Special crowd identification method of improved yov8 blind guiding robot dog
By improving the YOLOv8 model and hardware optimization, the problem of insufficient accuracy and speed for guided robot dogs identifying special groups in complex environments is solved, and more efficient and accurate identification of special groups is achieved to ensure the safe travel of visually impaired people.
Patent Information
- Application Number
- CN202411599151.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-11
- Publication Date
- 2025-03-04
AI Technical Summary
Existing guide robot dogs have problems with insufficient recognition accuracy and speed when identifying special groups, especially in complex environments, it is difficult to effectively identify special groups such as blind people, doctors, firefighters and traffic police.
Improve the YOLOv8 model, use GIoU loss function to replace CIoU, and use SiLU activation function in the Conv layer, combine multi-scale feature fusion to optimize the network structure to improve the accuracy and speed of object detection. At the same time, hardware optimization is based on Unitree GO 1 quadruped robot dog, increasing battery life and reinforcement of limb joints, built-in high-performance computing power and a high-resolution camera to support high-speed real-time recognition.
It significantly improves the accuracy and speed of the identification of special groups by guided robot dogs, helps visually impaired people travel safely in complex environments, and can identify dangerous situations in a timely manner and seek help from special groups.
Smart Images
Figure CN119559665A_ABST
Abstract
Description
Technical Field
[0001] The technical solution of the present application relates to the field of face recognition technology, and more specifically, to a method for improving the recognition of special groups of people by the yolov8 guide robot dog. Background Art
[0002] Visually impaired people face many unsafe factors in their daily travel. Although traditional guide canes and guide dogs have alleviated these problems to a certain extent, they still have limitations. Therefore, research on guide robots based on artificial intelligence technology can effectively help visually impaired people reduce hidden dangers. How to use intelligent guide robot dogs to detect and identify special uniforms and blind people's families, and help visually impaired people seek help from special people in time when they are in danger has become a focus of widespread attention. Summary of the invention
[0003] The technical solution of the present application provides a method for improving the special population identification of the guide robot dog of yolov8, so as to overcome at least one technical problem existing in the prior art.
[0004] The technical solution embodiment of the present application provides a method for identifying special groups of people by improving the guide robot dog of yolov8, including:
[0005] Step 1: Create a special population dataset
[0006] Create the required dataset, which contains 740 images in total. According to the distribution ratio of 7:2:1, it is divided into 518 images for training set, 148 images for test set, and 74 images for validation set. There are 4 types of targets in total: family, doctor, firefighter, and traffic police. The robot dog camera takes 200 photos of family members, 215 photos of doctors are collected from the Internet, 175 photos of firefighters, and 150 photos of traffic police.
[0007] Step 2: Design the SG-YOLOv8 network based on the optimized YOLOv8 network structure
[0008] The structure of the SG-YOLOv8 network includes an input layer, a backbone network, a neck network, and a head network; the input layer accepts image input, the backbone network extracts image features, the neck network performs multi-scale fusion, and the head network processes the fused features and obtains the predicted category and the position coordinates of the target box; the backbone network mainly includes the C2f module and the SPPF module; the C2f module improves the expressiveness of features through cross-stage feature fusion and is used to extract low-level and mid-level features in the image, and the SPPF module is a spatial pyramid pooling fusion module. The SPPF module divides the feature map into sub-regions of different scales, performs pooling operations separately, and splices the pooling results of all sub-regions together;
[0009] SG-YOLOv8 replaces the original loss function CIoU of YOLOv8 with the GIoU loss function; at the same time, the activation function ReLU is replaced with SiLU in the Conv layer to enhance the performance of the model;
[0010] (1) Improve the loss function
[0011] The loss function in the YOLOv8 model is divided into three parts: classification loss, target confidence loss, and positioning loss. The loss of the bounding box uses CIoU Loss. CIoU is an improvement on DIoU. It adds the loss of the predicted box and the loss of length and width, so that the detection is constantly adjusted and closer to the true value. However, CIoU is not sensitive enough when dealing with small targets. It is easy to ignore the shape change of the bounding box and only pay attention to the overlap between position and size, and has poor robustness.
