A wide-field-of-view, high-resolution multi-object detection method based on adaptive intelligent matching

Abstract

This invention discloses a wide-field-of-view, high-resolution multi-object detection method based on adaptive intelligent matching. The method includes: acquiring an original image containing pedestrians and constructing a target detection network model; inputting the original image into a backbone network for feature extraction to obtain deep semantic features; using a preset enhancement method to enhance the sensitivity of the deep semantic features to obtain feature enhancement results for the target boundaries; inputting the feature enhancement results into multiple detection heads to predict keypoints and regression sets based on the represented pedestrian bounding boxes to obtain keypoint prediction results and regression set prediction results; wherein, the represented pedestrian bounding boxes include boundary localization of the upper left corner keypoint of the pedestrian using the left and upper boundaries of the bounding boxes, and boundary localization of the lower right corner keypoint of the pedestrian using the right and lower boundaries. This invention can effectively improve the accuracy of automatic pedestrian detection in billion-pixel wide-field-of-view, high-resolution images.

Description

Technical Field

[0001] This invention relates to the field of multi-object detection technology, and in particular to a wide-field-of-view, high-resolution multi-object detection method based on adaptive intelligent matching. Background Technology

[0002] Most existing pedestrian detection methods are optimized variants based on center point detectors. These methods typically use the object's center point as a reference during the modeling process and then directly obtain the object's length and width through distance regression. There are also detection methods based on human body parts, which first detect the pedestrian's head, hands, legs, etc., separately, and then aggregate them into the final detection result through appropriate methods.

[0003] In billion-pixel scenes, pedestrians exhibit a wide variety of postures. The center point of a person's bounding box may fall into the background area and is prone to occlusion, indicating an unstable visual pattern for the center point. Furthermore, an accurate center point requires the joint localization of four boundaries, resulting in a high degree of freedom. Therefore, the robustness and accuracy of center point-based detection methods are limited. In large-scale scenes, the scale of pedestrians varies drastically at different distances from the camera. Detectors that directly regress the length and width of the bounding box are affected by long-distance regression errors, which can even influence the identification of positive samples by anchor-box-based detectors during training. Ultimately, the quality of the final detection box is also affected. Summary of the Invention

[0004] The present invention aims to at least partially solve one of the technical problems in the related art.

[0005] The purpose of this invention is to provide a wide-field-of-view, high-resolution multi-object detection method based on adaptive intelligent matching, which improves the accuracy of automatic pedestrian detection in billion-pixel wide-field-of-view high-resolution images and effectively handles detection challenges such as pedestrian poses in large-scale scenes, frequent occlusion, dense crowds, and different situations involving adults and children. Distance regression is decoupled from the quality of the final detection box, effectively addressing the impact of long-distance regression on the quality of the final detection box. A simple and effective adaptive corner matching strategy is proposed, which provides reliable guidance information for corner matching by designing pedestrian body shape encoding for each corner, thereby improving the reliability of corner matching.

[0006] Another objective of this invention is to propose a wide-field-of-view, high-resolution multi-object detection system based on adaptive intelligent matching.

[0007] To achieve the above objectives, this invention proposes a wide-field-of-view, high-resolution multi-object detection method based on adaptive intelligent matching, comprising:

[0008] Acquire the original image containing pedestrians and construct an object detection network model; wherein, the object detection network model includes a backbone network and multiple detection heads;

[0009] The original image is input into the backbone network for feature extraction to obtain deep semantic features;

[0010] The deep semantic features are enhanced with a preset enhancement method to obtain the feature enhancement result of the target boundary;

[0011] The feature enhancement results are input to the multiple detection heads respectively to predict key points and regression sets based on the represented pedestrian bounding boxes to obtain key point prediction results and regression set prediction results; wherein, the represented pedestrian bounding boxes include boundary positioning of the upper left corner of the pedestrian using the left and upper boundaries of the pedestrian bounding boxes, and boundary positioning of the lower right corner of the pedestrian using the right and lower boundaries.

[0012] In addition, the wide-field-of-view, high-resolution multi-object detection method based on adaptive intelligent matching according to the above embodiments of the present invention may also have the following additional technical features:

[0013] Furthermore, before inputting the original image into the backbone network, the method further includes:

[0014] The original image is downsampled to obtain a downsampled image;

[0015] The downsampled image is cropped using a sliding window of a preset size to obtain a slice image of the preset size.

