Obstacle detection method and readable medium for navigation assistance for visually impaired persons

Abstract

The present invention discloses an obstacle detection method and a readable medium for assisting navigation for the visually impaired, belonging to the technical field of target detection. The method inputs an obstacle image to be detected into an obstacle detection network for regression and classification tasks; a feature weighting module in the obstacle detection network weights the extracted features after the backbone network, and an adaptive bottleneck module is integrated in KAN‑PAFPN, which realizes effective fusion and optimization of features, thereby improving the robustness of the model in complex backgrounds, especially for the detection of occlusions and small targets. The enhanced feature attention head is responsible for the final target classification and regression tasks. By introducing a convolutional block attention module CBAM and a transposed convolution module, the accuracy of the model in target position regression and category judgment is enhanced. The readable medium stores computer instructions, which, when executed, cause the computer to execute the steps of the above-mentioned obstacle detection method for assisting navigation for the visually impaired.

Description

Technical Field

[0001] The present invention belongs to the technical field of target detection, and in particular, relates to an obstacle detection method and a readable medium for assisting navigation of visually impaired persons. Background Art

[0002] Most of the existing navigation aids for the visually impaired use traditional target detection algorithms, such as YOLO and SSD. Although these methods have the advantage of real-time performance, they have great limitations when dealing with small target detection and occlusion problems in complex scenes. Traditional methods have obvious deficiencies in small target detection accuracy, occlusion processing, and background complexity, making it difficult to meet the actual needs of the visually impaired.

[0003] Therefore, there is an urgent need for an improved obstacle detection method that can provide higher detection accuracy and robustness in complex scenarios. Summary of the invention

[0004] The purpose of the present invention is to overcome the problem of inaccurate target detection in complex scenes in the prior art, and to propose an obstacle detection method and a readable medium for navigation assistance for the visually impaired, which can effectively detect obstacles of different sizes, especially solving the problems of insufficient small target detection performance and poor occlusion processing in the prior art. By integrating the feature weighting module and the adaptive bottleneck module, the present invention not only improves the detection accuracy of the model, but also enhances its robustness in complex scenes, significantly improving the safety and reliability of navigation assistance equipment for the visually impaired during navigation.

[0005] The technical solution adopted by the present invention to achieve the above-mentioned purpose is:

[0006] In a first aspect, a method for obstacle detection for navigation assistance for visually impaired persons is provided, comprising:

[0007] Obtain an image of an obstacle to be detected;

[0008] The obstacle image to be detected is input into the obstacle detection network. First, the CSPDarknet-53 backbone network is used to extract multi-scale features of the input obstacle image to be detected. Then, the features extracted by the CSPDarknet-53 backbone network are weighted by the feature weighting module. Then, the Kolmogorov-Arnold Network Path Aggregation FeaturePyramid Network (KAN-PAFPN) is used for feature fusion processing. The KAN-PAFPN is equipped with an adaptive bottleneck module, which is integrated in the feature fusion path of the KAN-PAFPN. Specifically, it is used to replace the original DarknetBottleneck in PAFPN to enhance the feature extraction capability. The adaptive bottleneck module is located in the process of fusion of features from high-level semantic information to low-level spatial information. Its main function is to adaptively adjust features of different scales and levels to improve the feature representation capability of the network. The adaptive bottleneck module receives multi-scale features transmitted from KAN-PAFPN. These features contain multi-scale information from different network layers, including high-resolution detail features and low-resolution global semantic information. After being processed by multi-scale convolution, learnable normalization layer and learnable activation function layer in the adaptive bottleneck module, the enhanced multi-scale features are output, which retain the global information and important details of the original features. At the same time, the adaptive mechanism of the adaptive bottleneck module enhances the feature expression ability, so that the network can better cope with complex detection tasks; finally, the obtained features are subjected to the enhanced feature attention head for regression and classification tasks;

[0009] Get predictions for classification tasks and predictions for regression tasks;

[0010] Obtain an obstacle detection result according to the prediction of the classification task and the prediction of the regression task.

[0011] The feature weighting module is configured as follows:

[0012] First, average pooling (AvgPool) is performed on the features extracted by the CSPDarknet-53 backbone network. The feature dimension extracted by the CSPDarknet-53 backbone network is H×W×C, where H, W, and C are the height, width, and number of channels of the feature, respectively. The spatial dimension of the feature is reduced by taking the average value in the sliding window, so that the feature dimension becomes 1×1×C;

[0013] Then, a convolutional layer with a convolution kernel size of 1×1 is used to adjust the number of channels so that the feature dimension becomes 1×1×1.

