Visual SLAM method based on deep learning in dynamic environment

Abstract

The present invention discloses a visual SLAM method based on deep learning in a dynamic environment, the method comprising the following steps: using a GCNv2 network to extract feature points; using a method combining a semantic segmentation network and multi-view geometry to detect dynamic feature points and remove dynamic feature points, wherein an improved DeepLabV3plus network is used to perform semantic segmentation on RGB images, and the pixels corresponding to the prior dynamic object person are used as the dynamic area; using multi-view geometry to detect other dynamic objects other than people; then after tracking, local map optimization, closed-loop detection, finally using point cloud information and semantic information to combine to construct a semantic point cloud map and a semantic octree map. The present invention has a semantic segmentation model and a multi-view geometry method, which enables the visual SLAM algorithm to effectively reduce the impact of dynamic objects and improve the positioning accuracy of the system.

Description

Technical Field

[0001] The present invention relates to the fields of robotics, computer vision, and artificial intelligence technology, and in particular to a visual SLAM method based on deep learning in a dynamic environment. Background Art

[0002] Simultaneous Localization and Mapping (SLAM) enables autonomous positioning and mapping in unknown environments. It is a foundational technology for applications such as industrial robotics, autonomous navigation, and autonomous driving. Currently, the sensors used in SLAM are primarily lidar, cameras, and inertial measurement units (IMUs). In recent years, visual SLAM using cameras as sensors has become a major research area in the field.

[0003] Most current research on visual SLAM algorithms assumes a static and simple environment. However, real-world environments often contain moving objects, which can affect positioning accuracy and, in severe cases, even lead to positioning failure. This results in poor robustness of visual SLAM in dynamic scenes, with positioning and mapping accuracy susceptible to interference from dynamic objects. Furthermore, most SLAM algorithms lack semantic information, resulting in unintuitive maps and poor application in navigation. With the advancement of deep learning and the continuous improvement of computer performance, semantic segmentation has provided a new technical approach to addressing the poor robustness of visual SLAM algorithms in dynamic environments. Combining semantic segmentation with SLAM can effectively eliminate dynamic points and construct maps that incorporate semantic information. In the prior art, Yu et al. proposed the DS-SLAM algorithm, which utilizes a semantic segmentation network (SegNet) and motion consistency detection to eliminate dynamic feature points. Bescos et al. employed the instance segmentation network (Mask R-CNN) to segment dynamic objects a priori and detect potential moving objects using a multi-view integration approach. Zhong et al. employed the object detection method (SSD) to detect dynamic objects in frames. They propagated motion probabilities for dynamic feature points using feature matching and an extended region of influence, eliminating dynamic points from the tracking process. Xiao et al. also used SSD to detect a priori moving objects and employed a selective tracking algorithm in the tracking thread to process feature points on dynamic targets. However, this method suffers from the following drawbacks: Because visual SLAM cannot distinguish between dynamic and static feature points in dynamic scenes, dynamic feature points are mistakenly identified as static feature points when estimating the camera pose. This results in a significant deviation between the estimated camera pose and the true pose, reducing the positioning and mapping accuracy of the visual SLAM system. Furthermore, most SLAM algorithms lack semantic information, making the constructed maps unintuitive and ineffective for navigation applications. Summary of the Invention

[0004] The purpose of the present invention is to overcome the above shortcomings and propose a visual SLAM method based on deep learning in dynamic environments that can locate and map in dynamic scenes without being disturbed by dynamic objects, has semantics, and has high accuracy, robustness and good real-time performance.

[0005] The present invention provides a visual SLAM method based on deep learning in a dynamic environment, the method comprising the following steps:

[0006] S1. Feature point extraction: The RGB image acquired by the camera is used as input for feature extraction. The features are then homogenized and non-maximum suppressed to obtain evenly distributed feature points. The feature points extracted in the current frame include key points and descriptors.

[0007] The feature extraction uses the GCNv2 network, first adjusting the RGB image to 320×240 size, and then performing feature extraction.