[0016] Furthermore, the prediction of keypoints and regression sets based on the pedestrian bounding box representation yields keypoint prediction results and regression set prediction results, including:

[0017] The key point prediction results are obtained by using a heatmap of the location of the key point at the upper left corner of the pedestrian bounding box predicted by the first detection head and a heatmap of the location of the key point at the lower right corner of the pedestrian bounding box predicted by the third detection head.

[0018] The regression set prediction result is obtained by using the second detection head to regress the human shape code set corresponding to the upper left corner key point and the fourth detection head to regress the human shape code set corresponding to the lower right corner key point.

[0019] Furthermore, the key point at the top left corner or the key point at the bottom right corner is mapped to a circular region with Gaussian decay, and the loss function of the corresponding location heatmap is obtained:

[0020]

[0021] Where N is the total number of pedestrians in the current image, p xy It is the true value of the key point at coordinates (x, y) on the heatmap, with a value range of 0-1, p` xyα is the network output prediction value at the corresponding position, α is a hyperparameter that adjusts the weights of easy and difficult samples, and β is a hyperparameter that adjusts the weights of the distance decay term.

[0022] Furthermore, the optimization objective of the object detection network model changes from single-segment long-distance regression to regression set, and the corresponding loss function is:

[0023]

[0024] Where N is the number of pedestrians in the current image, n is the number of elements in a regression set, and w' k,i w is the network output predicted value of the i-th regression element in the regression set for the k-th objective. k,i It is the truth value of the corresponding regression element.

[0025] Furthermore, the method also includes:

[0026] Heatmap decoding: The sigmoid function is used to activate the location heatmap. After activation, max pooling is used to process the location heatmap, and keypoints with a confidence level greater than a preset threshold are obtained on each location heatmap to obtain multiple top-left and bottom-right keypoints; and,

[0027] Human body shape decoding: The regression set is mapped to the original space using an exponential function. The statistical average of the regression elements of the current set is calculated based on the elements of the regression set of the top-left key point and the corresponding scaling factor, so as to decode the length and width of the body frame corresponding to the top-left key point.

[0028] Furthermore, after heatmap decoding and human body shape decoding, the method further includes:

[0029] The intersection-union ratio (CUNR) of the body frames corresponding to the multiple top-left corner key points and multiple bottom-right corner key points is calculated to obtain the CUNR calculation result;

[0030] The cross-union ratio (CUNR) calculation results are compared, and a preset number of CUNRs are selected for matching based on the comparison results to obtain a pair of diagonal key points.

[0031] To achieve the above objectives, another aspect of the present invention proposes a wide-field-of-view, high-resolution multi-object detection system based on adaptive intelligent matching, comprising:

[0032] The data acquisition module is used to acquire original images containing pedestrians and construct a target detection network model; wherein, the target detection network model includes a backbone network and multiple detection heads;

[0033] The feature extraction module is used to input the original image into the backbone network to extract deep semantic features;

[0034] The feature enhancement module is used to enhance the sensitivity of the deep semantic features using a preset enhancement method to obtain the feature enhancement result of the target boundary;

[0035] The result prediction module is used to input the feature enhancement results to the multiple detection heads respectively, so as to obtain the key point prediction results and regression set prediction results based on the represented pedestrian bounding box; wherein, the represented pedestrian bounding box includes boundary positioning of the upper left corner of the pedestrian using the left and upper boundaries of the pedestrian bounding box, and boundary positioning of the lower right corner of the pedestrian using the right and lower boundaries.

[0036] The wide-field-of-view, high-resolution multi-object detection method and system based on adaptive intelligent matching in this invention can effectively improve the accuracy of automatic pedestrian detection in billion-pixel wide-field-of-view, high-resolution images, effectively solve the impact of long-distance regression on the quality of the final detection box, and improve the reliability of corner point matching.

[0037] The beneficial effects of this invention are as follows:

[0038] (1) This invention relies on the key corner points of the upper left and lower right with low degrees of freedom to locate and detect target pedestrians. These two independent corner points have good inclusiveness for dense and occluded pedestrian scenes, which can effectively improve the accuracy of automatic pedestrian detection in billion-pixel wide field of view high-resolution images.

[0039] (2) In the representation of pedestrian detection boxes, this technical solution relies only on the key corner points of the upper left / lower right corners, which eliminates the influence of long-distance regression on the quality of pedestrian detection boxes and can generate high-quality detection boxes.