[0014] Then, after the LeakyReLU activation function layer and the Hard Sigmoid activation function layer, the values ​​between 0 and 1 are compressed and the feature dimension becomes , is a weighting factor between 0 and 1;

[0015] Next, multiply the output of the Hard Sigmoid activation function layer by the initial feature;

[0016] Finally, a residual structure is introduced to weight the features.

[0017] In the present invention, it should be noted that although the existing PAFPN structure can perform multi-scale feature fusion, it has deficiencies in the feature extraction part, especially in complex scenes. In order to solve this problem, the present invention improves it and proposes a new adaptive bottleneck module (Adaptive Bottleneck Block, ABB) to replace the original DarknetBottleneck in PAFPN. The innovation of the adaptive bottleneck module is that it introduces multi-scale convolution, learnable normalization layer and learnable activation function layer, which further improves the adaptability of the model to different scenes. The design of the adaptive bottleneck module is derived from Kolmogorov-Arnold Networks (KAN). The present invention incorporates some ideas in KAN into the adaptive bottleneck module, and uses learnable normalization layer and learnable activation function layer to automatically adjust network parameters, so that the model can be dynamically adjusted according to different input features. This improvement not only enhances the adaptability of the model to diversified scenes, but also improves the accuracy of obstacle detection in complex environments. Since the present invention incorporates the key design ideas of KAN into the original PAFPN and uses the adaptive bottleneck module in the feature extraction stage, this new structure is named KAN-PAFPN. This module not only inherits the advantages of PAFPN in multi-scale feature fusion, but also further improves the flexibility and robustness of feature extraction through the enhancement of the adaptive bottleneck module.

[0018] The adaptive bottleneck module is configured as follows:

[0019] First, multi-scale convolution is used with kernel sizes of 1×1, 3×3, and 5×5 to capture multi-scale features;

[0020] Then, the captured multi-scale features are merged in the channel dimension, that is, a concat operation is performed;

[0021] Next, after the learnable normalization layer, the calculation formula of the learnable normalization layer is as follows:

[0022] ;

[0023] in represents the result of the normalization operation, represents the normalization layer, Represents the original features obtained by the previous layer of network, and Represents learnable parameters, which are used to scale and translate the normalized features. This means that the normalized features will be multiplied by a learnable scaling factor , and then add a learnable offset This mechanism helps the model better preserve the representation ability of the original features, allowing the network to adjust adaptively during the training process. Usually initialized to 1, is initialized to 0 to ensure that no additional scaling or offset is applied to the input features at the beginning. As training proceeds, the model will learn the optimal and , in order to maximize the performance of the model. is the mean of the feature, is the variance. Both are calculated based on the entire mini-batch, adjusting the features to the mean. is 0, variance Mini-batch is a training method in deep learning. In the training process of neural networks, it is usually necessary to adjust the model parameters through optimization algorithms (such as gradient descent) to minimize the loss function. Specifically, the mean and variance The calculation method is as follows:

[0024] For all sample features in the current mini-batch, the mean It is obtained by summing the features in the channel dimension: , Indicates the number of samples in mini-batch, Indicates the first batch in the mini-batch Characteristics of a sample.

[0025] Mean After finding the variance It is used to measure the dispersion of eigenvalues. The specific calculation formula is:

[0026] ;

[0027] Because the mean and variance They are calculated based on the samples in the mini-batch, so their values ​​will change with the data in each mini-batch. This dynamic calculation method makes The layer can adapt to different input data distributions and ensure the standardization of features, which helps the network to converge faster during training and improve the generalization ability of the network.

[0028] Therefore, during each forward propagation, the mean is recalculated for each mini-batch. and variance , which enables the network to standardize features and stabilize the training process when faced with different input data;

[0029] is a constant, and the present invention takes a very small value 1e -5 , mainly to avoid numerical instability.