[0012] Therefore, in order to improve the positioning loss function of bounding box regression, the technical solution of this application introduces GloU Loss; GIoU can reflect the distance and overlap between two boxes, and has better robustness to the position offset and shape change of the bounding box. When training the target detection model, using GIOU as the loss function can bring better optimization effect. The calculation formula is as follows:
[0013]
[0014] Among them, A represents the detection box, B represents the real box, and C represents the smallest external rectangular box;
[0015] In addition, Inner-IoU Loss is used, and the scale factor ratio is set to 0.75 to control the generation of auxiliary borders of different scales for loss calculation; Inner-IoU is applied to the existing IoU-based border regression loss function, defined as formula (3), and faster and more effective regression results can be obtained;
[0016] L inner-GIoU =LGIoU +IoU-IoU inner (3)
[0017] (2) Improvements in activation functions
[0018] The activation function used by the YOLOv8 network is replaced by the ReLU activation function with SiLU, and its calculation formula is (5):
[0019]
[0020] Step 3: Hardware optimization based on Unitree GO 1 quadruped robot dog
[0021] Unitree GO 1 from Yushu Technology was selected as the quadruped robot dog used in the experiment. The robot dog itself contains 5 groups of fisheye binocular cameras, which are located on its head, side and abdomen. It has the industry's cutting-edge visual calling ability and is suitable as a robot for blind people. Considering the complex road conditions and time issues when blind people go out, the robot dog is optimized and improved. A larger battery life battery is used and the protection of the robot dog's limb joints is reinforced. In addition, the robot dog has three JetsonNano control method boards and two Raspberry controllers built in, with a total of 16-core CPU+GPU computing power, and an external camera IVG-5MP is added. With its high resolution, high pixel and easy access to the robot dog platform, the robot dog can perform high-speed real-time recognition in scenes where special people are identified, thereby assisting the blind in identifying special people.
[0022] Step 4: Verification experiment of special population identification based on SG-YOLOv8 method
[0023] (1) Evaluation indicators
[0024] The accuracy and effectiveness of the model were evaluated using precision, recall, F1-core and average precision. True positive TP: the true category of the sample is a positive example, and the model predicts it to be a positive example, so the prediction is correct; true negative TN: the true category of the sample is a negative example, and the model predicts it to be a negative example, so the prediction is correct; false positive FP: the true category of the sample is a negative example, but the model predicts it to be a positive example, so the prediction is wrong; false negative FN: the true category of the sample is a positive example, but the model predicts it to be a negative example, so the prediction is wrong;
[0025] Precision refers to the ratio of the number of correctly assigned positive samples to the total number of assigned positive samples, as shown in formula (6), where TP is the positive detection rate, which refers to the number of correctly detected targets; FP is the false detection rate, which refers to the number of incorrectly detected targets;
[0026]
[0027] The recall rate refers to the ratio of the number of correctly classified positive samples to the total number of positive samples, as shown in formula (7), where FN is the missed detection rate, that is, the number of targets that are not detected;
[0028]
[0029] F1-Score is the harmonic mean of precision and recall, and is calculated as follows:
[0030]
[0031] mAP50 (Mean Average Precision at 50) is one of the indicators used to evaluate the performance of target detection methods. It is obtained by calculating the average precision AP (Average Precision) under different confidence thresholds. In the calculation process, the predicted bounding boxes are first sorted according to the confidence, then the precision and recall rate under each confidence threshold are calculated, and the precision-recall curve is plotted. Finally, the mAP50 value is obtained by calculating the average area under the curve. The higher the mAP50 value, the better the accuracy and recall rate of the model in predicting bounding boxes. mAP50 is the average precision mean when the IoU threshold is 0.5, and mAP50-95 represents the average precision mean under different IoUs. The calculation formula is as follows:
[0032]
[0033]
[0034] Among them, AP is calculated by interpolation method to interpolate the precision-recall curve. The interpolation P in formula (9) interpolation is the maximum precision value between a given recall value r and the next recall value r' in the curve. In formula (10), r1, r2, ..., r i ,r i+1 ,...,r n-1 is the recall value corresponding to the first interpolation point in the precision interpolation segment arranged in ascending order. In formula (11), n represents the number of recognized target categories, i represents the i-th recognized category, and AP i Represents the average accuracy for the i-th category;
[0035] Step 5: Transplant the trained SG-YOLOv8 network to the quadruped robot dog to identify special people.
