A positioning method for track inspection vehicles based on the fusion of lidar and vision

Abstract

This invention relates to a positioning method for a track inspection vehicle based on the fusion of lidar and vision, belonging to the field of rail transit automatic driving technology, and solves the drift problem easily encountered in the positioning of existing track inspection vehicles. This method fully utilizes the advantages of lidar and vision sensors, aiming to provide high-precision and reliable autonomous positioning to support the stable operation of track inspection vehicles in complex environments.

Description

Technical Field

[0001] This invention belongs to the field of automatic driving technology for rail transit, and in particular relates to a positioning method for rail inspection vehicles based on the fusion of lidar and vision. Background Technology

[0002] Rail transit transports a large number of people daily, bringing great convenience to people's travel and daily lives. However, maintenance consumes a significant amount of manpower and resources. To achieve automated track inspection, track inspection vehicles are used to detect and identify defects, damage, and anomalies in the track system. Inspection vehicles play a potentially crucial role in track inspection and maintenance, contributing to improved operational safety and reliability. However, track inspection vehicles face several challenges in their missions. One of these is the problem of autonomous positioning under complex and harsh environmental conditions. Traditional positioning methods typically rely on satellite navigation signals, but in track tunnel environments, satellite signals may be interfered with or unavailable, affecting the positioning accuracy of the inspection vehicle. Furthermore, positioning under low-light conditions is another challenge for track inspection vehicles. At night or in insufficient light, sensor performance is limited, leading to inaccurate positioning. Therefore, it is necessary to develop a new positioning method that can overcome these challenges and provide high-precision, reliable autonomous positioning to support the stable operation of track inspection vehicles in complex environments.

[0003] Currently, many scholars are conducting research on the positioning of inspection vehicles. For example, CN110434831B records the coordinate information of multiple baffles set at intervals on the track as track data into the local database of the inspection vehicle. During the inspection, the position of the inspection vehicle is inspected and located in real time, and the positioning is calibrated according to the track data. However, the scanning process of the baffles is unstable and prone to missing scans, which may lead to positioning failure.

[0004] For example, CN114802358A discloses a positioning method and system for track inspection vehicles based on visual recognition. The problem with this method is that the odometer and positioning wheel are installed on the side of the track, which may prevent them from passing through narrow track junctions. Summary of the Invention

[0005] This invention provides a positioning method for track inspection vehicles based on the fusion of lidar and vision, which addresses the drift problem in existing track inspection vehicle positioning systems. This method fully utilizes the advantages of lidar and vision sensors, aiming to provide high-precision and reliable autonomous positioning to support the stable operation of track inspection vehicles in complex environments.

[0006] This invention provides a positioning method for a track inspection vehicle based on the fusion of lidar and vision, comprising the following steps:

[0007] S1: Collect images of damaged sections of the track and construct a damaged track dataset; collect image data of all parts of the tunnel, extract semantic information of markers in the image data, and construct a tunnel visual feature dataset;

[0008] S2: Equip light inspection vehicles with data acquisition hardware;

[0009] S3: Use acquisition hardware to acquire images of the inspection vehicle during operation; extract the features of each frame in the images; compare the features of each frame with the damaged track dataset and tunnel visual feature dataset from S1 to obtain global similarity and local similarity; identify the damaged track based on global similarity and local similarity.

[0010] S4: The inspection vehicle continues to move forward and continues to acquire global and local similarity data as in S3 until the damaged track is identified and the corresponding marker is matched; the inspection vehicle stops.

[0011] S5: After the inspection vehicle stops, it acquires tunnel point cloud data through the data acquisition hardware on the inspection vehicle, obtains the current pose of the inspection vehicle relative to the marker through the point cloud data, and publishes the current position of the inspection vehicle.

[0012] Optionally, global image features of the job image can be extracted using a MobileNetV3 convolutional neural network.

[0013] Optionally, the specific steps for extracting global image features of the task image using the MobileNetV3 convolutional neural network are as follows:

[0014] The inverse residual structure is used to upscale each frame of the image through 1×1 convolution;

[0015] Feature extraction is performed using 3×3 convolutions.

