Blind person travel auxiliary equipment based on deep learning
By improving the combination of YOLOv8n, YOLOv8n-seg network models and edge computing devices, the problems of high cost and incomplete functions of blind travel auxiliary equipment are solved, more accurate environmental perception and navigation are achieved, and the quality of life and independence of blind people are improved.
Patent Information
- Application Number
- CN202411001583.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-25
- Publication Date
- 2024-11-12
Smart Images

Figure CN118942065A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of deep learning and computer vision technology, and specifically relates to a travel assistance device for the blind based on deep learning. Background Art
[0002] According to the World Health Organization, 2.2 billion people worldwide suffer from visual impairment, which makes it difficult for them to carry out normal daily activities. It is estimated that the number of people with visual impairment will stabilize at 11 billion by 2100. Due to the lack of visual perception, blind people often face problems such as getting lost and having a poor sense of direction when traveling, especially in complex traffic environments and unfamiliar areas, where they often feel uneasy and fearful.
[0003] One of the traditional ways to assist the blind is to use guide dogs to help them avoid obstacles. However, due to the high cost of animal training and pairing, the number of guide dogs that can be put into use is limited, and they have certain limitations in certain actions and places, so it is particularly important to develop suitable assistive devices. Auxiliary tools such as guide sticks can also help the blind determine the direction of travel, but their design structure is simple, the help they provide is limited, and they can only perceive the closer road conditions and have difficulty coping with the complex and changing external environment. Therefore, in order to improve the quality of life and autonomy of the blind, the research and development of travel assistance equipment for the blind has become an important topic.
[0004] In recent years, the development of science and technology has brought a more intelligent and convenient experience to assistive devices for the blind. Many new assistive tools and electronic assistive systems have appeared on the market. These new assistive tools have greatly improved the quality of life of the blind, but they also have disadvantages such as high cost, large errors, incomplete functions, and low level of intelligent interaction, making them difficult to be widely popularized. With the rapid development of artificial intelligence, deep learning methods can extract more abstract and deeper features, and have become one of the most promising technologies for promoting the development of travel assistive devices for the blind. Summary of the invention
[0005] In response to the above-mentioned research problems, the purpose of the present invention is to propose a deep learning-based travel assistance device for the blind to provide accurate environmental perception for the blind, thereby improving their quality of life and independence.
[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is as follows:
[0007] A travel assistance system for the blind based on deep learning, comprising the following steps:
[0008] S1: Data sets are collected through major websites such as Kaggle, Hugging Face, Roboflow, and web crawler technology, including three categories of data sets for road segmentation: blind paths, crosswalks, and ordinary roads, and 12 categories of data sets for obstacle detection: pedestrians, bicycles, motorcycles, cars, buses, trash cans, stone piers, fire hydrants, cones, puddles, green lights at pedestrian intersections, and red lights at pedestrian intersections. The data sets are expanded and the expanded data are sampled for data enhancement.
[0009] Furthermore, the specific steps of S1 are:
[0010] S1.1: Road segmentation and obstacle detection datasets are produced by searching major dataset websites and crawling the web, as well as using the D435i depth camera for field collection. The collected datasets include different angles, different environments, and different lighting conditions.
[0011] S1.2: Collect pictures and videos of roads and various obstacles on the roads through multiple scenes, and capture pictures from the videos frame by frame. Then use the semi-automatic labeling software X-Anylabeling to manually label the pictures, and merge them with the data sets found on the Internet to form a new segmentation and detection data set. Finally, the data set is further expanded through rotation, scaling, translation, color transformation, contrast transformation and other processing.
[0012] S1.3: To address the problem of low visibility of obstacles under low light conditions, the SCI low-light image enhancement algorithm is first used to enhance the input image, which improves the image brightness and contrast while retaining the image details. The enhanced image is then sent to the model for training, which helps the model better learn the characteristics of targets in dark environments and better ensure the safety of blind people's travel.
[0013] The above steps can solve the problem of scarce road segmentation and obstacle detection datasets.
