An adaptive city person finding method and system based on edge computing

Abstract

This invention discloses an adaptive urban person-finding method and system based on edge computing, relating to the field of urban person-finding technology. The method includes the following steps: acquiring video of a monitoring area; preprocessing the video to obtain keyframes of the monitoring area; allocating the keyframes using fuzzy logic control and performing facial recognition on the allocated keyframes to obtain pedestrian images; inputting the pedestrian images into an improved FastReID model to output a feature vector of a pedestrian to be matched; acquiring the feature vector of a target pedestrian; calculating the similarity between the target pedestrian feature vector and the feature vector of the pedestrian to be matched to obtain the target pedestrian similarity; comparing the target pedestrian similarity with a preset threshold to filter and obtain a list of pedestrian matching results, thereby achieving urban person-finding.

Description

Technical Field

[0001] This invention relates to the field of urban missing persons technology, specifically to an adaptive urban missing persons method and system based on edge computing. Background Technology

[0002] Urban public safety emergency missing person searches are a key application scenario for smart city governance. With the comprehensive coverage of city-wide video surveillance networks, massive amounts of surveillance video data have become the core support for missing person searches. Urban surveillance scenarios are characterized by complex environments, high pedestrian mobility, and varied shooting angles. Missing person search tasks place stringent demands on the accuracy of pedestrian feature extraction and the robustness of feature matching. Efficient and reliable pedestrian recognition technology has become the core support for ensuring the efficiency and accuracy of urban missing person searches.

[0003] Traditional pedestrian identification and matching methods in urban missing person search scenarios employ traditional CNN pedestrian re-identification models based on global feature extraction to extract pedestrian feature vectors. These models use ResNet50 as the backbone feature extraction network and directly generate a single global feature vector by performing global spatial feature extraction and global average pooling aggregation on pedestrian images. Then, pedestrian matching is determined by calculating the cosine similarity of the global feature vector.

[0004] However, traditional models rely on a single global feature to represent pedestrian features, which cannot accurately focus on and enhance the features of key local semantic parts such as the head, torso, and limbs of pedestrians. As a result, they are prone to losing key local details of pedestrians, ultimately leading to insufficient discriminative power of pedestrian features and a significant reduction in the accuracy of pedestrian re-identification and matching. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an adaptive urban person-finding method and system based on edge computing to solve the problems existing in the background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an adaptive urban person-finding method and system based on edge computing, comprising the following steps: Step S1: Acquire video of the monitoring area, preprocess the video of the monitoring area to obtain key frames of the monitoring area; Step S2: Assign keyframes to the monitoring area using fuzzy logic control, and perform human face recognition on the assigned keyframes to obtain pedestrian images; Step S3: Input the pedestrian image into the improved FastReID model and output the feature vector of the pedestrian to be matched; Step S4: Obtain the feature vector of the target pedestrian, calculate the similarity between the feature vector of the target pedestrian and the feature vector of the pedestrian to be matched, and obtain the similarity of the target pedestrian; compare the similarity of the target pedestrian with a preset threshold, and filter to obtain a list of pedestrian matching results, thereby realizing the search for people in the city.

[0007] Preferably, the step of acquiring the monitoring area video and preprocessing the monitoring area video to obtain the monitoring area keyframes includes the following specific steps: Video of the monitored area is acquired using camera capture devices within the monitored area; A Gaussian mixture model is established based on the video of the monitored area, and background filtering is performed based on the Gaussian mixture model to finally obtain the key frames of the monitored area.

[0008] Preferably, the step of establishing a Gaussian mixture model based on the video of the monitored area includes the following steps: The video of the monitoring area is converted to grayscale and filtered by Gaussian to obtain processed continuous video frames; a Gaussian mixture model is established for each pixel position in the processed continuous video frames.

[0009] Preferably, the allocation of keyframes in the monitoring area using fuzzy logic control includes the following steps: The state variables collected at time t include: the network bandwidth B(t) between the edge server and the cloud server at time t; the average latency Te(t) for the edge server to process a single frame at time t; and the number of keyframes Qe(t) waiting to be processed in the task queue of the edge server at time t. A trapezoidal membership function is established for each state variable. Based on the trapezoidal membership function and the preset fuzzy rules, key frames of the monitoring area are assigned to obtain the assigned key frames of the monitoring area.

[0010] Preferably, obtaining the pedestrian image includes the following steps: Human image recognition is performed on the keyframes of the assigned monitoring area to obtain pedestrian target bounding boxes. The keyframes of the assigned monitoring area are then cropped based on the pedestrian target bounding boxes to obtain pedestrian images.

[0011] Preferably, the step of inputting the pedestrian image into the improved FastReID model and outputting the pedestrian feature vector to be matched includes the following steps: Build an improved FastReID model; Iterate through all pedestrian images, inputting each image into the improved FastReID model, and output the feature vector of the pedestrian to be matched. .

[0012] Preferably, the construction of the improved FastReID model includes the following specific steps: The improved FastReID model includes: a backbone feature extraction network, a feature fusion layer, a dual-branch extraction layer, a feature aggregation layer, and a restoration layer.