[0008] S2. Dynamic point removal: Dynamic feature points are detected and removed using a combination of semantic segmentation networks and multi-view geometry. The specific steps are as follows:

[0009] S2.1 performs semantic segmentation on the image using a semantic segmentation network: the semantic segmentation network is an improved DeepLabV3plus network. The improved DeepLabV3plus network is based on the traditional DeepLabV3plus network, adopts MobileNet V2 as the backbone network, uses transfer learning to load the pre-trained model, replaces the standard convolution with depthwise separable convolution in the atrous spatial pyramid pooling module ASPP, and references the SE attention mechanism to increase the weights of feature channels with good network performance and suppress the weights of feature channels that cause poor network performance;

[0010] The improved DeepLabV3plus network is used to perform semantic segmentation on the RGB image to obtain pixel-level semantic segmentation results. The pixels corresponding to the prior dynamic object person are used as the dynamic area. The specific steps are as follows:

[0011] The input image is extracted using the lightweight backbone network MobilenetV2, resulting in two output values: a feature map containing high-level semantic information and a feature map containing low-level features. The feature map containing high-level semantic information first passes through the Atrous Spatial Pyramid Pooling module (ASPP), then uses the SE attention mechanism to improve training efficiency, and then uses 1x1 convolution to adjust the number of channels, and performs four upsampling operations. The feature map containing low-level features uses 1x1 convolution to adjust the number of channels, and then stacks it with the feature map of high-level semantic information. Finally, after channel adjustment using 3x3 convolution, it is four times upsampled to obtain the final pixel-level semantic segmentation result, and the pixels corresponding to the prior dynamic object people are used as the dynamic area.

[0012] Among them, the atrous spatial pyramid pooling module ASPP performs depth-separable atrous convolution and global average pooling operations at different resolutions to capture rich contextual information. The atrous convolution algorithm can be expressed as:

[0013]

[0014] Among them, y[i] is the feature map of the input hole convolution size i, w[k] represents the size of the convolution kernel k×k, w is the convolution kernel, k is the size of the convolution kernel, r is the expansion rate, and x is the output feature map; among them, the expansion rate is the stride size of the convolution kernel when working, and the size of the receptive field changes with the expansion rate.

[0015] S2.2 uses multi-view geometry to detect dynamic objects other than people: determine the key frame in the image, project the key points in the key frame to the current frame, and use the parallax angle and depth value change to distinguish the target object into dynamic and static areas; the process is as follows: first, project the key point p in the key frame to the current frame to obtain the projected point p′ and the projected depth D proj , then calculate the disparity angle α between the back projection of p and p′, and at the same time, calculate the depth value difference ΔD; use the disparity angle α and the depth value difference ΔD to determine whether the current feature point is a dynamic feature point, so that the dynamic feature points of other dynamic objects detected except people are regarded as dynamic areas;

[0016] When α>30°, it can be judged as a dynamic point;

[0017] At the same time, the depth value difference ΔD is calculated. If ΔD = 0, the point is considered static. If ΔD > 0, the point is considered dynamic.

[0018] Where ΔD = D proj -D′, D′ represents the depth of the key point in the current frame.

[0019] S2.3 Dynamic point removal: The two dynamic regions detected by the semantic segmentation network and the multi-view geometry are respectively taken as a union, and all feature points in the final dynamic region union are regarded as dynamic feature points and removed;

[0020] S3. Tracking: After removing feature points, the camera pose is initialized using the remaining static feature points. After obtaining the initial pose, more accurate tracking is performed in the local map, and keyframes are optionally inserted.

[0021] The basis for selecting the inserted key frame is: more than 20 frames have passed since the last inserted key frame, the number of feature points tracked by the current frame is less than 50, and the number of feature points tracked by the current frame is less than 90% of the reference key frame.

[0022] S4. Local map optimization: perform BA optimization on the newly inserted keyframe, all co-view frames of the newly inserted keyframe, and map points in the co-view frames;

[0023] S5. Closed loop detection: Use the dictionary model to compare the similarity between the current frame and the historical frames with the threshold to determine the existence of closed loops. The closed loop detection results are transmitted to the backend for optimization processing.

[0024] The back-end optimization optimizes the camera pose at each time point obtained by the front-end and the loop detection data to achieve real-time visual SLAM tracking.