[0016] Dimensionality reduction is performed using 1×1 convolutions to obtain the feature vectors of each frame.

[0017] The feature vectors of each frame are processed to obtain the global image features of the target image.

[0018] Optionally, the specific steps of S3 are as follows:

[0019] S31: Acquire operational images of the forward field of view during patrol vehicle operations;

[0020] S32: Scale each frame of the task image to a preset size and use a convolutional neural network to extract global image features of the task image;

[0021] S33: Extract ORB local image features from the job image;

[0022] S34: Obtain the Euclidean distance between the global feature vectors of the images in the tunnel visual feature dataset and the global image features of the task images, and obtain the global similarity based on the Euclidean distance;

[0023] S35: Use the bag-of-words model to perform image retrieval on ORB local image features to obtain local similarity;

[0024] S36: Identify damaged orbits based on global and local similarity.

[0025] Optionally, the specific steps for extracting the ORB local image features of the job image in S33 are as follows:

[0026] The FAST algorithm is used to obtain the pixel values ​​of multiple pixels surrounding the candidate key point in the task image. If the pixel values ​​of several or more consecutive pixels surrounding the candidate key point are greater than or less than the pixel value of the candidate key point, then the candidate key point is selected. The ORB local image features are obtained based on the selected key point.

[0027] Optionally, in S34, the normalized global features of the global image features of the job image are obtained, and the Euclidean distance between the global feature vector of the image in the tunnel visual feature dataset and the normalized features of the job image is obtained.

[0028] Optionally, the specific steps of step 35 are as follows:

[0029] S351: Generating a feature image dictionary based on a tunnel visual feature dataset;

[0030] S352: Based on the feature image dictionary, obtain the low-influence feature weights and high-influence feature weights of words in the images of the tunnel visual feature dataset;

[0031] S353: Obtain the word weight of the corresponding word based on the low-impact feature weight and the high-impact feature weight;

[0032] S354: Obtain the bag-of-words vectors of the words in the image;

[0033] S355: Obtain the local similarity between a single frame image in the assignment image and a single frame image in the tunnel visual feature dataset based on the bag-of-words vectors of words in the image.

[0034] Optionally, a global feature similarity threshold and a local feature similarity threshold are set. If the global similarity and local similarity of the current frame image in the operation image are both within the set threshold range, it is considered that the current frame image matches the target in the damaged track dataset and visual feature dataset in S1, and the track damage and markers in the field of view of the inspection vehicle in the current frame are identified; the inspection vehicle stops.

[0035] Optionally, the specific steps of S5 are as follows:

[0036] S51: After the inspection vehicle stops, it acquires tunnel point cloud data using the lidar on the inspection vehicle;

[0037] S52: Obtain the predicted pose of the inspection vehicle relative to the marker;

[0038] S53: Obtain the predicted point cloud of the inspection vehicle based on the predicted pose of the inspection vehicle and the observed point cloud data;

[0039] S54: Obtain the pose change relationship at the current moment and the proportion of time in the current resampling time;

[0040] S55: Based on the proportion of the current resampling time, perform linear interpolation on the total sampling time t to obtain the extended sampling time;

[0041] S56: Obtain the current pose of the radar;

[0042] S57: Project the radar's current pose onto the predicted point cloud in S53 to achieve intra-frame correction of the predicted point cloud, and obtain and publish the current position of the inspection vehicle.

[0043] Optionally, the hardware for data acquisition in the light inspection vehicle includes a power supply, lidar, and cameras.

[0044] Compared with the prior art, the present invention has at least the following beneficial effects:

[0045] 1. The method of the present invention uses the matching method of the current frame point cloud and the local point cloud map to correct the cumulative error, thereby reducing the detection error caused by the observation noise of the sensor during the operation of the inspection vehicle.