[0030] Finally, after the learnable activation function layer, the calculation formula of the learnable activation function is as follows:

[0031] ;

[0032] in represents the output features after being processed by the adaptive bottleneck module, represents a learnable activation function, and is a learnable parameter, It is used to control the overall amplitude of the activation function output, and Used to adjust the contribution of each basis function. Initially, Set to 1, is a random value, which is then continuously updated through the back propagation of the model to ensure that the activation function can be flexibly adjusted according to the distribution of the input data. Representing basis functions, in the present invention, polynomial basis functions and B-spline basis functions are selected to enhance the ability of the activation function to represent nonlinear features, thereby improving the detection effect of targets in complex scenes.

[0033] The enhanced feature attention head is configured as:

[0034] After the features are fused by KAN-PAFPN, they enter two different branches: the regression head Reg-head and the classification head Cls-head;

[0035] In the regression head Reg-head, first, the first convolution module ConvModule is used to adjust the input features to obtain the adjusted preliminary feature representation; then, the transposed convolution module Transposed ConvModule is used to perform upsampling operations to enhance the spatial resolution; then, the second convolution module ConvModule is used to adjust the features back to the original size; then, the first convolutional layer Convolutional Layer is used to generate the output required for the regression task, and finally the IoU union intersection loss is used to evaluate the final output of the regression head Reg-head;

[0036] In the classification head Cls-head, first, the input features are refined by the third convolution module ConvModule to obtain refined high-resolution features; then, the convolution block attention module CBAM uses the attention mechanism to enhance the features in the channel and spatial dimensions, so that the network focuses on the most important features; then, the output of CBAM is processed by the second convolutional layer ConvolutionalLayer to prepare features for classification, and finally, the cross-entropy loss Cross-Entropy Loss is used to evaluate the output of the regression head Reg-head.

[0037] In a second aspect, a readable medium is provided, wherein the readable medium stores computer instructions, and when the computer instructions are executed, the computer performs the steps of the above-mentioned obstacle detection method for navigation assistance for the visually impaired.

[0038] Through the above design scheme, the present invention can bring the following beneficial effects:

[0039] The present invention uses a feature weighting module that assigns different weights to features according to their importance, using higher weights to highlight significant features and weakening less important or redundant features. This method is particularly important for assisting the visually impaired, as accurately identifying and emphasizing key obstacle features can improve the reliability and safety of detection.

[0040] The present invention uses an adaptive bottleneck module, which consists of multiple components, including multi-scale convolution kernels for capturing features of different scales, flexible activation functions for processing different scenes and features, and adaptive normalization methods for adapting to different data distributions. This design improves the detection of small and large objects, enhances the recognition of high-contrast edges and fine textures, and adapts to significant lighting changes or complex backgrounds, thereby improving the overall accuracy and robustness of the system.

[0041] The present invention adopts an enhanced feature attention head, which consists of a classification head and a regression head. The classification head introduces CBAM to apply attention mechanism from channel and spatial dimensions to enhance features, helping the network to focus on the most important features. The regression head enhances the spatial resolution through a transposed convolution module, and then adjusts the features back to their original size through a convolution module, thereby improving the regression task of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present application. The exemplary embodiments of the present invention and their descriptions are used to understand the present invention and do not constitute improper limitations of the present invention. In the drawings:

[0043] Figure 1 The overall network structure diagram of the network on which the obstacle detection method for navigation assistance for the visually impaired proposed in the present invention is based;

[0044] Figure 2 This is the network structure diagram of the feature weighting module;

[0045] Figure 3 This is the network structure diagram of the adaptive bottleneck module;

[0046] Figure 4 Network structure diagram for enhanced feature attention head. DETAILED DESCRIPTION

[0047] In order to make the purpose, features and advantages of the present invention more obvious and easy to understand, the technical scheme in the present invention is clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the present invention is not limited to the following embodiments, and the specific implementation method can be determined according to the technical scheme of the present invention and the actual situation. In order to avoid confusing the essence of the present invention, the known methods, processes and procedures are not described in detail.

[0048] The obstacle detection method for navigation assistance for the visually impaired is implemented as follows:

[0049] like Figure 1As shown in the figure, the obstacle detection method for assisting navigation for the visually impaired first performs multi-scale feature extraction on the input image through the backbone network, then weights the features extracted by the backbone network through the feature weighting block (FWB), and then enters the Kolmogorov-Arnold Network Path Aggregation Feature Pyramid Network (KAN-PAFPN), and performs feature fusion processing in the Kolmogorov-Arnold Network Path Aggregation Feature Pyramid Network (KAN-PAFPN), and the key adaptive bottleneck module (Adaptive Bottleneck The features are further processed by a series of convolutions and learnable parameters in Block (ABB), and finally the regression and classification tasks are performed through the Enhanced Feature Attention Head.