[0025] S6. Semantic map construction: combining point cloud information with semantic information to construct a semantic point cloud map and a semantic octree map, specifically comprising: combining the semantic information extracted by the improved DeepLabV3plus network with the point cloud information to construct a point cloud map, and completing the conversion of the point cloud to the octree map through the Octo library.

[0026] In the process of constructing the point cloud map, the probability logarithm is used to describe whether the node is occupied, and the octree map is continuously updated, which is expressed as:

[0027]

[0028] Here, let y∈R (a set of real numbers) represent the logarithmic value of probability, and the numerical range of occupancy probability p is [0, 1].

[0029] Compared with the prior art, the present invention has obvious beneficial effects. It can be seen from the above scheme that the network structure of the improved DeepLabV3plus, compared with the traditional DeepLabV3plus, selects MobileNetV2 as the backbone network instead of the original Xception, and adopts transfer learning to load the pre-trained model, which improves the model segmentation performance while reducing the model complexity; in the ASPP module, depthwise separable convolution is used to replace the standard convolution, and the SE attention mechanism is referenced to enhance the weights of feature channels with good network performance, suppress the weights of feature channels that make the network performance poor, improve training efficiency, and thus obtain better segmentation effects.

[0030] The ASPP architecture is constructed by connecting convolution kernels with different dilation rates in parallel. A large dilation rate and large kernel size facilitate the segmentation of large objects, while a small dilation rate and small kernel size facilitate the segmentation of small objects. The ASPP module uses depthwise separable convolution instead of standard convolution. This convolution consists of two parts: depthwise convolution and pointwise convolution. For each channel of the input feature map, a 3×3 convolution with a kernel size is performed to obtain three channel-wise separated feature attributes. Each channel of the output is then subjected to a 1×1 convolution to obtain the output. In standard convolution, the network learns both spatial and channel-wise correlations of the input feature map, combining all inputs in a single operation to produce a new set of outputs. While depthwise separable convolution produces the same output as standard convolution, it can learn features with fewer parameters, effectively reducing the number of model parameters. After the ASPP output feature map, the SE attention mechanism is used to enhance the model's representational power, enrich the image's geometric details, and improve model training efficiency.

[0031] This visual SLAM method is based on semantic segmentation and multi-view geometry. Based on the ORB-SLAM2 framework, the existing three threads of visual odometry, local mapping, and loop closure detection are supplemented with a semantic segmentation module, a dynamic feature point detection module, and a thread for building a semantic octree map. First, data from the RGB-D camera is fed into the tracking thread, where the GCNV2 network is used to extract feature points from the input image frame. The extracted feature points in the current frame include keypoints and descriptors. A modified DeepLabV3plus network model is then used to obtain a per-pixel segmentation result for the RGB image. Moving objects are then filtered out, and dynamic feature points, such as those of walking people, are initially removed. Multi-view geometry algorithms are then used to further remove dynamic objects, and the remaining static features are used for pose estimation. Finally, the semantic map construction thread uses the semantic information extracted by the modified DeepLabV3plus to build a map. The point cloud is then converted to an octree map using the Octo library.

[0032] In summary, the present invention uses a semantic segmentation model and a multi-view geometry method to enable the visual SLAM algorithm to effectively reduce the impact of dynamic objects and improve the positioning accuracy of the system.

[0033] The beneficial effects of the present invention are further illustrated below through specific implementation methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is a flow chart of the present invention;

[0035] Figure 2 This is a framework diagram of the semantic segmentation process of the present invention;

[0036] Figure 3 This is a multi-view geometric principle diagram of the present invention. DETAILED DESCRIPTION

[0037] The following, in conjunction with the accompanying drawings and preferred embodiments, describes in detail the specific implementation, features and efficacy of a deep learning-based visual SLAM method in a dynamic environment proposed by the present invention.

[0038] See also Figure 1 The present invention provides a visual SLAM method based on deep learning in a dynamic environment, comprising the following specific steps:

[0039] S1. Feature point extraction: The GCNv2 network is used for feature extraction. GCNv2 receives a single-channel input image and resizes it to 320×240. The network receives the resized image, extracts features from it, and performs homogenization and non-maximum suppression on the features to obtain evenly distributed feature points and their corresponding descriptors.