[0014] S2: Improve the YOLOv8n and YOLOv8n-seg network models. First, use the improved state space model SSM as the backbone feature extraction network, slightly increase the number of parameters while improving the model accuracy. Secondly, use the fourth generation DCNv4 with variable convolution to improve the C2f module in the neck, reducing nearly 1 / 3 of the model parameters while maintaining the model accuracy. Finally, integrate the linear attention mechanism combined with the Mamba architecture after the three detection heads and segmentation heads to provide the necessary position information for model training, improving the inference speed and model performance.
[0015] Furthermore, the specific steps of S2 are:
[0016] S2.1: The backbone feature extraction network of YOLOv8n and YOLOv8n-seg is replaced with the improved SSM model for feature extraction. Unlike the original backbone network which uses 3x3 convolution with a stride of 2, the improved backbone network downsamples through 4x compressed point-wise convolution, thereby retaining the feature maps selected from the previous layer. In addition, the ODSS Block is proposed, in which LSBlock effectively extracts the local spatial information of the input feature map to make up for the local modeling ability of SSM. The RG Block combines gated aggregation with effective convolution and residual connectivity to effectively capture local dependencies and enhance the robustness of the model.
[0017] S2.2: Add the fourth generation of deformable convolution DCNv4 to the neck C2f module to more effectively handle deformation and dynamic changes in the image. Since each convolution kernel of DCNv4 can adaptively adjust its aggregation weight, the response to each pixel is no longer limited by a fixed range, thereby improving the performance of the model when processing objects of different resolutions and scales. In addition, the convolution has a lightweight effect. Combining a DCNv4 with C2f reduces the number of parameters by about 15W parameters, which is very suitable for model deployment.
[0018] S2.3: Combined with the attention mechanism of the Mamba architecture, MLLA, the model performance is improved by adding the MLLA attention module after the three-layer segmentation head and the detection head. Specifically, MLLA mainly integrates the two key factors of "forget gate" and "module design" in Mamba, and replaces the forget gate with position encoding, thereby providing the necessary position information for model training while maintaining parallel computing and fast reasoning. Experiments have shown that this attention mechanism has the effect of reducing the amount of model calculation and parameters, which is conducive to lightweight design of the model.
[0019] Through the above steps, we can obtain a road segmentation and obstacle detection algorithm with a small number of model parameters and good accuracy and speed.
[0020] S3: Based on the improved YOLOv8n and YOLOv8n-seg network models, the network models are deployed to edge computing devices and combined with intelligent auxiliary devices to better assist blind people in traveling.
[0021] Furthermore, the specific steps of S3 are:
[0022] S3.1: Deploy the optimized YOLOv8n and YOLOv8n-seg models to the edge computing device - NVIDIA Jetson nano, configure the deep learning environment, connect the edge device to the D435i depth camera through the USB serial port, and display the road segmentation and obstacle detection effect interface on the display screen in real time through the HDMI / DP serial port.
[0023] S3.2: The data collected by the depth camera is calculated by the computing device, and the output result is sent to the Arduino microcontroller module through the UART interface.
[0024] S3.3: The Arduino microcontroller module is connected to two SG90 servos through the GPIO pins. It vibrates left and right to warn the blind to walk along the designated road section and avoid obstacles on the road, ensuring the safety of the blind.