[0046] 2. The method of the present invention uses a camera to identify the semantic information of markers such as tunnel wall mileage markers and support segment number markers, and combines this with the pose of a lidar odometer. This improves the accuracy of positioning, taking into account the characteristics of tunnel environment having few features and being prone to feature degradation.

[0047] 3. The method of the present invention proposes a lightweight network that improves the accuracy and speed of image feature extraction by fusing and matching global and local image features of the forward view image of the inspection vehicle. Attached Figure Description

[0048] The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of the invention.

[0049] Figure 1 This is a flowchart of the positioning method for the track inspection vehicle of the present invention;

[0050] Figure 2This is a flowchart illustrating the process of extracting global image features from job images using the MobileNetV3 convolutional neural network, as described in this invention. Detailed Implementation

[0051] To better understand the above-described objectives, features, and advantages of the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments of the present invention and the features thereof can be combined with each other. Furthermore, the present invention can be implemented in other ways different from those described herein; therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0052] A specific embodiment of the present invention, such as Figure 1-2 A positioning method for a track inspection vehicle based on the fusion of lidar and vision is disclosed, including the following steps:

[0053] S1: Collect images of damaged sections of the track and construct a damaged track dataset; collect image data of all parts of the tunnel, extract semantic information of markers such as odometer signs and support segment number signs from the image data, and construct a tunnel visual feature dataset;

[0054] S2: Equip light inspection vehicles with data acquisition hardware;

[0055] Specifically, the hardware includes a power supply, LiDAR, and a camera;

[0056] Furthermore, the power supply is used to ensure that the inspection vehicle can provide the necessary power and data acquisition capabilities during operation;

[0057] S3: Collect images of the inspection vehicle during operation; extract features from each frame of the images; compare the features of each frame with the damaged track dataset and tunnel visual feature dataset from S1 to obtain global and local similarity; identify the damaged track based on global and local similarity.

[0058] S31: Use the camera's webcam to capture images of the inspection vehicle's forward field of view during operation;

[0059] S32: Scale each frame of the task image to a preset size, and use the MobileNetV3 convolutional neural network to extract global image features of the task image;

[0060] Specifically, such as Figure 2As shown, the MobileNetV3 convolutional neural network uses an inverse residual structure. First, it increases the dimensionality of each frame image through 1×1 convolution, then uses 3×3 convolution for feature extraction, and then uses 1×1 convolution for dimensionality reduction to obtain the feature vector of each frame image. The feature vector of each frame image is then processed to obtain the global vector of the working image (i.e., global image features).

[0061] The expression for processing the feature vector of each frame is:

[0062]

[0063] Where h-swish(x) represents the activation function; x represents the feature vector of the current frame; and ReLU6 is a variant rectified linear unit. The ReLU6 variant truncates the negative parts of the feature vector x of each frame to zero and truncates values ​​greater than 6 to 6. The specific expression is:

[0064] RELU6(x) = min(max(0,x),6).

[0065] This invention utilizes inverted residual structures to reduce information loss caused by activation functions. Simultaneously, the use of depthwise separable convolutions significantly reduces network computation. Furthermore, the h-swish activation function further reduces computation while improving network performance.

[0066] S33: Extract ORB local image features from the job image;

[0067] Specifically, the FAST algorithm is used to obtain the pixel values ​​of 16 pixels surrounding the candidate key point p in the task image. If the pixel values ​​of 8 or more consecutive pixels surrounding the candidate key point p are greater than or less than the pixel value of the candidate key point p, then the candidate key point p is selected as the key point. Based on the selected key point, the ORB local image features are obtained.

[0068] Furthermore, since an image contains multiple FAST feature points, the ORB algorithm is used to select N corner points from these FAST feature points in each frame as FAST corner points. Simultaneously, to describe the directionality of the FAST feature points in an image, the direction vector between the image's geometric center O and centroid C is used. First, the key points are described using the BRIEF algorithm. BRIEF is a binary descriptor. Each binary digit represents the size relationship between two pixels, p and q. By randomly selecting the positions of p and q using probability, a 1024-dimensional BRIEF feature descriptor (i.e., an ORB feature descriptor) can be obtained.