[0040] S2. Dynamic point removal: Dynamic feature points are detected by combining semantic segmentation network and multi-view geometry. The specific method is as follows:

[0041] S2.1 uses the improved DeepLabV3plus network to perform semantic segmentation on the RGB image to obtain pixel-level semantic segmentation results. The pixels corresponding to the prior dynamic object people are regarded as dynamic areas, and the feature points within them are removed as dynamic feature points.

[0042] Specifically, the following steps are involved: After the image is input into the network, features are extracted using the lightweight MobilenetV2 backbone network, resulting in two output values: a feature map containing high-level semantic information and a feature map containing low-level features. The feature map containing high-level semantic information first passes through the Atrous Spatial Pyramid Pooling (ASPP) module, then utilizes the SE attention mechanism to improve training efficiency. The number of channels is then adjusted using 1x1 convolution, and four upsampling operations are performed. The feature map containing low-level features uses 1x1 convolution to adjust the number of channels, is then stacked with the feature map containing high-level semantic information, and finally uses 3x3 convolution to adjust the channels before being upsampled four times to obtain the final segmentation result.

[0043] Among them, ASPP captures rich contextual information by performing depth-separable convolution and global average pooling operations at different resolutions. The dilated convolution algorithm can be expressed as:

[0044]

[0045] Where y is the input feature map, w[k] represents the size of the convolution kernel (k×k), w is the convolution kernel, k is the size of the convolution kernel, r is the expansion rate, and x is the output feature map. The expansion rate can be understood as the stride size of the convolution kernel during operation, and the size of the receptive field changes with the expansion rate.

[0046] For each channel of the input feature map, convolution with a kernel size of 3×3 is performed step by step to obtain the features of each channel, and the output 3 channels are then subjected to 1×1 convolution one by one to obtain the output, which can effectively reduce the amount of calculation and shorten the running time.

[0047] After ASPP outputs the feature map, the SE attention mechanism is used to enhance the model representation, enrich the geometric detail information of the image, and improve the model training efficiency.

[0048] S2.2 uses multi-view geometry to detect dynamic objects other than people. The process of detecting dynamic objects using multi-view geometry is as follows: First, the key point p in the key frame is projected to the current frame to obtain the projected point p′ and the projection depth D proj , then calculate the disparity angle α between the back projection of p and p′, and at the same time, we also need to calculate the depth value difference ΔD. The disparity angle α and the depth value difference ΔD are used to determine whether the current feature point is a dynamic feature point.

[0049] When α>30°, it can be judged as a dynamic point.

[0050] At the same time, we also need to calculate the depth value difference ΔD. If ΔD = 0, the point is considered to be static. If ΔD > 0, the point is considered to be dynamic.

[0051] Where ΔD = Dproj -D′, D′ represents the depth of the key point in the current frame.

[0052] S2.3 takes the union of the two dynamic regions detected by the semantic segmentation network and the multi-view geometry method respectively, and all feature points in the final dynamic region union are regarded as dynamic feature points and are eliminated.

[0053] S3. Tracking: After removing the feature points, the remaining static feature points are used to initialize the camera pose. After obtaining the initial pose, more accurate tracking is performed in the local map.

[0054] S4. Insert new keyframe: The basis for selecting a new keyframe to insert is as follows: more than 20 frames have passed since the last inserted keyframe, the current frame tracks fewer than 50 feature points, and the current frame tracks fewer than 90% of the reference keyframe's feature points.

[0055] S5. Local map optimization: perform BA optimization on the newly inserted keyframe, all co-view frames of the newly inserted keyframe, and map points in the co-view frames.

[0056] S6. Loop closure detection: Use a bag-of-words (Bow) dictionary model to compare the similarity between the current frame and the historical frames with a threshold to determine the existence of a closed loop and whether the robot has reached a place it has visited before. The loop closure detection results are then transmitted to the backend for optimization processing.

[0057] S7. Back-end optimization: further optimizes and processes the camera pose and loop detection data at each time point obtained by the front-end to achieve real-time visual SLAM tracking.