[0025] Compared with the prior art, the beneficial effects of the present invention are: the data set is collected with the help of the Internet and web crawler technology, and the data is expanded through field photography, manual annotation and some data enhancement methods. In view of the problems that the YOLOv8n and YOLOv8n-seg network models have large parameters, complex calculations, slow detection speed, and are not easy to deploy in hardware, and cannot meet daily needs, an improved segmentation and detection model is proposed to reduce the model parameters and improve the model detection speed. The optimized network model is deployed in hardware and applied in the real life of the blind. A servo is designed to assist the blind in traveling, and the offset signal of the person's position on the road and the distance signal between the person and the obstacle are converted into an Arduino control electrical signal, and then the servo vibration is used to warn the blind to walk in a fixed road section and avoid obstacles. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] In order to more clearly illustrate the embodiments of the present invention or the technical solutions 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 present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0027] Figure 1 It is a structural block diagram of the present invention;
[0028] Figure 2 It is a structural diagram of the improved YOLOv8n and YOLOv8n-seg algorithms in the present invention;
[0029] Figure 3 It is the module structure diagram of ODSS Block, LS Block and RG Block in the present invention;
[0030] Figure 4 This is a structural diagram of the C2f-DCNv4 module in the present invention;
[0031] Figure 5 This is a structural diagram of the MLLA attention module in the present invention;
[0032] Figure 6 This is a control system diagram of the travel assistance device for the blind in the present invention;
[0033] Figure 7 It is the hardware control system integration diagram of the present invention;
[0034] Figure 8 It is the overall flow chart of the software in the present invention. DETAILED DESCRIPTION
[0035] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention is described below by the specific embodiments shown in the accompanying drawings. However, it should be understood that these descriptions are only exemplary and are not intended to limit the scope of the present invention. The structures, proportions, sizes, etc. illustrated in the drawings of this specification are only used to match the contents disclosed in the specification for people familiar with this technology to understand and read, and are not used to limit the limiting conditions that the present invention can be implemented, so they have no technical substantive significance. Any structural modification, change in proportional relationship or adjustment of size, without affecting the effects that the present invention can produce and the purposes that can be achieved, should still fall within the scope of the technical content disclosed by the present invention. In addition, in the following description, the description of known structures and technologies is omitted to avoid unnecessary confusion of the concept of the present invention.
[0036] It should also be noted that, in order to avoid obscuring the present invention due to unnecessary details, only structures and / or processing steps closely related to the solutions according to the present invention are shown in the accompanying drawings, while other details that are not closely related to the present invention are omitted.
[0037] The specific implementation method of the present invention adopts the following technical solutions: including data collection, data preprocessing, road segmentation and obstacle detection model construction, experimental result visualization, and blind travel assistance equipment control system construction. The data preprocessing includes field shooting for data expansion, processing the data set format, and data enhancement by means of random rotation, cropping, color transformation, contrast transformation, etc. The experimental results are visualized using PyQt to design a 2x2 display interface, which displays the original image, depth image, road segmentation result map, and obstacle detection result map on the display screen from the upper left corner to the lower right corner. The blind travel assistance equipment includes an edge computing device, a microcontroller, and a servo mechanical structure, such as Figure 1 shown.
[0038] Data collection:
[0039] Data collection determines the application scope and function of the database. In order to better simulate the real environment of blind people's travel, the road segmentation dataset in this invention uses three types of roads that blind people walk on in their daily travel, namely, blind paths, crosswalks, and ordinary roads. The blind paths are a public dataset - TP-Dataset, the crosswalks are a dataset on the Kaggle platform, and the ordinary roads are all-weather road image segmentation datasets - UAS. The obstacle detection dataset covers most of the obstacles that blind people encounter on the road in their daily travel, including pedestrians, bicycles, motorcycles, cars, buses, trash cans, stone piers, fire hydrants, cones, puddles, green lights at pedestrian intersections, and red lights at pedestrian intersections. The dataset is collected from major websites such as HuggingFace and Roboflow and obtained using web crawler technology.
[0040] Data preprocessing:
[0041] Data preprocessing mainly includes data expansion, data enhancement, video frame-by-frame image extraction, image format adjustment, and manual image labeling. First, use a camera to record nearby scenes, then use OpenCV-Python to extract the captured video frame by frame, and use some data enhancement methods to achieve data expansion, and then manually label the data set. In addition, since some road segmentation data sets downloaded from the Internet are mask types, OpenCV-Python is needed to perform edge detection and convert mask labels into coordinate labels suitable for YOLOv8 training.
[0042] Road segmentation and obstacle detection model construction:
[0043] A neural network model is proposed to correct the direction of blind people's travel and help them avoid obstacles in their daily life accurately and in real time. Considering the complexity of obstacles in the blind's travel process and the interference of a large amount of redundant information in the real environment, the improved YOLOv8n and YOLOv8n-seg algorithm structures are designed, such as Figure 2 shown.