[0013] Preferably, the step of calculating the similarity between the target pedestrian feature vector and the pedestrian feature vector to be matched, to obtain the target pedestrian similarity, includes the following specific steps: Obtain the target pedestrian feature vector, and perform L2 normalization processing on the target pedestrian feature vector and the pedestrian feature vector to be matched; The cosine similarity between the feature vector of the target pedestrian and the feature vector of the pedestrian to be matched is calculated to obtain the similarity of the target pedestrian.

[0014] Preferably, the calculation of the cosine similarity between the target pedestrian feature vector and the pedestrian feature vector to be matched specifically involves: ; in, This represents the similarity between the nth pedestrian feature vector to be matched and the target pedestrian feature vector. This represents the feature vector of the nth pedestrian to be matched. The feature vector of the target pedestrian.

[0015] An adaptive urban missing person search system based on edge computing includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above method.

[0016] This invention provides an adaptive urban missing person search method based on edge computing, involving machine learning and deep learning technologies, which has the following beneficial effects: (1) The FastReID model, as a mature person re-identification (ReID) architecture, provides a reliable feature extraction foundation for this method. In its unmodified form, its core advantage lies in its ability to extract discriminative global feature vectors from pedestrian images, providing standardized feature inputs for subsequent similarity comparisons. This forms the basis of the person re-identification task, making it possible to search for target pedestrians by calculating the similarity between feature vectors.

[0017] (2) This improvement introduces a Local Attention Guided Network (LAG-Net) structure into the backbone network of the traditional FastReID model and integrates a center loss into the training loss function, which brings significant benefits. Specifically, the improved model can automatically generate region interest maps for key semantic parts such as the head, neck, and torso through LAG-Net, achieving dynamic focusing on effective local features and suppressing redundant background information, thus effectively solving the problem of incomplete feature extraction caused by occlusion. At the same time, the introduced center loss function can constrain the feature vectors of similar pedestrians to be more compact in the feature space, while the features of different pedestrians are more discrete, greatly enhancing the representational ability and discriminative power of the features. These improvements enable the model to better adapt to the complex environment of urban surveillance and significantly improve the accuracy and anti-interference ability of pedestrian feature extraction.

[0018] (3) The core of adaptive computing scheduling is to dynamically allocate key frames of the monitoring area to edge servers or cloud servers for processing using fuzzy logic control. This method takes three key state variables—real-time network bandwidth, edge server processing latency, and task queue length—as inputs and dynamically calculates the optimal task allocation ratio through a fuzzy rule base and inference mechanism. The advantage of this design is that it can intelligently adjust the distribution of computing tasks according to real-time changes in network conditions and computing load. For example, when network bandwidth is sufficient and edge servers are idle, more tasks can be allocated to the cloud to utilize its powerful computing power; while when the network is congested or the edge server load is high, the proportion of local processing is increased to avoid transmission delays. This approach achieves elastic collaboration and load balancing between edge and cloud computing resources, avoids processing bottlenecks caused by single-point overload, and significantly improves the system's processing efficiency and real-time task response in dynamic environments. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the steps of an adaptive urban person-finding method based on edge computing proposed in this invention. Figure 2 This is a framework diagram of the improved FastReID model in the adaptive urban missing person search method based on edge computing proposed in this invention; Figure 3This is a hierarchical diagram of the steps involved in obtaining the feature vector of the pedestrian to be matched in an adaptive urban person-finding method based on edge computing proposed in this invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figures 1-3 This invention provides a technical solution: an adaptive urban person-finding method based on edge computing.

[0023] Step S1: Acquire video of the monitoring area, preprocess the video of the monitoring area to obtain key frames of the monitoring area.

[0024] The real-time video stream output by the camera equipment within the monitoring area is continuously acquired and decoded at 25fps to obtain the video data of the monitoring area. For each frame of the monitoring area image in the video data, it is first converted into a grayscale image to reduce the complexity of subsequent calculations, and then Gaussian filtering is performed using a 3×3 Gaussian kernel (standard deviation of 1.0) to suppress high-frequency noise, finally obtaining the processed continuous video frames.

[0025] Subsequently, a Gaussian mixture model was used for background modeling, and the processed continuous video frames were analyzed: a model was established for each pixel position in the video. indivual( The preferred value is 3 to 5. Values ​​that are too small, such as 2, cannot characterize multimodal backgrounds such as changes in illumination and background disturbances; Values ​​that are too large, such as 7 or above, will significantly increase computational and storage overhead and are prone to overfitting to noise. Values ​​between 3 and 5 can better adapt to the dynamic complexity of most monitoring scenarios. A mixture model composed of a Gaussian distribution is used to describe the background grayscale value and its fluctuation at the pixel location; for each pixel in the new input frame, the pixel's The Gaussian distributions are sorted by weight from largest to smallest, and the top few high-weight Gaussian distributions (with a cumulative weight sum of ≥90%) are selected as background candidate distributions.