[0058] S8. Semantic map construction: point cloud information is combined with semantic information to construct semantic point cloud maps and semantic octree maps. Specifically, the semantic information extracted by the improved DeepLabV3plus is combined with point cloud information to construct a point cloud map, and the point cloud to octree map conversion is completed through the Octo library.

[0059] During the graph construction process, we choose to use the probability logarithm to describe whether a node is occupied, which is expressed as:

[0060]

[0061] Here, let y∈R (a set of real numbers) represent the logarithmic value of probability, and the numerical range of occupancy probability p is [0, 1].

[0062] Furthermore, the above formula can be reversibly transformed into:

[0063]

[0064] The observation probability of a node n at time T is P(n|Zt ), Z represents the observed data, then the probability of it being occupied is P(n|Z 1：T ) is expressed as:

[0065]

[0066] Among them, P(n) represents the prior probability that node n is occupied, P(n|Z 1：T-1 ) represents the estimated probability of n nodes from the start to time T-1. We usually set the prior probability P(n) to 0.5, so the above formula is converted into the probability pair form L(n|Z 1：T ), represents the logarithmic value of the probability of node n from the start to time T, then the time T+1 is:

[0067] L(n|Z 1：T+1 )=L(n|Z 1：T-1 )+L(n|Z T )

[0068] Among them, L(n|Z 1：T-1 ) and L(n|Z T ) represent the logarithmic probability of node n being occupied before time T and at time T, respectively. When a node is repeatedly observed to be occupied, its logarithmic probability increases; otherwise, it decreases. Based on this information, the node's occupation probability can be dynamically adjusted, continuously updating the octree map.

[0069] (1) Dynamic target information extraction based on semantic segmentation

[0070] In order to extract pixel-level semantic information from pixels in the image and then determine the prior moving target, the present invention uses the improved segmentation network DeepLabV3plus to complete the semantic segmentation task of the image frame. In recent years, many scholars have continuously proposed new semantic segmentation networks, such as PSPNet with a pixel accuracy of 0.9293 and BiSeNet with a pixel accuracy of 0.9337. DeepLabV3plus adds a decoder module to the framework of DeepLabv3, and integrates multi-scale information in the spatial pyramid pooling layer (ASPP) module based on void convolution. In the decoder architecture, more accurate object boundaries are obtained by restoring spatial information, optimizing the segmentation results, and the pixel accuracy reaches 0.9431. Therefore, the present invention selects DeepLabV3plus and makes improvements on this basis. Taking into account the main application in indoor scenes, the present invention DeepLabv3+ adopts VOC data set for training, and mainly targets people, displays, and chairs as the main segmentation targets, among which people are used as prior dynamic objects. Feature point-based SLAM uses two consecutive frames to initialize the camera pose and removes unmatched dynamic points with the help of the RANSAC (Random Sample Consensus) method. However, when there are many dynamic points, this method performs poorly. In order to obtain a more reliable camera pose, the present invention uses the categories defined in the VOC dataset to mark the targets in the RGB image with different colors, and then uses the results obtained by semantic segmentation to remove dynamic feature points of the object, and uses the remaining static points to initialize the camera pose. Figure 2 The figure shows the semantic segmentation algorithm. After the image is input to the network, DCNN feature extraction produces two output values: feature map 1 containing high-level semantic information and feature map 2 containing low-level features (low_level_features). Map 1 first passes through the ASPP module, and then uses 1×1 convolution to adjust the number of channels to produce map 1′. Map 2 uses 1×1 convolution to adjust the number of channels to produce map 2′. Map 1′ is upsampled four times and then stacked (concat) with map 2′. Finally, after using 3×3 convolution to adjust the channels, the final segmentation result is upsampled four times.

