Method and device for detecting attitude of small target in dense crowd in real time and medium

Abstract

The invention relates to the field of machine vision, and particularly discloses a real-time attitude detection method and device for small targets in dense crowds and a medium, a real-time attitude detection model is constructed by adopting a modified Alpharose framework, and the method specifically comprises the following steps: replacing an end side target detection algorithm in the Alpharose framework with a YOLOV10 algorithm; setting a human body detection confidence coefficient threshold value in the YOLOV10 algorithm as a minimum value; performing post-processing on the YOLOV10 algorithm by adopting a non-maximum suppression algorithm based on a CIOU loss function; and respectively carrying out skeleton point detection on n human body detection frames pointing to the same target through a Fastpose skeleton point detection network. By improving the post-processing logic of the human body detector and skeleton point detection, the small target posture detection is realized, the detection efficiency is high, and the target is not easy to lose.

Description

Technical Field

[0001] The present invention relates to the field of machine vision, and in particular to a method, device and medium for real-time posture detection of small targets in a dense crowd. Background Art

[0002] With the rapid development of machine vision, the diversity of various human detection algorithms is becoming more and more abundant, and the recognition efficiency and accuracy of human targets, expressions, postures and other data are getting higher and higher. At present, the realization of human posture detection mainly includes two directions: end-to-end posture detection and two-stage posture detection; the two stages are human body detection and human skeleton point detection. However, in actual project applications, it is found that there are still some problems that cannot be solved in the existing human posture detection algorithms, especially the Alphapose algorithm. Although its recognition accuracy in multi-person posture estimation in natural scenes with dense crowds is higher than other similar models, it consumes more computing resources and has a slow detection speed. To achieve real-time detection of the real-time posture of small targets in dense crowds, it requires extremely high computing platform support.

[0003] The end-to-end posture detection neural network consists of a human body detection network branch and a skeleton point detection branch. Compared with the two-stage posture detection, the end-to-end posture detection is constrained by both branches at the same time, resulting in poor generalization for small targets. The accuracy of the two-stage posture detection algorithm is greatly affected by the human body detection results.

[0004] The original Alphapose algorithm (multi-person pose estimation algorithm) uses the YOLOX human detector or the YOLOV3-SPP human detector. The network parameters of the above human detectors are relatively large. Although the recall rate of small targets has been improved, the reasoning time has also increased. Although there are attempts to replace other detectors to improve detection efficiency in the prior art, the detection efficiency is still not ideal, and the calculation of conventional hardware is still difficult to meet sufficient recognition efficiency. In addition, in actual scenarios, since the human body detection frame does not necessarily correspond to a complete human body, it is easy to cause the human skeleton point detection to lose some skeleton points, resulting in inaccurate final detection results. Summary of the invention

[0005] In order to overcome the above problems, the present invention provides a method, device and medium for real-time posture detection of small targets in a dense crowd.

[0006] The present invention provides a method for real-time posture detection of small targets in a dense crowd, comprising: using a modified Alphapose framework to construct a real-time posture detection model;

[0007] Use multiple pose-labeled images, divided into training set, validation set and test set;

[0008] The real-time posture detection model is trained using the training set and the validation set until the real-time posture detection model passes the test of the test set;

[0009] Use the tested real-time posture detection model to perform real-time posture detection on small targets in dense crowds and output the real-time posture detection results;

[0010] The modified Alphapose framework is specifically:

[0011] Use the YOLOV10 algorithm to replace the end-side object detection algorithm in the Alphapose framework to obtain the human detection frame in the image;

[0012] Set the human detection confidence threshold in the YOLOV10 algorithm to the minimum value;

[0013] The non-maximum suppression algorithm based on CIOU loss function is used for post-processing of YOLOV10 algorithm to screen n human detection boxes pointing to the same target;

[0014] Through the Fastpose skeleton point detection network, skeleton point detection is performed on n human detection frames pointing to the same target, and n skeleton point detection results of the target are extracted;

[0015] Traverse the 26 bone points corresponding to the n bone point detection results, take the bone point with the largest confidence, and reassemble it into a posture detection result corresponding to the 26 human bone points.