[0026] It should be noted that the selection of the top few high-weight Gaussian distributions (with a cumulative weight sum ≥ 90%) as background candidate distributions refers to the frequency of grayscale values ​​described by the corresponding Gaussian distribution in the observation. The higher the weight, the more stable and common the background state of that pixel (such as a static road surface or building). By sorting and selecting high-weight distributions, the most likely stable patterns representing the background can be adaptively filtered out, while automatically ignoring those distributions with low weights that may be caused by transient foreground elements (such as birds or drifting leaves) or noise. Setting the cumulative weight threshold to 90% (instead of 100%) is a key balance between model representation ability, computational efficiency, and noise resistance: it can cover the vast majority (90%) of stable background situations, ensuring the integrity of the background model; at the same time, it deliberately excludes the last 10% of distributions with extremely low weights, which usually correspond to accidental and unrepresentative interference. Excluding them can prevent the model from overfitting to instantaneous noise or small perturbations, thereby significantly improving the robustness of foreground detection and reducing the computational load during subsequent frame matching.

[0027] The grayscale value of the pixel is matched with the first few high-weight Gaussian distributions. For each high-weight Gaussian distribution: if the grayscale value of the current pixel satisfies |current pixel grayscale value - high-weight Gaussian mean| ≤ ​​2.5 × high-weight Gaussian standard deviation (matching the candidate Gaussian), then the pixel is determined to be the background, and the mean, variance and weight of the high-weight Gaussian distribution are updated simultaneously. If all high-weight Gaussian distributions do not match, then the pixel is determined to be the foreground.

[0028] The Gaussian mixture model dynamically updates the parameters (mean and variance) of the corresponding Gaussian distribution based on the matching results and adjusts the model weights, enabling the model to progressively adapt to changes in lighting and the introduction of new persistent stationary objects (such as parked vehicles) in the scene. This process ultimately outputs a binarized foreground mask. The number of pixels identified as foreground in this mask is counted, and the proportion of foreground pixels to the total pixels in the frame is calculated to obtain the foreground pixel ratio of the frame.

[0029] A foreground pixel ratio threshold is set (preferably 0.5%, which can be adjusted according to the noise level and target detection sensitivity requirements of the actual monitoring scene). When the foreground pixel ratio exceeds the threshold, the frame is determined to be a key frame of the monitoring area containing a significant moving target (pedestrian). For images determined to be key frames, image size normalization is further performed, uniformly scaling them to a fixed resolution of 256 pixels high and 128 pixels wide to meet the input requirements of the subsequent pedestrian recognition model. After the above preprocessing, a series of monitoring area key frames that have removed static background redundancy and contain effective dynamic targets are finally output.

[0030] Step S2: Allocate the key frames of the monitoring area through the fuzzy logic control method, and perform human face recognition on the allocated key frames of the monitoring area to obtain pedestrian images.

[0031] Connect the sequence of key frames of the monitoring area output in Step S1. The current processing time is t, and the total number of key frames to be processed accumulated at this time is (t).

[0032] In Step S2, first calculate the dynamic task allocation ratio through a fuzzy logic controller. This controller takes three state variables collected in real time as inputs, including the network bandwidth B(t) between the edge server and the cloud server at time t, in Mbps; the average delay Te(t) of the edge server to process a single frame at time t, in seconds; and the number of key frames Qe(t) waiting to be processed in the task queue of the edge server at time t.

[0033] Define three fuzzy linguistic values of low, medium, and high for each input variable, and use the trapezoidal membership function to complete the fuzzification of the exact value. For the bandwidth B(t), the membership function of the bandwidth B(t) is a continuous triangular function, and its membership function is configured as follows. For the fuzzy set "low", its membership function is defined as: when B(t) ≤ , = 1; when <B(t)< , = ; when B(t) ≥ , = 0. For the fuzzy set "medium", its membership function is defined as: when B(t) ≤ or B(t) > , = 0; when <B(t)< , = ; when ≤B(t)≤ , = 1; when <B(t)< , = . For the fuzzy set "high", its membership function is defined as: when B(t) ≤ , = 0; when <B(t)< , = ; when B(t) ≥ hour, =1. Where... , , , To preset threshold parameters based on the actual network environment, The preferred bandwidth is 2Mbps. When B(t)≤2Mbps, the network bandwidth is clearly considered "low". At this time, the transmission cost is high, and data transmission to the cloud should be avoided as much as possible. The preferred speed is 5Mbps, this value is the same as They jointly defined the transition range between "low" and "medium" ( to (2 to 5 Mbps) The preferred speed is 10 Mbps, this value is the same as They jointly defined the stable range of the "medium" bandwidth. to (5 to 10 Mbps) The preferred bandwidth is 20Mbps. When B(t)≥20Mbps, the network bandwidth is clearly defined as "high", which can make full use of cloud computing resources.