[0069] The extracted ORB feature descriptors are ultimately used to represent keypoints, exhibiting rotation, translation, and scaling invariance, and efficiently describing image features. The feature vectors extracted using the ORB feature descriptors serve as local image vectors (i.e., local image features).

[0070] S34: Obtain the Euclidean distance between the global feature vectors of the images in the tunnel visual feature dataset and the global image features of the task images, and obtain the global similarity based on the Euclidean distance.

[0071] The expression for Euclidean distance is:

[0072]

[0073] in, This represents the global feature vector of the m-th image in the tunnel visual feature dataset; D represents the global image features of the current frame of the task image; m Let Euclidean distance be the Euclidean distance between the global feature vector of the m-th image in the image dataset and the global feature vector of the current frame of the task image.

[0074] The similarity between each frame of the task image and the images in the tunnel visual feature dataset is represented by the calculated Euclidean distance.

[0075] Preferably, the global feature vector of the m-th image is obtained through a neural network.

[0076] By iterating through each frame of the task image and each frame of the tunnel visual feature dataset, the global similarity between the task image and the tunnel visual feature dataset is obtained using Euclidean distance.

[0077] Furthermore, in S34, the normalized global features of the global image features of the task image are obtained, the Euclidean distance between the global feature vector of the image in the tunnel visual feature dataset and the normalized features of the task image is obtained, and the global similarity is obtained based on the Euclidean distance.

[0078] The steps for obtaining normalized global features are as follows: The global image features of the task image in regularization step S32 are processed to obtain normalized global features, expressed as:

[0079]

[0080] Among them, v n This represents the nth dimension of the global image features, which are the global image features of the job image extracted by the neural network in S32. R represents the normalized global features, which serve as the global feature vector of the current frame of the image in the subsequent formulas; R represents a real number.

[0081] The Euclidean distance between the global feature vectors of the images in the tunnel visual feature dataset and the normalized global features of the task image is obtained, and the global similarity is obtained based on the Euclidean distance. At this point, the global image features of the task image in the current frame are... Normalized global features obtained after regularization processing

[0082] S35: Use the bag-of-words model to perform image retrieval on ORB local image features to obtain local similarity;

[0083] This invention employs a bag-of-words model to perform image retrieval using local image features of ORB, which can efficiently and accurately match the required local features.

[0084] S351: Generate a feature image dictionary based on a tunnel visual feature dataset.

[0085] Furthermore, the K-means++ algorithm is used to cluster the feature image dictionary to obtain a simplified dictionary; and a dictionary tree structure is constructed to facilitate image retrieval.

[0086] S352: Based on the feature image dictionary, obtain the low-influence feature weights and high-influence feature weights of words in the images of the tunnel visual feature dataset;

[0087] Understandably, each image is composed of many words. When a certain feature frequently appears in an image dataset, its contribution to image classification is relatively low. For example, features such as tracks appear in the operating environment of an inspection vehicle. Such features cannot reflect the characteristics of the inspection vehicle's operating environment, so their weight should be relatively low, which is a low-impact feature weight. On the other hand, if a certain feature does not frequently appear in an image, then this feature is more able to reflect the characteristics of this image, and its discriminative power is higher, which is a high-impact feature weight.

[0088] The expression for the low-impact feature weights of words is:

[0089]

[0090] Among them, IDF j The low-impact feature weights of word j in the image are represented by N, where N is the number of words in the tunnel visual feature dataset, and n is the number of words in the dataset. j Let be the number of words j in the tunnel feature dataset.

[0091] The expression for the weight of high-impact features of a word is:

[0092]

[0093] Among them, TF jThe high-influence feature weight of word j in the image is represented by ; N is the number of words in the tunnel visual feature dataset, n j Let be the number of words j in the tunnel visual feature dataset.

[0094] S353: Obtain the word weight of the corresponding word based on the low-impact feature weight and the high-impact feature weight, the expression is:

[0095] w j =IDF j ×TF j

[0096] Among them, w j This represents the word weight of word j in the image.