[0016] Preferably, in the non-maximum suppression algorithm based on the CIOU loss function, the calculation formula of the CIOU loss function is as follows:

[0017]

[0018] Among them, w' and h' are the width and height of the intersection of the two human detection frames, w1, h1, w2 and h2 are the width and height of the two human detection frames respectively, ρ is the distance between the center points of the two human detection frames, and c is the length of the hypotenuse of the bounding box composed of the two targets.

[0019] Preferably, the YOLOV10 algorithm specifically adopts one of the YOLOV10n algorithm or the YOLOV10s algorithm.

[0020] Preferably, the non-maximum suppression algorithm based on the CIOU loss function is used to post-process the YOLOV10 algorithm to screen n human body detection frames pointing to the same target, which specifically includes the following steps:

[0021] Step a1, export all human detection frames pointing to one of the same targets in the human detection frames identified by the YOLOV10 algorithm with the human detection confidence threshold set to the minimum value as a list BoxLists, and create a new list ResultBoxLists for saving the final results;

[0022] Step a2: sort each human body detection box in the list BoxLists in descending order according to the confidence level to obtain the list SortBoxLists;

[0023] Step a3, traverse each human detection frame in the list SortBoxLists, calculate the CIOU loss function for the first human detection frame ranked first in the list SortBoxLists and the remaining human detection frames, and record the human detection frames whose CIOU loss function calculation results are greater than 0.25;

[0024] Step a4, inserting the human body detection box ranked first in the list SortBoxLists into the list ResultBoxLists;

[0025] Step a5, deleting the first human body detection frame in the list SortBoxLists and the human body detection frame recorded in step a3 from the list SortBoxLists;

[0026] Step a6, looping through steps a3 to a5 until the number of human body detection boxes in the list SortBoxLists is less than or equal to one;

[0027] Step a7: insert the remaining human body detection frames in the list SortBoxLists into the list ResultBoxLists, and output n human body detection frames pointing to the target; wherein n is the number of human body detection frames finally included in the list ResultBoxLists.

[0028] Preferably, the post-processing of the YOLOV10 algorithm using a non-maximum suppression algorithm based on a CIOU loss function to screen n human body detection frames pointing to the same target also includes the following steps:

[0029] In the human detection frame identified by the YOLOV10 algorithm with the human detection confidence threshold set to the minimum value, all human detection frames pointing to one of the same targets are calculated using the IOU loss function, and the human detection frames with calculation results greater than 0.8 are merged into a new human detection frame.

[0030] The present invention also provides a device for real-time posture detection of small targets in dense crowds, comprising: a building module, a training module, a testing module and a detection module;

[0031] The building module is used to build a real-time posture detection model using a modified Alphapose framework;

[0032] The training module is used to use a plurality of posture-labeled images, which are divided into a training set, a validation set and a test set;

[0033] The testing module is used to train the real-time posture detection model using the training set and the validation set until the real-time posture detection model passes the test of the test set;

[0034] The detection module is used to use a tested real-time posture detection model to perform real-time posture detection on small targets in a dense crowd and output real-time posture detection results;

[0035] The modified Alphapose framework is specifically:

[0036] The YOLOV10 algorithm is used to replace the end-side target detection algorithm in the Alphapose framework to obtain the human detection frame in the image; the human detection confidence threshold in the YOLOV10 algorithm is set to the minimum value; the non-maximum suppression algorithm based on the CIOU loss function is used to post-process the YOLOV10 algorithm to screen n human detection frames pointing to the same target; the Fastpose skeleton point detection network is used to perform skeleton point detection on the n human detection frames pointing to the same target, and the n skeleton point detection results of the target are extracted; the skeleton points corresponding to the 26 points in the n skeleton point detection results are traversed, the skeleton points with the largest confidence are taken, and they are recombined into posture detection results corresponding to the 26 human skeleton points.

[0037] The present invention provides a computer-readable storage medium, which includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute the above-mentioned method for real-time posture detection of small targets in dense crowds.

[0038] The beneficial effects of the present invention are:

[0039] (1) By using the two-stage pose detection Alphapose network, the detector is replaced with YOLOV10n or YOLOV10s, and the detection confidence threshold is set to 0 to retain more small object boxes. Finally, the NMS with CIOU loss function is used to filter the duplicate boxes, which not only solves the problem of long reasoning time, but also solves the problem of small object loss, thereby improving the recall rate of the detector.