[0044] First, the backbone feature extraction network of YOLOv8n and YOLOv8n-seg is replaced with the improved SSM model for feature extraction. Unlike the original backbone network that uses 3x3 convolution with a stride of 2, the improved backbone network downsamples through 4x compressed point-wise convolution, thereby retaining the feature map selected from the previous layer. In addition, the ODSS Block is proposed, in which LSBlock effectively extracts the local spatial information of the input feature map to make up for the local modeling ability of SSM. RG Block combines gated aggregation with effective convolution and residual connectivity to effectively capture local dependencies and enhance the robustness of the model. The module structures of ODSS Block, LS Block, and RG Block are as follows: Figure 3 shown.
[0045] ODSS Block is the core module of the backbone feature extraction network. After a series of processing at the input stage, the network can learn deeper and richer feature representations, while maintaining the efficiency and stability of the training and reasoning process through batch normalization. As shown in formula (1).
[0046]
[0047] Where: is the activation function SiLU. The layer normalization and residual link design of ODSS Block draws on the Transformer Blocks style architecture, which allows the model to flow efficiently in the presence of deep stacking. Training is performed in the presence of deep stacking. The calculation formula is as follows.
[0048] Z l-1 =SS2D(LN(LS(Z l-2 )))+Z l-2 (2)
[0049] Z l =RG(LN(Z l-1 ))+Z l-1 (3)
[0050] Where Z l-3 and Z l Represent the input and output features respectively, LS and RG represent LS Block and RG Block respectively, Z l -1 Indicates the intermediate state after 2D-Selective-Scan.
[0051] The fourth generation of deformable convolution DCNv4 is added to the neck C2f module to more effectively handle deformation and dynamic changes in the image. Since each convolution kernel of DCNv4 can adaptively adjust its aggregation weight, the response to each pixel is no longer limited by a fixed range, thereby improving the performance of the model when processing objects of different resolutions and scales. The overall structure of C2f-DCNv4 is as follows: Figure 4 shown.
[0052] DCNv4 is an improved convolutional network operator designed to improve the efficiency and effectiveness of computer vision applications. It improves on the previous version DCNv3 in two main aspects: first, it removes the softmax normalization in spatial aggregation, which enhances its dynamic characteristics and expressiveness; second, it optimizes the memory access process to reduce redundant operations and thus speed up processing. These improvements enable DCNV4 to show faster convergence and higher processing efficiency in various visual tasks, such as image classification and generation.
[0053] Combined with the attention mechanism of the Mamba architecture, MLLA, the model performance is improved by adding the MLLA attention module after the three-layer segmentation head and the detection head. Specifically, MLLA mainly integrates the two key factors of "forget gate" and "module design" in Mamba, and uses position encoding to replace the forget gate, thereby providing the necessary position information for model training while maintaining parallel computing and fast reasoning. Experiments have shown that this attention mechanism has the effect of reducing the amount of model calculation and parameters, which is conducive to lightweight design of the model. MLLA attention module such as Figure 5 shown.
[0054] Visualization of experimental results:
[0055] A 2x2 display interface was designed using PyQt, which displays the original image, the depth image collected by the D435i depth camera, the road segmentation result map, and the obstacle detection result map from the upper left corner to the lower right corner. The road segmentation result map displays the mask information of the road segmentation and the angle information of the road centerline offset in real time, and the obstacle detection result map displays the category and confidence information of the obstacle detection, as well as the obstacle distance information and angle information in real time.
[0056] Construction of control system for blind travel assistance equipment:
[0057] In the design of the control system for the travel assistance device for the blind, after comparing various AI computing devices and combining the functional requirements of real-time and accuracy, the image processing and network computing parts of the present invention use the edge computing device - NVIDIA Jetson nano. Figure 6As shown in the figure. First, the depth camera D435i is used to obtain image information in real time. Then, the road segmentation and obstacle detection algorithms are run simultaneously. The network output is sent to the Arduino microcontroller through the UART interface. The Arduino microcontroller is connected to two SG90 servos through GPIO pins, and vibrates respectively under the lower part of the human ear to remind the blind to adjust the driving route and avoid obstacles.