[0040] (3) The human body shape encoding based on regression sets designed in this invention can provide reliable measurement information for the adaptive matching process. This regression set maps long distances to short distances of different proportions, reducing the overall regression error, and has good robustness, especially for large-scale pedestrian detection in a large spatial range.

[0041] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0042] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0043] Figure 1 This is a flowchart of a wide-field-of-view, high-resolution multi-object detection method based on adaptive intelligent matching according to an embodiment of the present invention;

[0044] Figure 2 This is a schematic diagram of the model architecture according to an embodiment of the present invention;

[0045] Figure 3 This is a schematic diagram of the regression set according to an embodiment of the present invention;

[0046] Figure 4 This is a schematic diagram of the body frame corresponding to the upper left and lower right corner points according to an embodiment of the present invention;

[0047] Figure 5 This is a schematic diagram of the structure of a wide field-of-view, high-resolution multi-object detection system based on adaptive intelligent matching according to an embodiment of the present invention. Detailed Implementation

[0048] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0049] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0050] The following description, with reference to the accompanying drawings, describes a wide-field-of-view, high-resolution multi-object detection method and system based on adaptive intelligent matching according to embodiments of the present invention.

[0051] Figure 1 This is a flowchart of a wide-field-of-view, high-resolution multi-object detection method based on adaptive intelligent matching, according to an embodiment of the present invention.

[0052] like Figure 1 As shown, the method includes, but is not limited to, the following steps:

[0053] S1, acquire the original image containing pedestrians and construct an object detection network model; wherein, the object detection network model includes a backbone network and multiple detection heads;

[0054] S2, Input the original image into the backbone network to extract deep semantic features;

[0055] S3, using a preset enhancement method to enhance the sensitivity of deep semantic features to obtain the feature enhancement result of the target boundary;

[0056] S4, the feature enhancement results are input to multiple detection heads respectively, and the key points and regression sets are predicted based on the pedestrian bounding boxes to obtain the key point prediction results and regression set prediction results; wherein, the pedestrian bounding boxes include using the left and upper boundaries of the pedestrian bounding boxes to locate the key points in the upper left corner of the pedestrian, and using the right and lower boundaries to locate the key points in the lower right corner of the pedestrian.

[0057] Understandably, because the original image is a billion pixels in size, it needs to be downsampled to fit the hardware computing resources. Then, a fixed-size sliding window is used to crop it to a suitable size slice before inputting it into the network. In this embodiment of the invention, to avoid the target being segmented and lost at the boundaries of adjacent slices, a certain overlap area is retained between adjacent slices to preserve the complete target as much as possible.

[0058] Specifically, the main model architecture of this invention is as follows: Figure 2 As shown, the system comprises a backbone network and four independent detection heads. The input sliced ​​image is first processed by the backbone network to extract its deep semantic feature representation. Because this invention uses two key corner points (top left and bottom right) to represent the bounding box of a pedestrian, and the top left corner point needs to be located by the left and top boundaries, while the bottom right corner point needs to be located by the right and bottom boundaries, the model has high requirements for object boundary information. Therefore, corner pooling is used after the backbone network to enhance the sensitivity to target boundaries. Finally, the enhanced features are fed into the four detection heads for keypoint and regression set prediction.

[0059] Furthermore, the first detection head is responsible for predicting the location heatmap of the key point in the upper left corner of the pedestrian bounding box, and the second detection head is responsible for regressing the human shape (length and width) encoding set corresponding to the upper left corner. The third detection head is responsible for predicting the location heatmap of the key point in the lower right corner of the pedestrian bounding box, and the fourth detection head is responsible for regressing the human shape (length and width) encoding set corresponding to the lower right corner.

[0060] The following description, in conjunction with the accompanying drawings, details the wide-field-of-view, high-resolution multi-object detection method based on adaptive intelligent matching according to embodiments of the present invention.

[0061] Top-left / bottom-right keypoint prediction: As described above, the network needs to predict the heatmap of the top-left / bottom-right corner points. The ground truth value for each corner point is set to 1, while other locations are considered negative samples. However, keypoints near corner points still have the potential to decode high-quality bounding boxes. Therefore, to reduce the penalty for negative samples around corner points, this embodiment maps a keypoint to a Gaussian decaying circular region and uses a distance-based Focalloss to supervise the learning of the keypoint heatmap. The specific loss function formula is shown below:

[0062]

[0063] Where N is the total number of pedestrians in the current image, p xy It is the true value of the key point at coordinates (x, y) on the heatmap, with a value range of 0-1, p` xy α is the network output prediction value at the corresponding position, β is a hyperparameter used to adjust the weight of easy and difficult samples, and β is a hyperparameter used to adjust the weight of the distance decay term.