[0040] (2) By merging the skeleton point results of multiple Fastpose detections, Fastpose detection is performed on multiple incomplete human detection frames to obtain a skeleton point set. In the 26 skeleton point sets, the skeleton point with the highest confidence in each point in the set is extracted to merge into a new human skeleton. This improves the accuracy of posture detection for incomplete human bodies in the two-stage method and also improves the recognition rate of human skeleton points of the posture detection algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The present invention will be further described below in conjunction with the accompanying drawings, wherein:

[0042] Figure 1 This is a flow chart of a method according to Embodiment 1 of the present invention;

[0043] Figure 2 This is a comparison chart of the results before and after screening of the non-maximum suppression algorithm based on the CIOU loss function based on different inputs in Example 2 of the present invention;

[0044] Figure 3 This is a schematic diagram of the skeleton point fusion process of the Fastpose posture detection network according to the second embodiment of the present invention. DETAILED DESCRIPTION

[0045] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0046] See also Figure 1 As a first embodiment of the present invention, a method for real-time posture detection of small targets in a dense crowd is disclosed, comprising the following steps:

[0047] S1. Establishment: Use the modified Alphapose framework to build a real-time posture detection model;

[0048] S2, training: multiple pose-labeled images are used and divided into training set, validation set and test set;

[0049] S3, testing: using the training set and the validation set to train the real-time posture detection model until the real-time posture detection model passes the test of the test set;

[0050] S4, detection: Use the tested real-time posture detection model to perform real-time posture detection on small targets in dense crowds and output the real-time posture detection results;

[0051] The modified Alphapose framework is specifically:

[0052] S11. Use the YOLOV10 algorithm to replace the end-side object detection algorithm in the Alphapose framework to obtain a human detection frame in the image.

[0053] S12, setting the human detection confidence threshold in the YOLOV10 algorithm to the minimum value (i.e., 0);

[0054] S13, using the non-maximum suppression algorithm (NMS algorithm) based on the CIOU loss function to post-process the YOLOV10 algorithm and screen n human detection frames pointing to the same target;

[0055] S14, using the Fastpose skeleton point detection network, respectively performing skeleton point detection on n human body detection frames pointing to the same target, and extracting n skeleton point detection results of the target;

[0056] S15, traversing the skeleton points corresponding to the 26 points in the n skeleton point detection results, taking the skeleton point with the largest confidence, and recombining them into posture detection results corresponding to the 26 human skeleton points.

[0057] This embodiment also discloses a device for real-time posture detection of small targets in dense crowds, including: a building module, a training module, a testing module and a detection module;

[0058] The building module is used to build a real-time posture detection model using a modified Alphapose framework;

[0059] The training module is used to use a plurality of posture-labeled images, which are divided into a training set, a validation set and a test set;

[0060] The testing module is used to train the real-time posture detection model using the training set and the validation set until the real-time posture detection model passes the test of the test set;

[0061] The detection module is used to use a tested real-time posture detection model to perform real-time posture detection on small targets in a dense crowd and output real-time posture detection results;

[0062] This embodiment uses the two-stage posture detection Alphapose network, replaces the detector with YOLOV10n or YOLOV10s, and retains more small target frames by setting the detection confidence threshold to 0. Finally, the NMS of the CIOU loss function is used to filter the duplicate frames, thereby solving the problems of long reasoning time and high computing resource consumption, solving the problem of small target loss, and improving the recall rate of the detector, providing the possibility for real-time reasoning of two-stage posture detection embedded terminal deployment.

[0063] By merging the skeleton point results of multiple Fastpose detections, Fastpose detection is performed on multiple incomplete human detection frames to obtain a skeleton point set. In the 26 skeleton point sets, the skeleton point with the highest confidence in each point in the set is extracted to merge into a new human skeleton. This improves the accuracy of posture detection for incomplete human bodies in the two-stage detection, and also improves the recognition rate of human skeleton points of the posture detection algorithm.