[0034] The fuzzification processing of the average delay Te(t) and the number of keyframes Qe(t) adopts the same architecture and process as the bandwidth B(t), that is, it is performed by using four preset ordered threshold parameters (for Te(t) as...). , , , For Qe(t) as , , , Each of the three trapezoidal membership functions, "low," "medium," and "high," is defined on its respective numerical domain. Specifically, when the input value is less than or equal to the first threshold, it belongs entirely to "low." When the input value is between the first and second thresholds, the membership of "low" decreases linearly to 0, while the membership of "medium" increases linearly to 1. When the input value is between the second and third thresholds, it belongs entirely to "medium." When the input value is between the third and fourth thresholds, the membership of "medium" decreases linearly to 0, while the membership of "high" increases linearly to 1. When the input value is greater than or equal to the fourth threshold, it belongs entirely to "high." Therefore, the fuzzy quantization descriptions of the average delay Te(t) and queue length Qe(t) are processed using the same mathematical method as that used for the bandwidth B(t).

[0035] The , , , , The optimal latency is 0.05 seconds. A value lower than this indicates extremely low latency and system idleness. The optimal value is 0.1 seconds, which is the boundary between low and medium delay, and also the ideal center value for medium-delay fuzzy sets. The preferred value is 0.2 seconds, which is the dividing point between medium and high latency. The optimal latency is 0.3 seconds. Higher values ​​result in high latency and may overload the server.

[0036] The , , , , Ideally, it uses 2 frames, resulting in a short queue and ample edge resources. The optimal value is 5 frames, which is the boundary between the short and medium queues and also the ideal center value of the fuzzy set of the medium queue. The preferred frame size is 15 frames, the dividing point between the medium and long queues. The optimal frame rate is 25 frames per second. If the frame rate is higher than this, the queue will be severely congested and the waiting time may exceed the limit.

[0037] The controller's output variable is the proportion of tasks assigned to the edge server, P(t), with a value range of [0,1]. The corresponding fuzzy linguistic values ​​are defined as few, medium, and many (few corresponds to P(t)∈[0,0.3], medium to P(t)∈[0.3,0.7], and many to P(t)∈[0.7,1]). The controller has a pre-defined rule base containing R fuzzy rules, each in the form of "if [precondition] then [conclusion]". A typical rule is: if B(t) is high, Te(t) is low, and Qe(t) is low, then P(t) is many. Using the Mamdani inference method, for a given set of exact inputs [B(t), Te(t), Qe(t)], the membership degree to each rule's precondition and the activation strength of each rule are calculated. Take the minimum membership degree of all the prerequisites of the rule (the AND operation takes the smaller operator).

[0038] Complete all rules to activate strength After calculation, the activation intensity is selected. Rules with a value greater than 0 are used as activated rules. Based on the corresponding output fuzzy linguistic values ​​(few / medium / many) of the activated rules, the conclusions of all activated rules are aggregated into a fuzzy set about the output variable P(t) through an "OR" operation (taking the larger value). The aggregated fuzzy set is then defuzzified using the centroid method to obtain the precise allocation ratio P(t), which is calculated using the following formula:

[0039] The integration domain is the range of values ​​for the output variable P, [0, 1]. The membership function of the aggregated fuzzy set is used to calculate the precise proportion P(t) of keyframes assigned to the edge server for processing at time t.

[0040] According to the allocation ratio Perform keyframe task allocation: The number of keyframes allocated to the edge server is Ne(t), Ne(t)=ceil(P(t)*N(t)), where ceil() is the floor function.

[0041] The remaining number of keyframes allocated to the cloud server is Nc(t), where Nc(t) = N(t) - Ne(t), and Nc(t) ≥ 0. If the calculation result is negative, then Nc(t) = 0. Following the keyframe time sequence output in step S1, the first Ne(t) frames are transmitted to the pedestrian detection module on the edge server, and subsequent Nc(t) frames are uploaded to the pedestrian detection module on the cloud server.

[0042] In one embodiment, within the edge server and cloud server, keyframes are allocated for each frame. (Where k is the frame number, k=1,2,...,Ne(t) or k=1,2,...,Nc(t)), both using the YOLOv5s model (YOLOv5s balances speed and accuracy, adaptable to heterogeneous deployments in edge / cloud environments) to perform target pedestrian recognition. This model employs a multi-scale grid partitioning strategy, dividing the input image frame into three groups of S×S grids: 80×80, 40×40, and 20×20 (80×80 is preferred for a single scale). Each grid cell is responsible for predicting B bounding boxes, with B preferably being 3. These three bounding boxes can respectively match pedestrian targets of small, medium, and large scales, such as small pedestrians at a distance, normal adults, and large pedestrians at close range, covering the core differences in pedestrian size in urban surveillance scenarios.

[0043] Each bounding box outputs three sets of predicted values, namely the offset of the bounding box center coordinates relative to the top left corner of the grid cell. , ), scaling of bounding box width and height relative to prior anchor box size ( , ), Confidence score of pedestrian target contained in the box The confidence score is calculated as follows:

[0044] in, This represents the probability that a pedestrian is present within the bounding box, with a value ranging from [0,1]. The intersection-union ratio (IU) of the predicted bounding box and the real pedestrian target box ranges from [0,1]. This score is used to comprehensively characterize the probability of a pedestrian target being present in the box, as well as the positional matching accuracy between the predicted bounding box and the real target box.

