD model; the hyperparameters include the initial learning rate, batch size, optimizer type, momentum factor, and weight decay coefficient. Step 3.2: Input the images in the training set into the SCAD-SSD model in batches and output the infrared UAV target detection results. Compare the infrared UAV target detection results with the label values ​​determined by the structured annotation file and prior box matching. Calculate the loss value through the multi-task loss function and update the network weights using backpropagation. Step 3.2.1: Input the images in the training set into the SCAD-SSD model in batches. After preprocessing by spatial scale reconstruction and mean-standard deviation correction, the images are processed by the SCB module in the feature extraction network for multi-scale feature extraction and target background environment perception. The extracted prediction layer feature maps are then input into 7 prediction branches. The ADR module in the prediction branches is used to mine salient features to suppress infrared background false alarm noise. Finally, the detection head outputs preliminary detection results containing probability confidence scores and location regression parameters. Step 3.2.2: Using the absolute spatial reference benchmark generated in Step 2.2.4, the location information of the real target boxes recorded in the structured annotation file in Step 1.2 is mapped to the corresponding feature map space; by calculating the intersection-union ratio of the prior boxes and the real target boxes, the prior boxes that meet the preset threshold conditions are established as positive sample prior boxes and real offset labels are generated; a hard negative sample mining strategy is used to filter the background region to obtain negative sample prior boxes and generate corresponding category labels; Step 3.2.3: Compare the position regression parameters and probability confidence scores output by the detector head in Step 3.2.1 with the true offset labels and category labels determined in Step 3.2.2; calculate the total loss value using the multi-task loss function. The weights of the SCAD-SSD model are updated layer by layer through error backpropagation; the formula for calculating the total loss is as follows: Where N is the number of positive prior boxes that meet the preset threshold condition. For confidence loss, To locate the loss; This is the balance coefficient; Step 3.3: Monitor the changing trend of the validation set loss value and determine whether it has reached the global optimal convergence state; if it is satisfied, then stop training and obtain the optimal detection model with UAV target recognition capability; otherwise, repeat the training logic of step 3.2 by adjusting the hyperparameters of step 3.

1.