[0050] The backbone network (Backbone) adopts the CSPDarknet-53 backbone network, which is responsible for extracting the preliminary features of the image. The CSPDarknet-53 backbone network is used to extract multi-scale features of the input image to ensure strong adaptability to obstacles of different scales.

[0051] Feature Weighting Block (FWB): This module weights the extracted features after the CSPDarknet-53 backbone network, focusing on the importance of features of different sizes, thereby enhancing the model's ability to recognize various types of obstacles. The specific implementation method of the Feature Weighting Block (FWB) is as follows Figure 2As shown in the figure: the feature dimension initially extracted by the CSPDarknet-53 backbone network is H×W×C, where H, W and C are the height, width and number of channels of the feature respectively. First, average pooling (AvgPool) is performed to reduce the spatial dimension of the feature by taking the average value in the sliding window, and the feature dimension becomes 1×1×C. Then, a convolution layer with a convolution kernel size of 1×1 is used to adjust the number of channels so that the feature dimension becomes 1×1×1. Then, the LeakyReLU activation function layer and the HardSigmoid activation function layer are passed. The LeakyReLU activation function is used to apply a small non-zero slope to negative input values ​​to ensure that all inputs contribute to the output, thereby preventing neurons from losing activity during training. The Hard Sigmoid activation function is a piecewise linear approximation of the standard Sigmoid function. By truncating and scaling the input values, the input values ​​can be quickly mapped to between 0 and 1, which significantly reduces the computational complexity. This is a prior art and will not be described in detail. After that, the values ​​between 0 and 1 will be compressed, and the feature dimension becomes , is a weighting factor between 0 and 1, and then the result is multiplied by the initial feature. The multiplication process is performed element by element, which means that each pixel of the feature of each channel is multiplied by the corresponding weighting factor. This operation can realize adaptive scaling of features, that is, amplification or suppression according to the "importance" of different features. After the weighting operation, the present invention introduces a residual structure to ensure that the original feature information can be partially retained when weighting is performed, so as to avoid the loss of useful features during the processing.

[0052] The specific implementation method of the adaptive bottleneck module (Adaptive Bottleneck Block, ABB): Figure 3 As shown in the figure, first, multi-scale convolution is used with convolution kernel sizes of 1×1, 3×3, and 5×5 to capture multi-scale features. Then, these features are merged in the channel dimension, that is, a concat operation is performed.

[0053] Next, after the learnable normalization layer, the calculation formula of the learnable normalization layer is as follows:

[0054] ;

[0055] in represents the result of the normalization operation, represents the normalization layer, Represents the original features obtained by the previous layer of network, and Represents learnable parameters, which are used to scale and translate the normalized features. This means that the normalized features will be multiplied by a learnable scaling factor , and then add a learnable offset This mechanism helps the model better preserve the representation ability of the original features, allowing the network to adjust adaptively during the training process. Usually initialized to 1, is initialized to 0. The reason for this is to ensure that no additional scaling or offset is applied to the input features at the beginning. As training proceeds, the model will learn the optimal and , in order to maximize the performance of the model. is the mean of the feature, is the variance. Both are calculated based on the entire mini-batch, adjusting the features to the mean. is 0, variance Mini-batch is a training method in deep learning. In the training process of neural networks, it is usually necessary to adjust the model parameters through optimization algorithms (such as gradient descent) to minimize the loss function. Specifically, the mean and variance The calculation method is as follows:

[0056] For all sample features in the current mini-batch, the mean It is obtained by summing the features in the channel dimension: , Indicates the number of samples in mini-batch, Indicates the first batch in the mini-batch Characteristics of a sample.

[0057] Mean After finding the variance It is used to measure the dispersion of eigenvalues. The specific calculation formula is:

[0058] ;

[0059] Because the mean and variance They are calculated based on the samples in the mini-batch, so their values ​​will change with the data in each mini-batch. This dynamic calculation method makes The layer can adapt to different input data distributions and ensure the standardization of features, which helps the network to converge faster during training and improve the generalization ability of the network.