[0058] Principle analysis and hardware integration:
[0059] Principle analysis of walking direction adjustment and obstacle avoidance:
[0060] After the road is segmented, the road is white pixels at this time, and other areas are black pixels. Orange color blocks are covered in the areas corresponding to the white pixels on the original image to represent the segmented road. By setting two reference lines, the boundaries and center point coordinates of the white area are determined on the reference lines. According to the slope K of the line connecting the center points, when the blind person walks on the road, the extension direction of the road can be determined: when K < -10°, the road extends to the left, and the blind person needs to be reminded to adjust to the left front; when -10° < K < 10°, the road extends straight ahead, and the blind person does not need to be reminded to make adjustments; when K > 10°, the road extends to the right, and the blind person needs to be reminded to adjust to the right front. At the same time, the center point at the bottom of the road segmentation bounding box is connected to the center of the object detection bounding box to obtain the slope K1, and the direction of the obstacle in front is judged: when 75° < K1 < 90°, an obstacle appears in the right front; when 90° < K1 < 105°, an obstacle appears in the left front, and the blind person is warned to avoid it in combination with the distance information.
[0061] Hardware integration:
[0062] The hardware integration of the blind travel assistance device mainly consists of an edge computing device for processing images and deep learning - the NVIDIA Jetson nano embedded platform, the D435i depth camera, the Arduino microcontroller, two SG90 servos, a 13.3 - inch touch liquid crystal display, and a 24V lithium battery. Its hardware control system integration diagram is as Figure 7 shown.
[0063] AI computing embedded platform Jetson nano:
[0064] NVIDIA Jetson nano is a small, cost-effective embedded platform designed for artificial intelligence (AI) and IoT applications, equipped with a 128-core Maxwell architecture GPU, a quad-core ARM Cortex-A57 processor, and 4GB of memory. It supports a wide range of development tools and AI frameworks, and is suitable for scenarios such as robots, smart monitoring, and IoT devices. Jetson Nano is easy to use, provides powerful computing power with the JetPack SDK, and has an active developer community, making it an ideal choice for getting started and developing AI and embedded system applications.
[0065] Arduino Uno Microcontroller Module:
[0066] Arduino Uno is a development board based on the ATmega328P microcontroller, with 14 digital I / O pins, 6 analog input pins and 32KB flash memory, an operating voltage of 5V and a clock speed of 16MHz. It is easy to use, suitable for programming and basic electronics teaching, rapid prototyping and various DIY projects, supports power supply via USB or DC power supply, and has rich community resources and open source hardware design.
[0067] SG90 servo:
[0068] There are three wires on the SG90 servo, namely VCC, GND, and signal line. The control signal of SG90 generally requires a PWM signal with a period of 20ms. In addition, VCC and GND need to be connected to the driver to power the servo, and they need to share the same ground with the development board. In this specific implementation, the 5V stable voltage provided by Arduino Uno is used to connect the SG90 to power it, and PB1 and PB2 generate PWM signals to control the rotation of the servo.
[0069] Software design and process:
[0070] The software part of the blind travel assistance device mainly includes the network design module and the steering gear control module. The overall process is as follows: Figure 8 shown.
[0071] Taking into account the program development environment and model buildability, the network model uses YOLOv8n and YOLOv8n-seg. Based on the Pytorch deep learning framework, the GPU version of Pytorch1.13.0 and the Torchvision version of 0.14.0 are selected. The model is also lightweighted to facilitate its deployment on edge computing devices.
[0072] The servo control part uses geometric knowledge in mathematics and combines distance information to take angle and distance as the judgment conditions for the direction of travel and obstacles, converts them into Arduino control electrical signals, and then uses the servo vibration to warn the blind to walk in a fixed road section and avoid obstacles.
[0073] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, the embodiments should be considered exemplary and non-restrictive in all respects, and the scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims be included in the present invention.