[0064] See also Figures 2 to 3 As the second embodiment of the present invention, which is a more specific embodiment of the first embodiment, the detector is replaced with YOLOV10n or YOLOV10s based on Alphapose, and the posture detection uses the 26 human skeleton point detection network of Fastpose. More small targets are retained by reducing the YOLOV10 human detection confidence threshold, and then the repeated frames are filtered based on the non-maximum threshold (NMS) of CIOU, and finally the skeleton points facing the same target are screened out and the skeleton points are merged.

[0065] The improvement of the Alphapose framework in this embodiment is divided into the following two steps:

[0066] Step 1: Improvement of detector;

[0067] (11)Replace the detector with YOLOV10n or YOLOV10s.

[0068] YOLOv10 is an object detection algorithm trained with NMS without post-processing. Use YOLOV10n or YOLOV10s to replace Alphapose's YOLOX-x or YOLOV3-SPP detection. Compared with YOLOX-x or YOLOV3-SPP, YOLOV10n or YOLOV10s is a smaller network, and the model inference speed can be several times faster.

[0069] (12) The confidence threshold of target detection is minimized, retaining more small target boxes.

[0070] After actual testing and verification, when the size of the network and input is fixed, as the confidence threshold increases, the loss of small targets becomes more serious; in particular, the confidence of small targets is mostly less than 0.25. When the confidence threshold is fixed, YOLOV10N (640*640) is compared with YOLOV10N (1080*1080), and YOLOV10S (640*640) is compared with YOLOV10S (1080*1080). After actual testing, the conclusion is drawn: when high-resolution network input is recognized, although the neural network obtains more detection information in large-size input, it will also lose some small targets when the confidence threshold is less than 0.25. Combined with other actual test results, the confidence threshold is inversely proportional to the loss of small targets in human detection; when the confidence threshold is set to 0, the loss of small targets in human detection is the least. When the confidence threshold is reduced, more duplicate frames will appear when pointing to a target. To solve this problem, the present invention uses a non-maximum suppression algorithm (NMS) based on CIOU to filter out the redundant duplicate frames that appear in step (12). The effect after filtering is as follows: Figure 2 shown.

[0071] (13) The human detection frame is subjected to a non-maximum suppression algorithm to filter out frames pointing to the same target.

[0072] Specifically, the sub-step S31 of this embodiment specifically includes the following sub-steps:

[0073] Step a1, export all human detection frames pointing to one of the same targets in the human detection frames identified by the YOLOV10 algorithm with the human detection confidence threshold set to the minimum value as a list BoxLists, and create a new list ResultBoxLists for saving the final results;

[0074] Step a2: sort each human body detection box in the list BoxLists in descending order according to the confidence level to obtain the list SortBoxLists;

[0075] Step a3, traverse each human detection frame in the list SortBoxLists, calculate the CIOU loss function for the first human detection frame ranked first in the list SortBoxLists and the remaining human detection frames, and record the human detection frames whose CIOU loss function calculation results are greater than 0.25;

[0076] Step a4, inserting the human body detection box ranked first in the list SortBoxLists into the list ResultBoxLists;

[0077] Step a5, deleting the first human body detection frame in the list SortBoxLists and the human body detection frame recorded in step a3 from the list SortBoxLists;

[0078] Step a6, looping through steps a3 to a5 until the number of human body detection boxes in the list SortBoxLists is less than or equal to one;

[0079] Step a7: insert the remaining human body detection frames in the list SortBoxLists into the list ResultBoxLists, and output n human body detection frames pointing to the target; wherein n is the number of human body detection frames finally included in the list ResultBoxLists.

[0080] In this embodiment, in the non-maximum suppression algorithm based on the CIOU loss function, the calculation formula of the CIOU loss function is as follows:

[0081]

[0082] Among them, w' and h' are the width and height of the intersection of the two human detection frames, w1, h1, w2 and h2 are the width and height of the two human detection frames respectively, ρ is the distance between the center points of the two human detection frames, and c is the length of the hypotenuse of the bounding box composed of the two targets (that is, the leftmost boundary, rightmost boundary, topmost boundary and bottommost boundary of the two frames are taken to synthesize a new frame).

[0083] YOLOV10 is a target detection algorithm trained by NMS without post-processing, which achieves better robustness in small targets. When the confidence threshold is set to 0, CIOU NMS filtering is performed, and the reasoning process is basically the same as the YOLOX-x and YOLOV3-SPP detection algorithms. This embodiment introduces YOLOV10 lightweight network reasoning without changing the overall process of Alphapose reasoning, which reduces the model reasoning time and retains more small targets.