[0045] Target pedestrian bounding boxes are filtered using a non-maximum suppression algorithm: First, filtering boxes with confidence scores below the detection threshold are processed. (Preferred value is 0.5, which can be adjusted according to the scenario) bounding boxes, and then apply an IOU threshold to the remaining bounding boxes. =0.45 performs non-maximum suppression (deduplication of overlapping boxes), ultimately obtaining the set of pedestrian target bounding boxes for the current frame { |j=1,2,..., },in =( , , , The coordinates (x, y) of the center of the j-th bounding box, as well as its width and height, are respectively... This represents the total number of pedestrians identified in the current frame. Based on each bounding box... The coordinates are derived from the keyframe after size normalization in step S1. By cropping the corresponding image region, we obtain the pedestrian image set of the current frame. }; Traverse all allocated keyframes on the edge server and cloud server and complete the above recognition and cropping operations, finally summarizing to obtain the pedestrian image list output in step S2.

[0046] The pedestrian image list contains N pedestrian images, and the number of pedestrian images, N, is equal to the sum of the number of pedestrians detected by all processed keyframes on the edge server and the cloud server. Specifically, for each processed keyframe... The YOLOv5s model will detect A set of pedestrian images for that frame is obtained by cropping the images of each pedestrian. After traversing and processing all assigned keyframes, the pedestrian image sets of all frames are summarized to obtain a complete list of pedestrian images.

[0047] Step S3: Input the pedestrian image into the improved FastReID model and output the pedestrian feature vector to be matched.

[0048] Step S3 inputs the pedestrian image output from step S2 into the improved FastReID model, and outputs the pedestrian feature vector. The improved FastReID model uses ResNet50 as the backbone feature extraction network, introduces a Local Attention Guided Network (LAG-Net) structure, and integrates center loss in the training loss function to enhance feature discrimination.

[0049] The improved FastReID model includes: a backbone feature extraction network (initial convolutional layer and residual network), a feature fusion layer, a dual-branch extraction layer (local branch and global branch), a feature aggregation layer, and a restoration layer.

[0050] Step S2 outputs a total of N pedestrian images, forming the input image set. ,in Let represent the nth pedestrian image. For each input image... First, the image is scaled to a fixed height (H) and width (W) in pixels (H=256, W=128), and then pixel value normalization is performed to obtain the preprocessed image tensor. .

[0051] The preprocessed image tensor The image is then fed into an improved backbone feature extraction network for preliminary feature extraction. The input layer of the backbone feature extraction network receives the preprocessed image tensor. Its fixed dimensions are height H=256 pixels, width W=128 pixels, and 3 color channels (RGB). The batch size during training is preferably 32.

[0052] The image tensor (The input shape is (batch size, 3, 256, 128), and the format is batch size, channel, height, width.) First, it passes through the initial convolutional layer of the backbone feature extraction network. This initial convolutional layer uses 64 convolutional kernels with a size of 7×7, a stride of 2, and padding of 3 to perform two-dimensional convolution operations. Then, it is followed by batch normalization and ReLU activation function, and then passed through a max pooling layer with a size of 3×3 and a stride of 2 for downsampling. The final output is a feature map with a shape of (batch size, 64, 64, 32).

[0053] Subsequently, the feature map with shape (batch size, 64, 64, 32) is sequentially passed through a residual network consisting of four stages (stage1 to stage4) for deep feature extraction.

[0054] Stage 1 consists of three bottleneck residual blocks. Each residual block first passes through a 1×1 convolutional layer (64 kernels, stride 1) to reduce the number of input channels from 64 to 64, followed by batch normalization and ReLU activation. Then, it passes through a 3×3 convolutional layer (64 kernels, stride 1, padding 1) for spatial feature extraction, again followed by batch normalization and ReLU. Finally, it passes through a 1×1 convolutional layer (256 kernels, stride 1) to increase the number of channels to 256, followed by batch normalization. The main path output is added to the shortcut connection output of the identity mapping, and then processed by ReLU activation to complete the single-block computation. All convolutional strides in this stage are 1, with no spatial downsampling. The final output is a feature map with the shape (batch size, 256, 64, 32).

[0055] Stage 2 consists of four bottleneck residual blocks. The first residual block first reduces the number of input channels from 256 to 128 using a 1×1 convolutional layer (128 kernels, stride 1), followed by batch normalization and ReLU activation. Then, it passes through a 3×3 convolutional layer (128 kernels, stride 2, padding 1) for spatial feature extraction and downsampling, followed by batch normalization and ReLU. Finally, it passes through a 1×1 convolutional layer (512 kernels, stride 1) to increase the number of channels from 128 to 512, followed by batch normalization. Because the output space size (32, 16) and number of channels (512) of the main path do not match the input (64, 32) and number of channels (256), the shortcut connection simultaneously performs downsampling and channel boosting through a 1×1 convolutional layer (512 convolutional kernels, stride 2). After adding the outputs of the main path and the shortcut connection, ReLU is applied, and the output shape of this block is (batch size, 512, 32, 16). The subsequent three bottleneck residual blocks (maintaining size): the structure of each residual block is consistent with the core idea of ​​the first residual block, only the stride is adjusted, that is, 1×1 convolution (128 kernels, stride 1), 3×3 convolution (128 kernels, stride 1, padding 1), and 1×1 convolution (512 kernels, stride 1). Each layer is followed by batch normalization and ReLU (the last 1×1 convolution is followed only by batch normalization). Since the input / output space size (32, 16) and the number of channels (512) are exactly the same, the shortcut connections of these three blocks are all identity mappings. After the above processing, the final output feature map shape in this stage is (batch size, 512, 32, 16).