[0074] In addition, the above is only one embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. A travel assistance device for the blind based on deep learning, It is characterized in that It includes the following steps: S1: We collected data sets through major websites such as Kaggle, Hugging Face, and Roboflow, as well as web crawler technology, including three categories of data sets for road segmentation: blind paths, crosswalks, and ordinary roads, and 12 categories of data sets for obstacle detection: pedestrians, bicycles, motorcycles, cars, buses, trash cans, stone piers, fire hydrants, cones, puddles, green lights at pedestrian intersections, and red lights at pedestrian intersections. We also expanded the data sets and performed sampling data enhancement on the expanded data. S2: Improve the YOLOv8n and YOLOv8n-seg network models. First, use the improved state space model SSM as the backbone feature extraction network, slightly increase the number of parameters while improving the model accuracy. Secondly, use the fourth generation DCNv4 with variable convolution to improve the C2f module in the neck, reducing nearly 1 / 3 of the model parameters while maintaining the model accuracy. Finally, integrate the linear attention mechanism combined with the Mamba architecture after the three detection heads and segmentation heads to provide the necessary position information for model training, improve the reasoning speed and model performance; S3: Based on the improved YOLOv8n and YOLOv8n-seg network models, the network models are deployed to edge computing devices and combined with intelligent auxiliary devices to better assist blind people in traveling.
2. A deep learning-based travel assistance device for the blind according to claim 1, It is characterized in that The specific steps of step S1 are: S1.1: Road segmentation and obstacle detection datasets are produced by searching major dataset websites and crawling the web, as well as using the D435i depth camera for field collection. The collected datasets include different angles, different environments, and different lighting conditions; S1.2: Collect pictures and videos of roads and various obstacles on the roads through multiple scenes, and capture pictures from the videos frame by frame. Then use the semi-automatic labeling software X-Anylabeling to manually label the pictures, and merge them with the data sets found on the Internet to form a new segmentation and detection data set. Finally, the data set is further expanded through rotation, scaling, translation, color transformation, contrast transformation and other processing.
3. A deep learning-based travel assistance device for the blind according to claim 1, It is characterized in that The specific steps of step S2 are: S2.1: Replace the backbone feature extraction network of YOLOv8n and YOLOv8n-seg with the improved SSM model for feature extraction. Unlike the original backbone network that uses 3x3 convolution with a stride of 2, the improved backbone network downsamples through 4x compressed point-wise convolution, thereby retaining the feature map selected from the previous layer. In addition, the ODSS Block is proposed, in which LSBlock effectively extracts the local spatial information of the input feature map to make up for the local modeling ability of SSM. RG Block combines gated aggregation with effective convolution and residual connectivity to effectively capture local dependencies and enhance the robustness of the model; S2.2: The deformable convolution fourth generation DCNv4 is added to the neck C2f module to more effectively handle deformation and dynamic changes in the image. Since each convolution kernel of DCNv4 can adaptively adjust its aggregation weight, the response to each pixel is no longer limited to a fixed range, thereby improving the performance of the model when processing objects of different resolutions and scales. In addition, this convolution has a lightweight effect. Combining a DCNv4 with C2f reduces the number of parameters by about 15W parameters, which is very suitable for model deployment. S2.3: Combined with the attention mechanism of the Mamba architecture, MLLA, the model performance is improved by adding the MLLA attention module after the three-layer segmentation head and the detection head. Specifically, MLLA mainly integrates the two key factors of "forget gate" and "module design" in Mamba, and replaces the forget gate with position encoding, thereby providing the necessary position information for model training while maintaining parallel computing and fast reasoning. Experiments have shown that this attention mechanism can reduce the amount of model calculations and parameters, which is conducive to lightweight design of the model.
4. The deep learning-based travel assistance device for the blind according to claim 1, It is characterized in that The specific steps of step S3 are: S3.1: Deploy the optimized YOLOv8n and YOLOv8n-seg models to the edge computing device - NVIDIA Jetsonnano, configure the deep learning environment, connect the edge device to the D435i depth camera through the USB serial port, and display the road segmentation and obstacle detection effect interface on the display screen in real time through the HDMI / DP serial port; S3.2: The data collected by the depth camera is calculated by the computing device, and the output result is sent to the Arduino microcontroller module through the UART interface; S3.3: The Arduino microcontroller module is connected to two SG90 servos through the GPIO pins. It vibrates left and right to warn the blind to walk along the designated road section and avoid obstacles on the road, ensuring the safety of the blind.
Citation Information
Cited By
Small target detection method and device, equipment and storage medium
CN119851149A
Small target detection method, device, equipment and storage medium
CN119851149B
Lightweight drainage pipeline defect detection method based on improved YOLOv8
CN120355997A