[0064] Generation and Prediction of Regression Sets: It is understood that the solution in this invention is designed for pedestrian detection tasks on large-scale images with billions of pixels. Pedestrians are numerous, densely distributed, and have similar appearance features. Therefore, it is necessary to encode robust human shape (length and width) information for each keypoint to facilitate subsequent evaluation and identification of whether a pair of top-left / bottom-right corner points originate from the same object. However, directly using a single long-range regression is affected by regression errors, resulting in insufficient robustness of the decoded human shape and ultimately poor reliability of corner point matching. Therefore, this invention proposes a method using regression sets to encode human shapes, improving robustness and helping to improve the accuracy of subsequent corner point matching.

[0065] Specifically, each keypoint corresponds to two regression sets: one encoding the length of the pedestrian shape and the other encoding its width. The generation of a regression set and the loss function used in supervised learning will be explained in detail below.

[0066] As an example, this invention transforms a single distance regression into a regression set, thereby reducing the error in the final decoding length by decreasing its regression length and difficulty. Figure 3 As shown, a regression set is responsible for encoding a length equal to the width w of the pedestrian shape. The elements within this regression set are divided into g groups, each containing n1, n2, n3... Each subgroup is responsible for a regression length of w*d1, w*d2, w*d3, w*d4, where d1, d2, d3, and d4 are scaling factors. The number of elements in each group, the number of subgroups g, and the scaling factors d for different groups can all be tuned. Here, the parameters are set to g = 4, n1 = n2 = n3 = n4 = 3, d1 = 0.25, d2 = 0.5, d3 = 0.75, and d4 = 1.

[0067] This regression set contains multiple short-range and long-range regressions, transforming the model's optimization objective from the original single-segment long-range regression into a regression set. The specific loss function is shown below:

[0068]

[0069] Where N is the number of pedestrians in the current image, n is the number of elements in a regression set, and w' k,i w is the network output predicted value of the i-th regression element in the regression set for the k-th objective. k,i This represents the true value of the corresponding regression element. To reduce the learning difficulty of the network and further improve the accuracy of the regression, each regression element is mapped to the logarithmic space log(*) to compress the learning range.

[0070] Furthermore, during decoding, the statistical average of this regression set divided by the scaling factor is used as the length and width of the pedestrian body bounding box. This optimization objective includes shorter regression elements, resulting in a reduction in the overall regression error, thereby improving the accuracy and robustness of long-distance regression.

[0071] Furthermore, in the model's inference and prediction phase, the corner heatmap and regression set predicted by the network output need to be preprocessed, then decoded, and finally the top-left and bottom-right corner points with high similarity are matched into a detection box. Specifically:

[0072] Corner heatmap decoding: First, the corner heatmap is activated using the sigmoid function. Then, 3*3 max pooling is used to process the heatmap. Finally, the top-k confidence points on each heatmap are retained, thus obtaining k top-left corner points and k bottom-right corner points. In this embodiment of the invention, k=100.

[0073] Human body length and width decoding: First, the predicted regression set is mapped to the original space using the exponential function exp(*). Then, each element in the regression set corresponding to the top-left keypoint is divided by its corresponding scaling factor d. The statistical average of the regression elements in the current set is then calculated. This decodes the length and width of the body frame corresponding to the top-left keypoint. Similarly, the length and width of the body frame corresponding to the bottom-right keypoint can be obtained.

[0074] Understandably, since the top left and bottom right corners of a pedestrian are predicted independently, it is necessary to pair the corners by evaluating the similarity of the length and width of the human body corresponding to a pair of corners.

[0075] Furthermore, such as Figure 4 As shown, the intersection-union ratio (IUR) of the body bounding boxes corresponding to all top-left and bottom-right corner points is calculated. Boxes with higher IURs are matched as a pair of corner points. The generation of the final detection box is independent of the regressed body bounding box; it is decoded only based on the corresponding top-left and bottom-right keypoints. Therefore, the error in distance regression does not affect the quality of the final detection box. This invention can obtain high-quality detection boxes through accurate pair of corner points.