[0056] Stage 3 comprises six bottleneck residual blocks. The first block uses a 1×1 convolutional layer (256 kernels, stride 1) to reduce the input channel count from 512 to 256, followed by batch normalization and ReLU activation. Next, a 3×3 convolutional layer (256 kernels, stride 2, padding 1) extracts spatial features and downsamples the input, again followed by batch normalization and ReLU. Finally, a 1×1 convolutional layer (1024 kernels, stride 1) increases the channel count from 256 to 1024, followed by batch normalization. Since the output spatial size (after stride 2 convolution) and channel count (1024) of the main path in this block do not match the input (512, 32, 16), its shortcut connection uses an additional 1×1 convolutional layer (1024 kernels, stride 2) to simultaneously downsample the input feature map and increase the channel count to match the output of the main path. After adding the outputs of the main path and the shortcut connections, the output of this block is obtained by applying the ReLU activation function, with a shape of (batch size, 1024, 16, 8). The subsequent five bottleneck residual blocks (maintaining size): each block has the same structure as the first core block, namely, a 1×1 convolution (256 kernels, stride 1), a 3×3 convolution (256 kernels, stride 1, padding 1), and a 1×1 convolution (1024 kernels, stride 1). Each layer is followed by batch normalization and ReLU (the last 1×1 convolution is followed only by batch normalization). Since the input-output spatial dimensions (16, 8) and the number of channels (1024) are exactly the same, the shortcut connections of these five blocks are all identity mappings. After the above processing, the final output feature map shape in this stage is (batch size, 1024, 16, 8).

[0057] Stage 4 comprises three bottleneck residual blocks. The first block begins by using a 1×1 convolutional layer (512 kernels, stride 1) to reduce the number of input channels from 1024 to 512, followed by batch normalization and ReLU activation. Next, a 3×3 convolutional layer (512 kernels, stride 2, padding 1) extracts spatial features and downsamples the input, again followed by batch normalization and ReLU. Finally, a 1×1 convolutional layer (2048 kernels, stride 1) increases the number of channels from 512 to 2048, followed by batch normalization. Since the output spatial size (after stride 2 convolution) and number of channels (2048) of the main path in this block do not match the input (1024, 16, 8), its shortcut connection uses an additional 1×1 convolutional layer (2048 kernels, stride 2) to simultaneously downsample the input feature map and increase the number of channels to match the output of the main path. After adding the outputs of the main path and the shortcut connections, the output of this block is obtained by applying the ReLU activation function, with a shape of (batch size, 2048, 8, 4). The subsequent two blocks have similar structures, namely, 1×1 convolutions (512 kernels, stride 1), 3×3 convolutions (512 kernels, stride 1, padding 1), and 1×1 convolutions (2048 kernels, stride 1), with batch normalization and ReLU applied after each layer (except for the last 1×1 convolution layer, which is only followed by batch normalization). Since the input-output spatial dimensions (8, 4) and the number of channels (2048) are exactly the same, the shortcut connections of these two blocks are identity mappings. After the above processing, the final output feature map shape of this stage is (batch size, 2048, 8, 4).

[0058] Feature Fusion Layer: The feature maps output from stage 3 and stage 4 are concatenated along the channel dimension. Specifically, the feature map from stage 4 is upsampled to a spatial size of 16×8 using bilinear interpolation, and then concatenated with the feature map from stage 3 along the channel dimension, ultimately resulting in a fused feature map with the shape (batch size, 3072, 16, 8). This fused feature map will serve as the input for the subsequent local attention-guided network.

[0059] Dual-branch extraction layer: Subsequently, the fused feature map with a shape of (batch size, 3072, 16, 8) is input into the local branch composed of a local attention-guided network module. This module first generates Q region interest maps (Q=6, corresponding to the 6 core semantic parts of the pedestrian: head, neck, torso, left arm, right arm, and legs). Specifically, the feature map is processed through a squeezing and activation module. Model the inter-channel relationships: First, Global average pooling along the spatial dimension (H′×W′=16×8) yields a C-dimensional (3072-dimensional) channel statistics vector, which is then passed through two fully connected layers (hidden layer dimension). The activation function is ReLU, which outputs a C-dimensional channel attention weight vector. This weight vector is then broadcast along the spatial dimension to H′×W′ and compared with the original feature map. Element-wise multiplication is used to reweight the channels, thus obtaining the enhanced feature map. Next, a RIM generator consisting of 1×1 convolutional layers (6 kernels, 1 stride, no padding) is applied to... Output Q region interest maps , () is the mapping function for the RIM generator, where each The spatial attention weight distribution characterizes the semantic region q in the nth image.