[0071] (2) Multi-view geometry method

[0072] The semantic segmentation network can only detect a priori moving target objects such as people, ignoring the impact of static objects such as chairs and books on SLAM when they undergo passive movement. For example, targets such as books in people's hands or chairs pushed by humans should be regarded as dynamic target objects, but they are regarded as static objects, which will have a greater impact on SLAM. Therefore, the present invention adopts a dynamic feature point detection method of multi-view geometry: the map point cloud is projected to the current frame, and the target objects are distinguished into dynamic targets and static target objects by using the viewpoint difference and depth value change. The principle of multi-view geometry is as follows Figure 3 As shown, for each input frame, multiple frames with a high degree of overlap with the input image frame are selected, and the key point p in the key frame KF is projected to the current frame CF to obtain the projection point p′ and the projection depth D proj The 3D point corresponding to each key point is P, and the value of the parallax angle α (the angle α formed between the back projection of p and p') is obtained through calculation. According to the test experiment on the TUM dataset, when α>30°, it can be judged as a dynamic point. At the same time, we also need to calculate the depth value difference ΔD=D proj -D′, D′ represents the depth of the key point in the current frame. If ΔD=0, the point is considered to be static. If ΔD>0, the point is considered to be dynamic.

[0073] (3) Performance analysis

[0074] Training Dataset: Image semantic segmentation training is completed on the PascalVoc2012 dataset, which contains 20 labeled objects, including airplanes, people, chairs, sofas, etc. The SegmentationClass file in PascalVoc2012 is a semantically labeled image, 2913 images, and the semantically labeled images are different from the original Figure 1 The Annotations folder is primarily used for object detection and contains data labels. The Segmentation file within the ImageSets folder stores indexes for segmenting the training and test datasets. The JPEGImages file contains 17,125 RGB images. The SegmentationObject file stores semantically labeled images.

[0075] Results and Analysis: Table 1 shows a comparison of the segmentation performance of the DeeplabV3plus model before and after the improvements. The results show that the improved DeeplabV3plus model significantly improves both memory usage and single-image processing speed, thanks to its use of the lightweight MobileNetV2 convolutional network for real feature extraction and the replacement of the atrous convolution in the ASPP module with depthwise separable convolution for multi-feature fusion. The model's memory usage is reduced from 439MB to 13.1MB, a 97% reduction in model size. The average run time per image is reduced from 189.4ms to 21.2ms, a reduction of approximately 88%. This significantly optimizes both model memory usage and average image processing time, while maintaining no significant decrease in segmentation accuracy. The improved DeeplabV3plus model significantly improves image processing speed, meeting the real-time requirements for image segmentation. For specific applications, the semantic segmentation model is applied to the construction of 3D dense semantic maps to ensure the real-time performance of the resulting semantic map construction system.

[0076] Table 1 Comparison of model segmentation performance before and after improvement

[0077]

[0078] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of the technical solution of the present invention.