[0036] One embodiment of this specification can at least achieve the following beneficial effects: The technical solution of this application proposes a new YOLO (SG-YOLOv8) detection and recognition algorithm based on changing the loss function and the activation function. The focus is that in the YOLOv8 model, SiLU is used to replace the activation function ReLU, and GIoU is used to replace the loss function CIoU, and the data of four special groups of people, including blind family members, doctors, firefighters, and traffic police, are identified and detected. In addition, the Unitree Go1 quadruped robot is used as a carrier for practical application. The technical solution of this application aims at the problem of target recognition accuracy of guided quadruped robot dogs, effectively improves the accuracy and speed of recognition, and helps visually impaired people travel safely. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] In order to more clearly illustrate the embodiments of the technical solution of the present application or the technical solution in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the technical solution of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0038] Figure 1 A dataset distributed according to the length and width of labels;
[0039] Figure 2 This is the network structure diagram of SG-YOLOv8;
[0040] Figure 3 Optimized design for the appearance of a four-legged guide dog robot;
[0041] Figure 4 It is the F1 score curve graph;
[0042] Figure 5 This is a comparison chart of the four indicators of the SG-YOLOv8 and YOLOv8 models;
[0043] Figure 6 Schematic diagram of a four-legged robot dog identifying blind family members;
[0044] Figure 7 Schematic diagram of a four-legged robot dog identifying special groups such as doctors and firefighters. DETAILED DESCRIPTION
[0045] In order to make the purpose, technical solutions and advantages of one or more embodiments of this specification clearer, the technical solutions of one or more embodiments of this specification will be clearly and completely described below in combination with the specific embodiments of this specification and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of one or more embodiments of this specification.
[0046] It should be understood that although the terms first, second, third, etc. may be used in this application document to describe various information, this information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other.
[0047] In the embodiments of this specification, a method for identifying special populations of a guide robot dog that improves YOLOv8 is provided. In order to improve the recognition speed and accuracy of the guide robot dog for special populations, the technical solution of this application proposes a new target detection and recognition algorithm (SG-YOLOv8) based on the improvement of YOLOv8. This method replaces the loss function CIoU (Complete Intersection over Union) of YOLOv8 with GIoU (Generalized Intersection over Union), and increases the ratio used to adjust the GIoU loss function to 0.75, which takes into account both the overlap of the bounding boxes and the positional relationship between the bounding boxes, thereby more comprehensively evaluating the target detection performance. In addition, the activation function ReLU (Rectified Linear Unit) is replaced with SiLU (Sigmoid-weighted Linear Unit) to improve gradient propagation, nonlinear expressiveness and model stability while maintaining computational efficiency.
[0048] Although object detection based on visual image processing has made substantial progress in many aspects, most detections are still based on simple closed environments or single working environments. Therefore, there is still great research prospect in applying guide robots for object detection and recognition in open, changeable and complex environments. In order to improve the robustness of guide robots in complex environments, the technical solution of this application introduces multi-scale feature fusion in the SG-YOLOv8 algorithm to enhance the model's ability to capture fine-grained features and its attention to key target areas.
[0049] The technical solution of this application uses the SG-YOLOv8 algorithm and combines it with actual application needs to provide an efficient and robust target detection and recognition solution, providing technical support for the application of guide quadruped robots.
[0050] The purpose of the technical solution of this application is to provide a method for identifying special groups of people based on SG-YOLOv8, which is carried on the Unitree Go1 quadruped robot to help visually impaired people travel safely and seek help from special groups in time when they encounter danger. This method replaces the loss function CIoU (Complete Intersection over Union) of YOLOv8 with GIoU (Generalized Intersection over Union), and increases the ratio used to adjust the GIoU loss function to 0.75, which takes into account both the overlap of the bounding boxes and the positional relationship between the bounding boxes, thereby detecting the target more comprehensively. In addition, the activation function ReLU (Rectified Linear Unit) is replaced with SiLU (Sigmoid-weighted Linear Unit) to improve the stability of the model while maintaining computational efficiency.