[0060] For the interest map of the k-th region The corresponding local features are calculated through the local attention module. This module will generate regional interest maps. Broadcast along the channel dimension to channel C, and then compared with the initial feature map. Perform element-wise multiplication in the spatial dimension, and finally combine them through residual join, i.e. ,in This represents an element-wise multiplication broadcast operation along spatial dimensions.

[0061] For each local feature Global average pooling is performed along its spatial dimensions (height and width) to obtain a C-dimensional column feature vector. , that is, local feature vectors.

[0062] All Q local feature vectors By stacking along the row dimensions (outside the channel dimensions), a local attention feature matrix is ​​constructed. The matrix is ​​processed by a batch attention dropout module, which randomly selects a feature vector position corresponding to a region of interest map with a 50% probability and sets all channels to zero (the other 5 feature vectors remain unchanged) to enhance the model's robustness to local feature loss. The processed features are then subjected to attention feature regularization, a process that introduces a center loss term. The feature vector that constrains the category of the q-th semantic part In feature space, it is close to its learnable category center. This is to enhance the discriminative power of local features.

[0063] Local attention feature matrix After processing, it is compared with the global branch (which is parallel to the local branch composed of local attention-guided network modules, and the global branch is used to fuse feature maps). The global features extracted by global average pooling along the spatial dimension are fused to obtain the fused features.

[0064] Feature Aggregation Layer: The fused features are fed into the aggregation module, where generalized average (GeM) pooling is used for feature aggregation. Specifically, for each channel c of the feature map, c=1,2,3,...,C, its spatial position is calculated. Activation value of The mean of the powers, then open The right root yields the aggregated global feature vector. The formula for calculating the c-th component is: ,in The total number of spatial locations is (H′×W′=16×8=128). The shape hyperparameter for GeM pooling is preferably 3.

[0065] Reduction layer: The final global feature vector A reduction layer, consisting of a 1x1 convolutional layer (512 kernels, stride 1), a batch normalization layer, a ReLU activation function, and a dropout layer (dropout rate = 0.5) connected sequentially, reduces the feature dimension from C to D (D = 512), yielding the final pedestrian feature vector to be matched. Iterate through all input pedestrian images. For n=1,2,...,N, repeat the above process to finally obtain the corresponding set of pedestrian feature vectors to be matched. This is the output of step S3.

[0066] It should be noted that the traditional FastReID model relies on global feature extraction in urban missing person search scenarios. When faced with issues such as frequent partial occlusion, pose changes, and perspective differences in pedestrians in surveillance footage, it is prone to losing key local details and has insufficient feature discrimination power, leading to a decrease in pedestrian re-identification accuracy. This makes it difficult to meet the actual needs of urban missing person search for accurate matching and efficient identification. Therefore, a targeted improvement is made. This improvement introduces a Local Attention-Guided Network (LAG-Net) into the FastReID backbone feature extraction network, which accurately locates key semantic parts of the human body by generating region interest maps. The improved FastReID model dynamically focuses on effective local features and suppresses redundant information interference, solving the problem of incomplete feature extraction in occluded scenarios. It also introduces a center loss into the model training loss function to constrain similar pedestrian features to be more compact in the feature space and dissimilar pedestrian features to be more discrete, further enhancing the discriminative power and robustness of the features. The improved FastReID model is better adapted to the complex conditions of urban surveillance, significantly improving the accuracy and anti-interference ability of pedestrian feature extraction, providing reliable feature support for subsequent pedestrian feature similarity comparison, and effectively ensuring the recognition accuracy and search efficiency of the edge computing adaptive urban person-finding method.

[0067] Step S4: Obtain the feature vector of the target pedestrian, calculate the similarity between the feature vector of the target pedestrian and the feature vector of the pedestrian to be matched, and obtain the similarity of the target pedestrian; compare the similarity of the target pedestrian with a preset threshold, and filter to obtain a list of pedestrian matching results, thereby realizing the search for people in the city.

[0068] The similarity between the pedestrian feature vector and the target pedestrian feature vector is calculated, and the target pedestrians (i.e., the people to be found) are selected based on the similarity threshold, thereby realizing the search for people in the city. The specific implementation steps are as follows: The target pedestrian feature vector is... ,in The dimension of the feature vector is the feature vector of the target pedestrian, which is generated from the reference image of the target pedestrian through the feature extraction process from step S1 to step S3, ensuring that it is consistent with the extraction logic of the feature vector of the pedestrian to be matched.

[0069] For each feature vector in the pedestrian feature vector to be matched output in step S3 eigenvectors Corresponding to the nth pedestrian image To ensure the accuracy of similarity calculation, first analyze all feature vectors ( and Perform L2 normalization. Then, for each eigenvector in the set... Calculate its relationship with the target feature vector Similarity between target pedestrians The calculation formula is: , This represents the similarity between the nth pedestrian feature vector to be matched and the target pedestrian feature vector. The range of values ​​is The closer the value is to 1, the more similar the appearance of the pedestrians represented by the two vectors are.