6. A hardware acceleration method for an unmanned aerial vehicle (UAV) early warning system based on SCAD-SSD, characterized in that, The steps are as follows: Step 1: Deploy a hardware accelerator for the UAV early warning system based on SCAD-SSD on the programmable logic (PL) end of the field-programmable gate array (FPGA), and coordinate with the processing system (PS) end to execute the forward inference of the optimal detection model obtained in Step 3. The hardware accelerator consists of a multi-dimensional parallel memory access unit, a computing cluster composed of two general-purpose parameterized reconfigurable engines UPRE 0 and UPRE 1, and a global controller. During the forward inference process, the global controller implements task scheduling, controls the multi-dimensional parallel memory access unit to perform data transfer, and coordinates the computing cluster to switch between different modes to complete the inference of the target recognition network. Step 1.1: Construct a hardware accelerator for the UAV early warning system based on SCAD-SSD; among them, the multi-dimensional parallel memory access unit acts as the memory access hub, responsible for handling the data interaction between the system-side memory PS_DDR and the hardware accelerator; the general-purpose parameterized reconfigurable engine in the computing cluster acts as the core computing power unit, responsible for realizing the logical mapping of the initial sampling layer, SCB module, end-level dimensionality increase layer, ADR module and detection head in the SCAD-SSD model through the reconfiguration of hardware resources; the global controller acts as the command center, responsible for parsing control commands and coordinating the collaborative work of memory access and computing resources; Step 1.1.1: The multidimensional parallel memory access unit integrates a DMA controller, a bit-width converter, a row cache module, and a double-buffered architecture. The DMA controller handles data transfer between the system-side memory PS_DDR and the cache array on the programmable logic PL (Programmable Gate Array) of the FPGA. The bit-width converter enables bidirectional conversion between 64-bit and 8-bit data. The row cache module temporarily stores the pixel data of the serially input correction image and reconstructs it into a feature map data sliding window that meets the local receptive field requirements of the convolutional kernel. The double-buffered architecture includes Buffer 0 and Buffer 1, employing a ping-pong operation mechanism to asynchronously prefetch the convolutional kernel weight parameters, thereby achieving parallel pipelined memory access and computation. Step 1.1.2: The general-purpose parametric reconfigurable engine integrates a processing array, an interconnect control module, a nonlinear activation unit, a residual addition unit, and a weighting module. The processing array integrates a high-performance fixed-point multiplier and an addition tree logic, responsible for performing multiplication and addition operations on various convolutional layers in the optimal detection model. The interconnect control module dynamically adjusts the data flow according to the configuration signal, enabling the general-purpose parametric reconfigurable engine to adapt to the computational requirements of the SCB and ADR modules in the optimal detection model. The nonlinear activation unit performs activation mapping through an internally hardened lookup table (LUT) circuit. The residual addition unit is responsible for pixel fusion and overlay. The weighting module is responsible for requantizing and truncating the high-bit-width intermediate results produced by various convolutional layers, ensuring that the feature data or prediction results output to the system-side memory PS_DDR remain at an 8-bit fixed-point width. Step 1.1.3: The global controller drives the two general parameterized reconfigurable engines to switch between parallel working mode and collaborative working mode according to the network topology logic; at the same time, the global controller is responsible for configuring the memory access address and transmission length of the DMA controller, and coordinating the data synchronization status of the double buffer and the line buffer to realize on-chip pipelined inference of the optimal detection model. Step 1.2: After parsing the instructions from the processing system PS, the global controller coordinates the multi-dimensional parallel memory access units and the computing cluster to perform forward inference. During the forward inference process, the global controller drives the computing cluster to switch between parallel working mode and collaborative mode according to the data dependency and distribution characteristics of the network topology logic, and completes the computation tasks of the feature extraction backbone, extension layer and prediction branch. Finally, the prediction results produced by the computing cluster are written back to the system memory PS_DDR, and the processing system PS completes the post-processing of the results and coordinate output. Step 1.2.1: The global controller parses the inference start instruction from the processing system PS and begins to accelerate inference. First, the multidimensional parallel memory access unit asynchronously prefetches network weights and reads the pixels of the image to be detected. The computation cluster is configured to work in parallel mode, with UPRE 0 and UPRE 1 processing the initial sampling layer, SCB 1-6, SCB 8-13, SCB 15-17, and the final up-dimensional layer in the feature extraction network in sequence. At the same time, the multidimensional parallel memory access unit continuously performs asynchronous data prefetching on the weights according to the network level index to compensate for the memory access latency of the system-side memory PS_DDR. Step 1.2.2: When the pipeline advances to the prediction branch of the feature extraction backbone, UPRE 0 and UPRE 1 jointly complete the calculation of SCB 6, SCB13 or the final dimension-upgrading layer, producing the corresponding shallow prediction layer feature map or mid-level prediction layer feature map, resulting in two data flows: one flows into the corresponding next-level SCB 7, SCB 14 or E-SCB 1 in the extended layer, and the other flows out to the ADR module in the corresponding prediction branch; based on this data splitting characteristic, the global controller drives the computation cluster to switch to a collaborative mode; UPRE 0 is responsible for continuing to execute the calculation of the SCB module in the feature extraction network, and UPRE 1 is responsible for executing the ADR module and detection head operation of the corresponding prediction branch; Depending on the network depth, the computation cluster collaboration mode exhibits two waiting states: First, for the shallow prediction layer feature maps produced by SCB 6 and SCB13, UPRE 1 enters the idle state after completing the prediction branch operation to wait for UPRE0 to complete the subsequent SCB 7 and SCB 14 calculations in the feature extraction backbone; Second, for the mid-level prediction feature maps produced by the terminal up-dimensional layer, the computational amount required for E-SCB 1, which is handled by UPRE 0, is less than the computational amount required for the prediction branch handled by UPRE 1, so UPRE 0 enters the idle state in advance to wait for UPRE 1. Step 1.2.3: When the inference pipeline delves into the extended layer stage, the global controller drives the computation clusters to switch back completely and maintain parallel operation, sequentially executing computation tasks. UPRE 0 and UPRE 1 work together to sequentially complete the ADR module and detection head operation corresponding to E-SCB 1, E-SCB 2 operation, E-SCB 2's ADR module and detection head operation, E-SCB 3 operation, E-SCB 3's ADR module and detection head operation, E-SCB 4 operation, and E-SCB 4's ADR module and detection head operation. The global controller, through high-frequency task switching and dual-engine joint operation, quickly digests the dense prediction computation load of the extended layer, ensuring that the feature maps of the seven prediction layers are generated sequentially. Step 1.2.4: The prediction results output by each detection head are generated by the computing cluster and written back to the system-side memory PS_DDR; the processing system PS end combines the preset prior box to complete the coordinate decoding and non-maximum suppression (NMS) processing of the prediction results, and finally outputs the position coordinates and category confidence of the infrared weak target.

Images