[0051] To achieve the above objectives, the technical solution of this application provides the following technical solutions:
[0052] Step 1: Create a special population dataset
[0053] This method first uses a self-created dataset, which contains 740 images, divided into 518 training images, 148 test images, and 74 validation images in a 7:2:1 distribution ratio. There are 4 types of targets: family, doctor, firefighter, and traffic police. The robot dog camera took 200 family photos, 215 doctor photos, 175 firefighter photos, and 150 traffic police photos. Figure 1 It shows how to draw a histogram of width and height variables to show the distribution of the data set.
[0054] Step 2: Optimize and improve the YOLOv8 network structure
[0055] The structure of the SG-YOLOv8 network includes the input layer, the backbone network, the neck network, and the head network. The input layer accepts image input, the backbone network extracts image features, the neck network performs multi-scale fusion, and the head network processes the fused features and obtains the predicted category and the position coordinates of the target box. The backbone network mainly includes the C2f module and the SPPF module. The C2f module improves the expressiveness of features by cross-stage feature fusion, and is usually used to extract low-level and mid-level features in images. The SPPF module is a spatial pyramid pooling fusion module. Compared with traditional convolutional neural networks, the SPPF module divides the feature map into sub-regions of different scales, performs pooling operations separately, and splices the pooling results of all sub-regions together. This approach can capture richer spatial information at different scales and improve the network's perception of targets of different sizes.
[0056] The direction of the SG-YOLOv8 method structure is basically the same as that of YOLOv8. Compared with YOLOv8, the specific innovations are as follows: SG-YOLOv8 replaces the original loss function CIoU of YOLOv8 with the GIoU loss function, optimizes the regression of the target box, improves the regression accuracy, and enables the model to locate the target more accurately; at the same time, the activation function ReLU is replaced with SiLU in the Conv layer to enhance the performance of the model.
[0057] (1) Improvements in loss function
[0058] The loss function in the YOLOv8 model is divided into three parts: classification loss, target confidence loss, and positioning loss. The loss of the bounding box uses CIoU Loss. CIoU is an improvement on DIoU. It adds the loss of the predicted box and the loss of length and width, so that the detection is constantly adjusted and closer to the true value. However, CIoU is not sensitive enough when dealing with small targets. It is easy to ignore the shape change of the bounding box, only pay attention to the overlap between position and size, and has poor robustness.
[0059] Therefore, in order to improve the positioning loss function of bounding box regression, the technical solution of this application introduces GloU Loss. GIoU can reflect the distance and overlap between two boxes, and has better robustness to the position offset and shape change of the bounding box. When training the target detection model, using GIOU as the loss function can bring better optimization results. The calculation formula is as follows:
[0060]
[0061]
[0062] Among them, A represents the detection box, B represents the real box, and C represents the smallest enclosing rectangular box.
[0063] Inner-IoU Loss is also used, and the scale factor ratio is set to 0.75 to control the generation of auxiliary bounding boxes of different scales for loss calculation. Inner-IoU is applied to the existing IoU-based bounding box regression loss function, as defined in formula (3), and faster and more effective regression results can be obtained.