[0084] Step 2: Posture Detection Improvement

[0085] The posture detection network used in the present invention is Fastpose. The detection result of step 1 is obtained, multiple target frames pointing to the same person are screened out, posture detection is performed, and then the posture skeleton points are merged.

[0086] (21)26 Fastpose skeleton point detection algorithms.

[0087] Fastpose is an algorithm that detects 26 human skeleton points. The 26 skeleton points cover the entire human body from the top of the head to the feet.

[0088] (22) Through the CIOU NMS algorithm, the boxes pointing to the same target are filtered;

[0089] (23) Fastpose detection is performed on a set of boxes pointing to the same target, and then skeleton points are fused.

[0090] Traverse the list PointsLists of boxes pointing to the same target in the above step (21), perform Fasepose skeleton point detection on the n target boxes pointing to the same person respectively, and obtain n*26 skeleton points.

[0091] If n is 1, the 26 skeleton points obtained are the final skeleton points of the human body. If n is greater than 1, the 26 skeleton points are traversed, and the point with the highest confidence among the n 26 skeleton points is taken as the point of the human body at the corresponding position, and reassembled into 26 human skeleton points.

[0092] In other embodiments, sub-step S31 may further include the following sub-steps:

[0093] In the human detection frame identified by the YOLOV10 algorithm with the human detection confidence threshold set to the minimum value, all human detection frames pointing to one of the same targets are calculated using the IOU loss function, and the human detection frames with calculation results greater than 0.8 are merged into a new human detection frame; that is, the leftmost boundary, rightmost boundary, topmost boundary and bottommost boundary of the two frames are taken to synthesize a new human detection frame.

[0094] That is, the CIOU NMS filtering algorithm for the human detection frame when the confidence threshold is 0 can be replaced by merging the human detection frame with a confidence threshold of 0 with overlapping (IOU of two frames is greater than 0.8) (that is, taking the leftmost boundary, rightmost boundary, topmost boundary and bottommost boundary of two human detection frames to synthesize a new human detection frame).

[0095] The present invention also discloses a terminal device, including a processor and a storage device, the storage device is used to store one or more programs; when the one or more programs are executed by the processor, the processor implements the above-mentioned method for real-time posture detection of small targets in dense crowds. The processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc. The processor is the control center of the test equipment, and various interfaces and lines are used to connect the various parts of the entire test equipment.

[0096] The storage device can be used to store computer programs and / or modules. The processor realizes various functions of the terminal device by running or executing the computer programs and / or modules stored in the storage device and calling the data stored in the storage device. The storage device can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application required for a function, etc.; the data storage area can store data created according to the use of the terminal device, etc. In addition, the storage device can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device, or other volatile solid-state storage devices.

[0097] Among them, if the module / unit integrated in the device for real-time posture detection of small targets in dense crowds is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the present invention implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in at least one computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying computer program code, recording medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium, etc.

[0098] It should be noted that the embodiments of the devices and apparatuses described above are only illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place, or may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

Claims

1. A method for real-time posture detection of small targets in dense crowds, characterized in that: include: The modified Alphapose framework is used to build a real-time posture detection model; Use multiple pose-labeled images, divided into training set, validation set and test set; The real-time posture detection model is trained using the training set and the validation set until the real-time posture detection model passes the test of the test set; Use the tested real-time posture detection model to perform real-time posture detection on small targets in dense crowds and output real-time posture detection results; The modified Alphapose framework is specifically: Use the YOLOV10 algorithm to replace the end-side object detection algorithm in the Alphapose framework to obtain the human detection frame in the image; Set the human detection confidence threshold in the YOLOV10 algorithm to the minimum value; The non-maximum suppression algorithm based on CIOU loss function is used for post-processing of YOLOV10 algorithm to screen n human detection boxes pointing to the same target; Through the Fastpose skeleton point detection network, skeleton point detection is performed on n human detection frames pointing to the same target, and n skeleton point detection results of the target are extracted; Traverse the 26 bone points corresponding to the n bone point detection results, take the bone point with the largest confidence, and reassemble it into a posture detection result corresponding to the 26 human bone points.