[0076] In summary, this invention employs an architecture based on adaptive matching and corner modeling. First, it predicts the top-left and bottom-right corners of all pedestrians using heatmaps. Then, it adaptively matches the two corners of the same person to form a detection box. A corner only requires two boundaries for localization, resulting in fewer degrees of freedom than the center point. Predicting accurate corners is easier compared to the center point, avoiding the difficulty in accurately locating the center point. The adaptive matching-based detection box representation method proposed in this invention does not involve distance regression; errors in long-distance regression do not directly affect the quality of the final detection box. This invention proposes an adaptive and robust human shape encoding method. By including sets of short and long distances, it improves the accuracy of regression. The statistical average of the sets is used to decode and obtain the length and width of the human detection box, thus providing reliable guidance for matching the top-left / bottom-right corners. Note that distance regression here is only used to provide a benchmark for corner matching; the final detection box is still decoded from the top-left / bottom-right keypoints, so the quality of the detection box is not affected by distance regression errors.

[0077] The wide-field-of-view, high-resolution multi-object detection method based on adaptive intelligent matching according to embodiments of the present invention can effectively improve the accuracy of automatic pedestrian detection in billion-pixel wide-field-of-view, high-resolution images. It eliminates the influence of long-distance regression on the quality of pedestrian detection boxes, generating high-quality detection boxes. The designed human shape encoding based on regression sets provides reliable metric information for the adaptive matching process. This regression set maps long distances to short distances of different proportions, reducing the overall regression error, and exhibiting good robustness, especially for large-scale pedestrian detection in large spatial areas.

[0078] To achieve the above embodiments, such as Figure 5 As shown, this embodiment also provides a wide field-of-view high-resolution multi-object detection system 10 based on adaptive intelligent matching. The system 10 includes a data acquisition module 100, a feature extraction module 200, a feature enhancement module 300, and a result prediction module 400.

[0079] The data acquisition module 100 is used to acquire original images containing pedestrians and construct an object detection network model; wherein, the object detection network model includes a backbone network and multiple detection heads;

[0080] Feature extraction module 200 is used to input the original image into the backbone network to extract deep semantic features;

[0081] The feature enhancement module 300 is used to enhance the sensitivity of deep semantic features using a preset enhancement method to obtain the feature enhancement result of the target boundary;

[0082] The result prediction module 400 is used to input the feature enhancement results into multiple detection heads to predict key points and regression sets based on the pedestrian bounding boxes of the representation, thereby obtaining key point prediction results and regression set prediction results; wherein, the represented pedestrian bounding boxes include boundary localization of the pedestrian's upper left corner key point using the left and upper boundaries of the pedestrian bounding boxes, and boundary localization of the pedestrian's lower right corner key point using the right and lower boundaries.

[0083] Furthermore, prior to the aforementioned feature extraction module 200, a data preprocessing module is also included, used for:

[0084] The original image is downsampled to obtain a downsampled image;

[0085] The downsampled image is cropped using a sliding window of a preset size to obtain slice images of the preset size.

[0086] Furthermore, the aforementioned result prediction module 400 is also used for:

[0087] The location heatmap of the key point in the upper left corner of the pedestrian bounding box is predicted using the first detection head, and the location heatmap of the key point in the lower right corner of the pedestrian bounding box is predicted using the third detection head to obtain the key point prediction result.

[0088] The regression set prediction result is obtained by using the second detection head to regress the human shape code set corresponding to the upper left key point and the fourth detection head to regress the human shape code set corresponding to the lower right key point.

[0089] The wide-field-of-view, high-resolution multi-object detection system based on adaptive intelligent matching according to embodiments of the present invention can effectively improve the accuracy of automatic pedestrian detection in billion-pixel wide-field-of-view, high-resolution images. It eliminates the influence of long-distance regression on the quality of pedestrian detection boxes, generating high-quality detection boxes. The designed human shape encoding based on regression sets provides reliable metric information for the adaptive matching process. This regression set maps long distances to short distances of different proportions, reducing the overall regression error, and exhibiting good robustness, especially for large-scale pedestrian detection in large spatial areas.