[0064] L inner-GIoU =L GIoU +IoU-IoU inner (3)
[0065] (2) Improvements in activation functions
[0066] The activation function used by the YOLOv8 network is the ReLU activation function, and the calculation formula is shown in (4). Where x is a given element:
[0067] ReLU(x)=max(x,0) (4)
[0068] However, compared to the discontinuity of ReLU at zero, the SiLU function is smooth and has continuous derivatives. The use of the SiLU function can better process gradient information, reduce the possibility of gradient disappearance, and improve the training effect of the model. The output range of the ReLU function is [0, +∞), while the output range of the SiLU function is (0, 1], which means that the SiLU function can limit the input activation value to a smaller range, which helps the stability and generalization ability of the model. Therefore, the activation function ReLU is replaced by SiLU, and its calculation formula is (5):
[0069]
[0070] Step 3: Hardware optimization based on Unitree GO 1 quadruped robot dog
[0071] The technical solution of this application uses Unitree GO 1 of Yushu Technology Company as the four-legged robot dog used in the experiment. The robot dog itself contains 5 groups of fisheye binocular cameras, which are located on its head, side and abdomen. It has the industry's cutting-edge visual call capabilities and is suitable as a robot for blind people. Considering the complex road conditions and time issues for blind people to go out, the technical solution of this application chooses to optimize and improve the robot dog. Use a larger capacity battery and reinforce the protection of the robot dog's limb joints. In addition, the robot dog has three Jetson Nano control method boards and two Raspberry controllers built in, with a total of 16-core CPU+GPU (384Core+1.5TFLOPS) computing power to meet the computing power requirements for adding more functions in the future. In addition to the above improvement measures, an external camera IVG-5MP has been added. Taking advantage of its high resolution, high pixel and easy access to the robot dog platform, the robot dog can perform high-speed real-time recognition in scenes where special people are identified, thereby assisting the blind in identifying special people.
[0072] The optimized guide robot dog is as follows Figure 3 As shown, it can also cross various complex terrains to better provide safe travel services for the visually impaired.
[0073] Step 4: Verification experiment of special population identification based on SG-YOLOv8 method
[0074] (1) Evaluation indicators
[0075] The experiment of the technical solution of this application uses several evaluation indicators such as precision, recall, F1-core and average precision to evaluate the accuracy and effectiveness of the model. First, you need to understand the true positive (True Positive, TP): the true category of the sample is a positive example, and the result predicted by the model is also a positive example, and the prediction is correct; true negative (True Negative, TN): the true category of the sample is a negative example, and the model predicts it as a negative example, and the prediction is correct; false positive (False Positive, FP): the true category of the sample is a negative example, but the model predicts it as a positive example, and the prediction is wrong; false negative (False Negative, FN): the true category of the sample is a positive example, but the model predicts it as a negative example, and the prediction is wrong.
[0076] Precision refers to the ratio of the number of correctly assigned positive samples to the total number of assigned positive samples, as shown in formula (6), where TP is the positive detection rate, which refers to the number of correctly detected targets; FP is the false detection rate, which refers to the number of incorrectly detected targets.
[0077]
[0078] The recall rate refers to the ratio of the number of correctly classified positive samples to the total number of positive samples, as shown in formula (7), where FN is the missed detection rate, that is, the number of targets that are not detected.
[0079]
[0080] F1-Score is the harmonic mean of precision and recall. The calculation formula is as follows:
[0081]
[0082] mAP50 (Mean Average Precision at 50) is one of the indicators used to evaluate the performance of target detection methods. It is obtained by calculating the average precision AP (Average Precision) under different confidence thresholds. During the calculation process, the predicted bounding boxes are first sorted according to the confidence. Then, the precision and recall rate at each confidence threshold are calculated, and the precision-recall curve is plotted. Finally, the mAP50 value is obtained by calculating the average value of the area under the curve. The higher the mAP50 value, the better the accuracy and recall rate of the model when predicting bounding boxes. mAP50 is the average precision mean when the IoU threshold is 0.5, and mAP50-95 is the average precision mean under different IoU. The calculation formula is as follows:
[0083]
[0084] Among them, AP is calculated by interpolation method to interpolate the precision-recall curve. interpolation It is the maximum precision value between a given recall value r and the next recall value r' in the curve. i ,r i+1 ,...,r n-1 is the recall value corresponding to the first interpolation point in the precision interpolation segment arranged in ascending order. n in formula (11) represents the number of recognized target categories, which is 4 in this experiment, i represents the i-th recognized category, AP i Represents the average precision for the i-th category.