[0060] Therefore, during each forward propagation, the mean is recalculated for each mini-batch. and variance , which enables the network to standardize features and stabilize the training process when faced with different input data;

[0061] is a constant, and the present invention takes a very small value 1e -5 , mainly to avoid numerical instability.

[0062] Finally, after the learnable activation function layer, the calculation formula of the learnable activation function is as follows:

[0063] ;

[0064] in represents the output features after being processed by the adaptive bottleneck module, represents a learnable activation function, and is a learnable parameter, It is used to control the overall amplitude of the activation function output, and Used to adjust the contribution of each basis function. Initially, Set to 1, is a random value, which is then continuously updated through the back propagation of the model to ensure that the activation function can be flexibly adjusted according to the distribution of the input data. Representing basis functions, in the present invention, polynomial basis functions and B-spline basis functions are selected to enhance the ability of the activation function to represent nonlinear features, thereby improving the detection effect of targets in complex scenes.

[0065] Enhanced Feature Attention Head: This module is responsible for the final target classification and regression tasks. By introducing the Convolutional Block Attention Module (CBAM) and the transposed convolution module, the accuracy of the model in target position regression and category judgment is enhanced. The specific implementation method of the Enhanced Feature Attention Head is as follows: Figure 4 As shown in Figure 2, after the features are fused by KAN-PAFPN, they enter two different branches: the regression head (Reg-head) and the classification head (Cls-head). The following is how each branch processes the input features: In the regression head (Reg-head), first, the convolution module (ConvModule) is Figure 4The first convolution module is applied to the input features to adjust the features. The focus of this step is to refine the spatial information and prepare the features for further transformation. It then passes through the transposed convolution module (Transposed ConvModule), which essentially acts as an upsampling operation to enhance the spatial resolution, and then passes through the convolution module (ConvModule). Figure 4 The second convolution module in the image adjusts the features back to their original size, and then passes through the convolutional layer. Figure 4 The first convolutional layer in is used to generate the output required for the regression task, and the final output is evaluated using the IoU (intersection over union) loss, which is a standard loss function for regression tasks in object detection. In the classification head (Cls-head), similar to the regression head, the classification branch first converts the convolution module (ConvModule) Figure 4 The third convolutional module is applied to the input features to refine the features, and then passes through the convolutional block attention module (CBAM), which enhances the features by applying attention mechanisms in the channel and spatial dimensions, helping the network focus on the most important features. Then it passes through another convolutional layer, namely Figure 4 The second convolutional layer in processes the output of CBAM to prepare features for classification. Finally, the classification head output is evaluated using cross-entropy loss, which is a common loss function for multi-class classification tasks.

[0066] The present invention provides a readable medium storing computer instructions, which, when executed, cause the computer to execute the steps of the above-mentioned obstacle detection method for assisting navigation for the visually impaired. It should be understood by those skilled in the art that the embodiments of the present invention may be provided as a method, a system or a computer program product. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes. The scheme in the embodiments of the present invention may be implemented in various computer languages, for example, the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0067] The present invention is compared with several mainstream detectors in the current target detection field. Finally, it is found that the method proposed by the present invention has obvious advantages in detection accuracy. The comparison is shown in Table 1:

[0068] Table 1

[0069]

[0070] Among them, Models represents the model. Target detection is mainly divided into one-stage target detector (One-stage) and two-stage target detector (Two-stage). Faster-rcnn, Cascade-rcnn, Deformable-detr are classic two-stage target detectors, while Yolov5-s, Yolov7-tiny, Yolox-sSSD300, Yolov8-s are one-stage target detectors. Ours is the detection method proposed by the present invention. Bacbone is the backbone network. ResNet-50, YOLOv5CSPDarknet, E-ELAN, YOLOXCSPDarknet, SSDVGG, YOLOv8CSPDarknet, CSPDarknet are the corresponding backbone networks. AP is the detection accuracy of target detection. AP 0.5 is the detection accuracy when the confidence level is 0.5, APS, AP m , A.P. L It is the detection accuracy of small targets, medium targets and large targets. It can be seen from the table that the method of the present invention is superior to the above models in terms of detection accuracy on the selected data set.