2. The method for real-time posture detection of small targets in a dense crowd according to claim 1, characterized in that: In the non-maximum suppression algorithm based on the CIOU loss function, the calculation formula of the CIOU loss function is as follows: Among them, w' and h' are the width and height of the intersection of the two human detection frames, w1, h1, w2 and h2 are the width and height of the two human detection frames respectively, ρ is the distance between the center points of the two human detection frames, and c is the length of the hypotenuse of the bounding box composed of the two targets.

3. The method for real-time posture detection of small targets in a dense crowd according to claim 1, characterized in that: The YOLOV10 algorithm specifically adopts one of the YOLOV10n algorithm or the YOLOV10s algorithm.

4. The method for real-time posture detection of small targets in a dense crowd according to claim 1, characterized in that: The non-maximum suppression algorithm based on the CIOU loss function is used to post-process the YOLOV10 algorithm to screen n human detection frames pointing to the same target, which specifically includes the following steps: Step a1, export all human detection frames pointing to one of the same targets in the human detection frames identified by the YOLOV10 algorithm with the human detection confidence threshold set to the minimum value as a list BoxLists, and create a new list ResultBoxLists for saving the final results; Step a2: sort each human body detection box in the list BoxLists in descending order according to the confidence level to obtain the list SortBoxLists; Step a3, traverse each human detection frame in the list SortBoxLists, calculate the CIOU loss function for the first human detection frame ranked first in the list SortBoxLists and the remaining human detection frames, and record the human detection frames whose CIOU loss function calculation results are greater than 0.25; Step a4, inserting the human body detection box ranked first in the list SortBoxLists into the list ResultBoxLists; Step a5, deleting the first human body detection frame in the list SortBoxLists and the human body detection frame recorded in step a3 from the list SortBoxLists; Step a6, looping through steps a3 to a5 until the number of human body detection boxes in the list SortBoxLists is less than or equal to one; Step a7: insert the remaining human body detection frames in the list SortBoxLists into the list ResultBoxLists, and output n human body detection frames pointing to the target; wherein n is the number of human body detection frames finally included in the list ResultBoxLists.

5. The method for real-time posture detection of small targets in a dense crowd according to claim 1, characterized in that: The method adopts the non-maximum suppression algorithm based on the CIOU loss function to post-process the YOLOV10 algorithm to screen n human detection frames pointing to the same target, and also includes the following steps: In the human detection frame identified by the YOLOV10 algorithm with the human detection confidence threshold set to the minimum value, all human detection frames pointing to one of the same targets are calculated using the IOU loss function, and the human detection frames with calculation results greater than 0.8 are merged into a new human detection frame.

6. A device for real-time posture detection of small targets in dense crowds, characterized in that: include: Building modules, training modules, testing modules and detecting modules; The building module is used to build a real-time posture detection model using a modified Alphapose framework; The training module is used to use a plurality of posture-labeled images, which are divided into a training set, a validation set and a test set; The testing module is used to train the real-time posture detection model using the training set and the validation set until the real-time posture detection model passes the test of the test set; The detection module is used to use a tested real-time posture detection model to perform real-time posture detection on small targets in a dense crowd and output real-time posture detection results; The modified Alphapose framework is specifically: The YOLOV10 algorithm is used to replace the end-side target detection algorithm in the Alphapose framework to obtain the human detection frame in the image; the human detection confidence threshold in the YOLOV10 algorithm is set to the minimum value; the non-maximum suppression algorithm based on the CIOU loss function is used to post-process the YOLOV10 algorithm to screen n human detection frames pointing to the same target; the Fastpose skeleton point detection network is used to perform skeleton point detection on the n human detection frames pointing to the same target, and the n skeleton point detection results of the target are extracted; the skeleton points corresponding to the 26 points in the n skeleton point detection results are traversed, the skeleton points with the largest confidence are taken, and they are recombined into posture detection results corresponding to the 26 human skeleton points.

7. A computer-readable storage medium, characterized in that: The computer-readable storage medium includes a stored computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located is controlled to execute the method for real-time posture detection of small targets in a dense crowd as described in any one of claims 1 to 5.

Images