[0085] (2) Verification of experimental results
[0086] Through comparative experiments, it is found that the improved SG-YOLOv8 model comprehensively considers the indicators of precision and recall rate such as Figure 4: F1 score curve. The F1 score is the harmonic mean of precision and recall, and is a measure of classification effectiveness. The curve can be used to determine whether the model has achieved the expected goal. From the F1 curve, the confidence value parameter node of the optimized precision and recall rate of this model is 0.453. Before 0.453, as the confidence value increases, the F1 score gradually increases, reaching a maximum of 0.93 classification accuracy, and then begins to decline.
[0087] The experimental results are shown in Table 1. Under the same training conditions, the target detection accuracy of the SG-YOLOv8 model is improved by 3.3%, the recall rate is improved by 5.1%, the mAP50 is improved by 1.3%, and the mAP50-95 is improved by 2.8% compared with the original YOLOv8 model.
[0088] Table 1 Comparison of experimental results of different models
[0089]
[0090] in Figure 5 The comparison of four indicators, including precision, recall, mAP50 and Map50-95, between SG-YOLOv8 and YOLOv8 is shown. The curve results show that the overall performance of SG-YOLOv8 is better than that of YOLOv8, which proves the effectiveness of the improved model for guide robot dogs to identify special people.
[0091] Finally, the experiment applied the SG-YOLOv8 network to the quadruped robot dog through embedding and transplantation, and fed back to the computer through the IVG-5MP camera. Figure 6 The target family member portrait identified by the robot dog. Figure 7 The labels and accuracy of identification of special groups of people are displayed, and the identification accuracy generally reaches above 0.9.
[0092] Those skilled in the art can understand that the accompanying drawings are only schematic diagrams of one embodiment, and the modules or processes in the accompanying drawings are not necessarily necessary for implementing the technical solutions of the present application.
[0093] Those skilled in the art can understand that the modules in the device in the embodiment can be distributed in the device in the embodiment according to the description of the embodiment, or can be changed accordingly and located in one or more devices different from the embodiment. The modules in the above embodiment can be combined into one module, or can be further divided into multiple sub-modules.
[0094] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the technical solutions of the present application, rather than to limit them. Although the technical solutions of the present application have been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the technical solutions of the embodiments of the technical solutions of the present application.
Claims
1. A method for identifying special groups of people by improving the guide robot dog of yolov8, characterized in that: The method comprises: Step 1: Create a special population dataset Create the required dataset, which contains 740 images in total. According to the distribution ratio of 7:2:1, it is divided into 518 images for training set, 148 images for test set, and 74 images for validation set. There are 4 types of targets in total: family, doctor, firefighter, and traffic police. The robot dog camera takes 200 photos of family members, 215 photos of doctors are collected from the Internet, 175 photos of firefighters, and 150 photos of traffic police. Step 2: Design the SG-YOLOv8 network based on the optimized YOLOv8 network structure The structure of the SG-YOLOv8 network includes an input layer, a backbone network, a neck network, and a head network; the input layer accepts image input, the backbone network extracts image features, the neck network performs multi-scale fusion, and the head network processes the fused features and obtains the predicted category and the position coordinates of the target box; the backbone network mainly includes the C2f module and the SPPF module; the C2f module improves the expressiveness of features through cross-stage feature fusion and is used to extract low-level and mid-level features in the image, and the SPPF module is a spatial pyramid pooling fusion module. The SPPF module divides the feature map into sub-regions of different scales, performs pooling operations separately, and splices the pooling results of all sub-regions together; SG-YOLOv8 replaces the original loss function CIoU of YOLOv8 with the GIoU loss function; at the same time, the activation function ReLU is replaced with SiLU in the Conv layer to enhance the performance of the model; (1) Improve the loss function The loss function in the YOLOv8 model is divided into three parts: classification loss, target confidence loss, and positioning loss. The loss of the bounding box uses CIoU Loss. CIoU is an improvement on DIoU. It adds the loss of the predicted box and the loss of length and width, so that the detection is constantly adjusted and closer to the true value. However, CIoU is not sensitive enough when dealing with small targets. It is easy to ignore the shape change of the bounding box and only pay attention to the overlap between position and size, and