Claims

1. An obstacle detection method for assisting navigation for visually impaired persons, comprising obtaining an image of an obstacle to be detected; inputting the image of the obstacle to be detected into an obstacle detection network to perform regression and classification tasks; obtaining a prediction of the classification task and a prediction of the regression task; obtaining an obstacle detection result based on the prediction of the classification task and the prediction of the regression task; characterized in that: In the obstacle detection network, first, the input obstacle image to be detected is subjected to multi-scale feature extraction through the CSPDarknet-53 backbone network; then, the features extracted by the CSPDarknet-53 backbone network are weighted through the feature weighting module; then, KAN-PAFPN is entered for feature fusion processing, wherein KAN-PAFPN introduces an adaptive bottleneck module into the existing PAFPN structure to replace the original DarknetBottleneck in the PAFPN structure. The adaptive bottleneck module is located in the process of fusion of features from high-level semantic information to low-level spatial information. After being processed by multi-scale convolution, learnable normalization layer and learnable activation function layer in the adaptive bottleneck module, the enhanced multi-scale features are output. Finally, the obtained features are subjected to the enhanced feature attention head for regression and classification tasks; the adaptive bottleneck module is configured as follows: First, multi-scale convolution is used with kernel sizes of 1×1, 3×3, and 5×5 to capture multi-scale features; Then, the captured multi-scale features are merged in the channel dimension, that is, concat operation is performed; Next, after the learnable normalization layer, the calculation formula of the learnable normalization layer is as follows: ； in represents the result of the normalization operation, represents the normalization layer, Represents the original features obtained by the previous layer of network, and represents the learnable parameters, is a learnable scaling factor, is the learnable offset, and They are used to scale and translate the normalized features, respectively. During initialization, The initialization value is 1, The initialization value is 0; is the mean of the feature, is the variance. Both are calculated based on the entire mini-batch, adjusting the features to the mean. is 0, variance The distribution of 1, is a constant; Finally, after the learnable activation function layer, the calculation formula of the learnable activation function is as follows: ； in represents the output features after being processed by the adaptive bottleneck module, represents a learnable activation function, and is a learnable parameter, Used to control the overall amplitude of the activation function output, Used to adjust the contribution of each basis function; Represents basis functions.

2. The obstacle detection method for navigation assistance for the visually impaired according to claim 1, characterized in that: The feature weighting module is configured as follows: First, average pooling is performed on the features extracted by the CSPDarknet-53 backbone network. The feature dimension extracted by the CSPDarknet-53 backbone network is H×W×C, where H, W, and C are the height, width, and number of channels of the feature, respectively. The spatial dimension of the feature is reduced by taking the average value in the sliding window, so that the feature dimension becomes 1×1×C. Then, a convolutional layer with a convolution kernel size of 1×1 is used to adjust the number of channels so that the feature dimension becomes 1×1×1. Then, after the LeakyReLU activation function layer and the Hard Sigmoid activation function layer, the values ​​between 0 and 1 are compressed and the feature dimension becomes , is a weighting factor between 0 and 1; Next, the output of the Hard Sigmoid activation function layer is multiplied by the initial feature, and the multiplication process is performed element by element, that is, each pixel of the feature of each channel is multiplied by the corresponding weighting factor ; Finally, a residual structure is introduced to weight the features.

3. The obstacle detection method for navigation assistance for the visually impaired according to claim 1, characterized in that: The enhanced feature attention head is configured as: After the features are fused by KAN-PAFPN, they enter two different branches: the regression head Reg-head and the classification head Cls-head; In the regression head Reg-head, first, the first convolution module ConvModule is used to adjust the input features to obtain the adjusted preliminary feature representation; then, the transposed convolution module Transposed ConvModule is used to perform upsampling operations to enhance the spatial resolution; then, the second convolution module ConvModule is used to adjust the features back to the original size; then, the first convolutional layer Convolutional Layer is used to generate the output required for the regression task, and finally the IoU union intersection loss is used to evaluate the final output of the regression head Reg-head; In the classification head Cls-head, first, the input features are refined by the third convolution module ConvModule to obtain refined high-resolution features; then, the convolution block attention module CBAM is used to enhance the features in the channel and spatial dimensions using the attention mechanism; then, the output of CBAM is processed by the second convolutional layer ConvolutionalLayer to prepare features for classification, and finally, the cross-entropy loss Cross-Entropy Loss is used to evaluate the output of the classification head Cls-head.

4. A readable medium, characterized in that: The readable medium stores computer executable instructions, and the computer executable instructions are used to: execute the steps of the obstacle detection method for navigation assistance for the visually impaired according to any one of claims 1 to 3.

Images