Claims

1. A deep learning-based visual SLAM method in a dynamic environment, characterized by: The method comprises the following steps: S1. Feature Point Extraction: This process uses an RGB image as input to extract features, then performs homogenization and non-maximum suppression on the features to obtain evenly distributed feature points. The feature points extracted from the current frame of the image include key points and descriptors. S2. Dynamic point removal: Dynamic feature points are detected and removed using a combination of semantic segmentation networks and multi-view geometry. The specific steps are as follows: S2.1 performs semantic segmentation on the image using a semantic segmentation network: the semantic segmentation network is an improved DeepLabV3plus network. The improved DeepLabV3plus network is based on the traditional DeepLabV3plus network, adopts MobileNet V2 as the backbone network, uses transfer learning to load the pre-trained model, replaces the standard convolution with depthwise separable convolution in the atrous spatial pyramid pooling module ASPP, and references the SE attention mechanism to increase the weights of feature channels with good network performance and suppress the weights of feature channels that cause poor network performance; The improved DeepLabV3plus network is used to perform semantic segmentation on the RGB image to obtain pixel-level semantic segmentation results. The pixels corresponding to the prior dynamic object person are used as the dynamic area. The specific steps are as follows: After inputting an image and extracting image features using the lightweight backbone network MobilenetV2, two output values are obtained: a feature map containing high-level semantic information and a feature map containing low-level features; The feature map containing high-level semantic information first passes through the Atrous Spatial Pyramid Pooling (ASPP) module, then uses the SE attention mechanism to improve training efficiency. 1x1 convolution is then used to adjust the number of channels, and four upsampling operations are performed. The feature map containing low-level features uses 1x1 convolution to adjust the number of channels, and then is stacked with the feature map of high-level semantic information. Finally, after channel adjustment using 3x3 convolution, four times upsampling is performed to obtain the final pixel-level semantic segmentation result, and the pixels corresponding to the prior dynamic object people are used as the dynamic area. S2.2 uses multi-view geometry to detect dynamic objects other than people: determine the key frames in the image, project the key points in the key frames to the current frame, and use the parallax angle and depth value changes to distinguish the target object into dynamic and static areas; The process is as follows: First, the key point p in the key frame is projected to the current frame to obtain the projection point p' and projection depth D proj , calculate the disparity angle α between the back projection of p and p', and at the same time, calculate the depth value difference ΔD; The parallax angle α and the depth value difference ΔD are used to determine whether the current feature point is a dynamic feature point, so that the dynamic feature points of other dynamic objects detected except people are regarded as dynamic areas; S2.3 Dynamic point removal: The two dynamic regions detected by the semantic segmentation network and the multi-view geometry are respectively taken as a union, and all feature points in the final union of the dynamic regions are regarded as dynamic feature points and removed; S3. Tracking: After removing feature points, the camera pose is initialized using the remaining static feature points. After obtaining the initial pose, more accurate tracking is performed in the local map, and keyframes are optionally inserted. S4. Local map optimization: perform BA optimization on all common view frames and map points in the common view frames of the newly inserted key frame; S5. Closed loop detection: Use the dictionary model to compare the similarity between the current frame and the historical frames with the threshold to determine the existence of closed loops. The closed loop detection results are transmitted to the backend for optimization processing. S6. Semantic map construction: combining point cloud information with semantic information to construct a semantic point cloud map and a semantic octree map, specifically comprising: combining the semantic information extracted by the improved DeepLabV3plus network with the point cloud information to construct a point cloud map, and completing the conversion of the point cloud to the octree map through the Octomap library.

2. The visual SLAM method based on deep learning in a dynamic environment as claimed in claim 1, wherein: The feature extraction in step S1 adopts the GCNv2 network, and the RGB image is first adjusted to 320×240 size before feature extraction.

3. The visual SLAM method based on deep learning in a dynamic environment as claimed in claim 1, wherein: In step S2.1, the atrous spatial pyramid pooling module ASPP is used. ASPP performs depth-separable atrous convolution and global average pooling operations at different resolutions to capture rich contextual information. The atrous convolution algorithm can be expressed as: Among them, y[i] is the feature map of the input hole convolution size i, w[k] represents the size of the convolution kernel k×k, w is the convolution kernel, k is the size of the convolution kernel, r is the expansion rate, and x is the output feature map; among them, the expansion rate is the stride size of the convolution kernel when working, and the size of the receptive field changes with the expansion rate.

4. The visual SLAM method based on deep learning in a dynamic environment as claimed in claim 1, wherein: In step S2.2, when the parallax angle α is greater than 30°, it can be determined that the point is a dynamic point.

5. The visual SLAM method based on deep learning in a dynamic environment as claimed in claim 1, wherein: In step S2.2, the depth value difference ΔD is considered to be static if ΔD=0, and dynamic if ΔD>0. proj -D′, D′ represents the depth of the key point in the current frame.

6. The visual SLAM method based on deep learning in a dynamic environment as claimed in claim 1, characterized in that: The basis for selecting the key frame to be inserted in step S3 is that the time since the last key frame insertion is more than 20 frames and the number of feature points tracked in the current frame is less than 50.

7. The visual SLAM method based on deep learning in a dynamic environment as claimed in claim 1, characterized in that: The back-end optimization in step S5 is to optimize the camera pose and loop detection data at each time point obtained by the front-end to achieve real-time tracking.

8. The visual SLAM method based on deep learning in a dynamic environment as claimed in claim 1, wherein: In the process of constructing the point cloud map in step S6, the probability logarithm is used to describe whether the node is occupied, and the octree map is continuously updated, which is expressed as: Here, let y∈R represent the logarithmic probability value, R is a set of real numbers, and the numerical range of the occupancy probability p is [0, 1].

Images