has poor robustness. Therefore, in order to improve the positioning loss function of bounding box regression, the technical solution of this application introduces GloU Loss; GIoU can reflect the distance and overlap between two boxes, and has better robustness to the position offset and shape change of the bounding box. When training the target detection model, using GIOU as the loss function can bring better optimization effect. The calculation formula is as follows: Among them, A represents the detection box, B represents the real box, and C represents the smallest external rectangular box; In addition, Inner-IoU Loss is used, and the scale factor ratio is set to 0.75 to control the generation of auxiliary borders of different scales for loss calculation; Inner-IoU is applied to the existing IoU-based border regression loss function, defined as formula (3), and faster and more effective regression results can be obtained; L inner-GIoU =L GIoU +IoU-IoU inner (3) (2) Improvements in activation functions The activation function used by the YOLOv8 network is replaced by the ReLU activation function with SiLU, and its calculation formula is (5): Step 3: Hardware optimization based on Unitree GO 1 quadruped robot dog Unitree GO 1 from Yushu Technology was selected as the quadruped robot dog used in the experiment. The robot dog itself contains 5 groups of fisheye binocular cameras, which are located on its head, side and abdomen. It has the industry's cutting-edge visual calling ability and is suitable as a robot for blind people. Considering the complex road conditions and time issues when blind people go out, the robot dog is optimized and improved. A larger battery life battery is used and the protection of the robot dog's limb joints is reinforced. In addition, the robot dog has three JetsonNano control method boards and two Raspberry controllers built in, with a total of 16-core CPU+GPU computing power, and an external camera IVG-5MP is added. With its high resolution, high pixel and easy access to the robot dog platform, the robot dog can perform high-speed real-time recognition in scenes where special people are identified, thereby assisting the blind in identifying special people. Step 4: Verification experiment of special population identification based on SG-YOLOv8 method (1) Evaluation indicators The accuracy and effectiveness of the model were evaluated using precision, recall, F1-core and average precision. True positive TP: the true category of the sample is a positive example, and the model predicts it to be a positive example, so the prediction is correct; true negative TN: the true category of the sample is a negative example, and the model predicts it to be a negative example, so the prediction is correct; false positive FP: the true category of the sample is a negative example, but the model predicts it to be a positive example, so the prediction is wrong; false negative FN: the true category of the sample is a positive example, but the model predicts it to be a negative example, so the prediction is wrong; Precision refers to the ratio of the number of correctly assigned positive samples to the total number of assigned positive samples, as shown in formula (6), where TP is the positive detection rate, which refers to the number of correctly detected targets; FP is the false detection rate, which refers to the number of incorrectly detected targets; The recall rate refers to the ratio of the number of correctly classified positive samples to the total number of positive samples, as shown in formula (7), where FN is the missed detection rate, that is, the number of targets that are not detected; F1-Score is the harmonic mean of precision and recall, and is calculated as follows: mAP50 (Mean Average Precision at 50) is one of the indicators used to evaluate the performance of target detection methods. It is obtained by calculating the average precision AP (Average Precision) under different confidence thresholds. In the calculation process, the predicted bounding boxes are first sorted according to the confidence, then the precision and recall rate under each confidence threshold are calculated, and the precision-recall curve is plotted. Finally, the mAP50 value is obtained by calculating the average area under the curve. The higher the mAP50 value, the better the accuracy and recall rate of the model in predicting bounding boxes. mAP50 is the average precision mean when the IoU threshold is 0.5, and mAP50-95 represents the average precision mean under different IoUs. The calculation formula is as follows: Among them, AP is calculated by interpolation method to interpolate the precision-recall curve. The interpolation P in formula (9) interpolation is the maximum precision value between a given recall value r and the next recall value r' in the curve. In formula (10), r1, r2, ..., r i ,r i+1 ,...,r n-1 is the recall value corresponding to the first interpolation point in the precision interpolation segment arranged in ascending order. In formula (11), n represents the number of recognized target categories, i represents the i-th recognized category, and AP i Represents the average accuracy for the i-th category; Step 5: Transplant the trained SG-YOLOv8 network to the quadruped robot dog to identify special people.