[0090] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0091] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A wide field of view high resolution multi-object detection method based on adaptive intelligent matching, characterized in that, The method comprises the following steps: obtaining an original image containing pedestrians and constructing a target detection network model; wherein the target detection network model comprises a backbone network and a plurality of detection heads; inputting the original image into the backbone network for feature extraction to obtain deep semantic features; using a preset enhancement method to enhance the sensitivity of the deep semantic features to obtain a feature enhancement result of a target boundary; inputting the feature enhancement result into the plurality of detection heads respectively to predict key points and regression sets based on a represented pedestrian bounding box to obtain a key point prediction result and a regression set prediction result; wherein the represented pedestrian bounding box comprises boundary positioning of a pedestrian top-left corner key point using a left boundary and an upper boundary of the pedestrian bounding box and boundary positioning of a pedestrian bottom-right corner key point using a right boundary and a lower boundary of the pedestrian bounding box; the prediction of key points and regression sets based on the represented pedestrian bounding box to obtain the key point prediction result and the regression set prediction result comprises: predicting a position heat map of the top-left corner key point of the pedestrian bounding box using a first detection head and predicting a position heat map of the bottom-right corner key point of the pedestrian bounding box using a third detection head to obtain the key point prediction result; regressing a human shape code set corresponding to the top-left corner key point using a second detection head and regressing a human shape code set corresponding to the bottom-right corner key point using a fourth detection head to obtain the regression set prediction result; heat map decoding: using a sigmoid function to activate the position heat map, processing the position heat map after activation using maximum pooling, and obtaining key points with a confidence greater than a preset threshold on each position heat map to obtain a plurality of top-left corner key points and a plurality of bottom-right corner key points; and human shape decoding: mapping the regression set to the original space using an exponential function, calculating the statistical average of the regression elements of the current set based on the elements of the regression set of the top-left corner key point and the corresponding scaling factor to decode the length and width of the body frame corresponding to the top-left corner key point.

2. The method of claim 1, wherein, Before the original image is input into the backbone network, the method further comprises: down-sampling the original image to obtain a down-sampled image; cropping the down-sampled image using a sliding window of a preset size to obtain a slice image of a preset size.

3. The method of claim 1, wherein, mapping the top-left corner key point or the bottom-right corner key point to a circular region with Gaussian attenuation to obtain a loss function of the corresponding position heat map: where N is the total number of pedestrians in the current image, p xy is the ground truth of the key point at coordinate (x, y) on the heat map, taking a value in the range 0-1, p xy is the network output prediction value at the corresponding position, a is a hyperparameter adjusting the weight of difficult sample, and β is a hyperparameter adjusting the weight of distance decay term.

4. The method of claim 1, wherein, the optimization target of the target detection network model is changed from single-section long-distance regression to regression set, so the corresponding loss function is: where N is the number of pedestrians in the current image, n is the number of elements contained in a regression set, w k,i is the network output prediction value of the i-th regression element in the regression set of the k-th target, w k,i is the true value of the corresponding regression element.

5. The method of claim 4, wherein, after heat map decoding and human shape decoding, the method further comprises: calculating the intersection over union of the body frames corresponding to the plurality of top-left corner key points and the plurality of bottom-right corner key points to obtain an intersection over union calculation result; comparing the intersection over union calculation result, and selecting a preset number of intersection over unions for matching according to the comparison result to obtain a pair of corner key points.

6. A wide field high resolution multi-object detection system based on adaptive smart matching using the method as claimed in claim 1, wherein, comprise: a data acquisition module for obtaining an original image containing pedestrians and constructing a target detection network model; wherein the target detection network model comprises a backbone network and a plurality of detection heads; The feature extraction module is configured to input the original image into the backbone network to extract deep semantic features; The feature enhancement module is configured to use a preset enhancement method to enhance the sensitivity of the deep semantic features to obtain a feature enhancement result of the target boundary; The result prediction module is configured to input the feature enhancement result into the multiple detection heads respectively, to predict key points and a regression set based on a represented pedestrian bounding box to obtain a key point prediction result and a regression set prediction result; wherein the represented pedestrian bounding box includes boundary positioning of a pedestrian top-left corner key point using a left boundary and an upper boundary of a pedestrian bounding box, and boundary positioning of a pedestrian bottom-right corner key point using a right boundary and a lower boundary of the pedestrian bounding box.

7. The system of claim 6, wherein, Before the feature extraction module, the data preprocessing module is further included, which is configured to: down-sample the original image to obtain a down-sampled image; use a preset size sliding window to crop the down-sampled image to obtain a preset size slice image.

Images