[0070] Traverse all The similarity is calculated using each feature vector, resulting in the corresponding similarity set. Set a similarity matching threshold. Its value range is It can be adjusted according to the actual application requirements for recognition accuracy and recall. The preferred value is 0.7. Each calculated similarity... With threshold If a comparison is made, Then determine the first The pedestrian image contains the target pedestrian who matches the target pedestrian. This pedestrian image and its associated information (including but not limited to: pedestrian image) The keyframe timestamps from the source, the camera ID from the source, and the calculated similarity. Add to pedestrian matching results list In the middle; if If the match is not found, it will be excluded. After traversing the entire similarity set and completing the above judgment and filtering, the final list of matched pedestrians will be output. All entries in this list represent the pedestrian sighting records that the system has located in the city's surveillance network and that match the target being searched, thus completing the task of finding people in the city.

[0071] Furthermore, based on the above method embodiments, the present invention also provides an urban missing person search system, including a memory, a processor, and computer programs stored in the memory, which are adapted to be loaded and executed by the processor to implement the above-described adaptive urban missing person search method based on edge computing.

[0072] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, the phrase "comprising an element defined as..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0073] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An adaptive urban person-finding method based on edge computing, characterized in that: Includes the following steps: Step S1: Acquire video of the monitoring area, preprocess the video of the monitoring area to obtain key frames of the monitoring area; Step S2: Assign keyframes to the monitoring area using fuzzy logic control, and perform human face recognition on the assigned keyframes to obtain pedestrian images; Step S3: Input the pedestrian image into the improved FastReID model and output the feature vector of the pedestrian to be matched; Step S4: Obtain the feature vector of the target pedestrian, calculate the similarity between the feature vector of the target pedestrian and the feature vector of the pedestrian to be matched, and obtain the similarity of the target pedestrian; compare the similarity of the target pedestrian with a preset threshold, and filter to obtain a list of pedestrian matching results, thereby realizing the search for people in the city.

2. The adaptive urban person-finding method based on edge computing according to claim 1, characterized in that: The process of acquiring video of the monitoring area and preprocessing the video to obtain keyframes of the monitoring area includes the following specific steps: Video of the monitored area is acquired using camera capture devices within the monitored area; A Gaussian mixture model is established based on the video of the monitored area, and background filtering is performed based on the Gaussian mixture model to finally obtain the key frames of the monitored area.

3. The adaptive urban person-finding method based on edge computing according to claim 2, characterized in that: The process of establishing a Gaussian mixture model based on the video of the monitored area includes the following steps: The video of the monitoring area is converted to grayscale and filtered by Gaussian to obtain processed continuous video frames; a Gaussian mixture model is established for each pixel position in the processed continuous video frames.

4. The adaptive urban person-finding method based on edge computing according to claim 3, characterized in that: The allocation of keyframes in the monitoring area using fuzzy logic control includes the following steps: The state variables collected at time t include: the network bandwidth B(t) between the edge server and the cloud server at time t; the average latency Te(t) for the edge server to process a single frame at time t; and the number of keyframes Qe(t) waiting to be processed in the task queue of the edge server at time t. A trapezoidal membership function is established for each state variable. Based on the trapezoidal membership function and the preset fuzzy rules, key frames of the monitoring area are assigned to obtain the assigned key frames of the monitoring area.

5. The adaptive urban person-finding method based on edge computing according to claim 4, characterized in that: Obtaining the pedestrian image includes the following steps: Human image recognition is performed on the keyframes of the assigned monitoring area to obtain pedestrian target bounding boxes. The keyframes of the assigned monitoring area are then cropped based on the pedestrian target bounding boxes to obtain pedestrian images.

6. The adaptive urban person-finding method based on edge computing according to claim 5, characterized in that: The process of inputting the pedestrian image into the improved FastReID model and outputting the pedestrian feature vector to be matched includes the following steps: Build an improved FastReID model; Iterate through all pedestrian images, inputting each image into the improved FastReID model, and output the feature vector of the pedestrian to be matched. .

7. The adaptive urban person-finding method based on edge computing according to claim 6, characterized in that: The construction of the improved FastReID model includes the following specific steps: The improved FastReID model includes: a backbone feature extraction network, a feature fusion layer, a dual-branch extraction layer, a feature aggregation layer, and a restoration layer.

8. The adaptive urban person-finding method based on edge computing according to claim 7, characterized in that: The calculation of the similarity between the target pedestrian feature vector and the pedestrian feature vector to be matched, to obtain the target pedestrian similarity, includes the following specific steps: Obtain the target pedestrian feature vector, and perform L2 normalization processing on the target pedestrian feature vector and the pedestrian feature vector to be matched; The cosine similarity between the feature vector of the target pedestrian and the feature vector of the pedestrian to be matched is calculated to obtain the similarity of the target pedestrian.

9. The adaptive urban person-finding method based on edge computing according to claim 8, characterized in that: The calculation of the cosine similarity between the target pedestrian feature vector and the pedestrian feature vector to be matched is specifically as follows: ； in, This represents the similarity between the nth pedestrian feature vector to be matched and the target pedestrian feature vector. This represents the feature vector of the nth pedestrian to be matched. The feature vector of the target pedestrian.

10. An adaptive urban person-finding system based on edge computing, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1-9.

Images