[0097] S354: Obtain the bag-of-words vectors of words in the image, expressed as:

[0098]

[0099] in, n represents the bag-of-words vector of word j in image M; j w represents the number of words j in the tunnel feature dataset; j This represents the word weight of word j in the image.

[0100] S355: Local similarity is obtained between single-frame images in the assignment image and single-frame images in the tunnel visual feature dataset based on the bag-of-words vectors of words in the image. This is achieved by calculating the L1 norm of the difference between different bag-of-words vectors in the two images, expressed as:

[0101]

[0102] Among them, v A The bag-of-words vector set representing a single frame image in the task image; v B s(v) represents the bag-of-words vector set of a single frame image in the tunnel visual feature dataset; A -v B () represents the L1 norm of the difference between the different bag-of-words vectors of two images; This represents the bag-of-words vector of word j in a single frame of the task image; This represents the bag-of-words vector of word j in the tunnel visual feature dataset.

[0103] S36: Identify damaged orbits based on global and local similarity.

[0104] S4: The inspection vehicle continues to move forward and continues to acquire global and local similarity data as in S3 until the damaged track is identified and the corresponding marker is matched; the inspection vehicle stops.

[0105] Specifically, a global feature similarity threshold and a local feature similarity threshold are set. If the global similarity and local similarity of the current frame image in the operation image are both within the set threshold range, it is considered that the current frame image matches the target in the damaged track dataset and visual feature dataset in S1, and the track damage and markers in the field of view of the inspection vehicle in the current frame are identified.

[0106] S5: After the inspection vehicle stops, it acquires tunnel point cloud data through the lidar on the inspection vehicle, obtains the current pose of the inspection vehicle relative to the marker through the point cloud data, and publishes the current position of the inspection vehicle.

[0107] S51: After the inspection vehicle stops, it acquires tunnel point cloud data using the lidar on the inspection vehicle;

[0108] S52: Obtain the predicted pose of the inspection vehicle relative to the marker, the expression is:

[0109]

[0110] in, This represents the predicted pose of the inspection vehicle at time k+1; This represents the pose of the inspection vehicle relative to the initial coordinate system at time k; This represents the motion estimation result of the inter-frame matching between time k+1 and time k;

[0111] S53: Obtain the predicted point cloud of the inspection vehicle based on the predicted pose of the inspection vehicle and the observed point cloud data;

[0112] The constructed point cloud map is Q k ;

[0113] The inspection vehicle's predicted pose at time k+1 Observation point cloud Transform to the initial coordinate system to obtain the observation point cloud in the initial coordinate system;

[0114] The observed point cloud in the initial coordinate system is accumulated into the point cloud map to obtain the point cloud pose of the inspection vehicle at time k+1.

[0115] Furthermore, to reduce the computational load of point cloud registration, this invention extracts and saves a cubic region, and divides this cube into multiple sub-cube regions, retrieving the sub-cubes using an index. After obtaining the patrol vehicle's location, the cube index is calculated using the location coordinates. The local map corresponding to this index is matched with the corrected current frame to obtain the patrol vehicle's accurate location. Based on the patrol vehicle's location, the point cloud information of the surrounding environment collected in real time by the patrol vehicle is accumulated to construct a point cloud map of the surrounding environment during the patrol vehicle's operation.

[0116] S54: Obtain the pose change relationship at the current moment and the proportion of time in the current resampling time;

[0117] The expression for the pose change at the current moment is:

[0118]

[0119] Among them, R k R is the rotation matrix at time k (i.e., the current time); k-1 It is the rotation matrix at time k-1 (i.e., the previous time step); T is the rotation matrix between the current time step and the previous time step; k T is the translation matrix at the current time; k-1 This is the translation matrix from the previous time step; It is the translation matrix between the previous time step and the previous time step.

[0120] The expression for the proportion of time in the current resampling time is:

[0121]

[0122] Where t represents the total sampling time; t k It is the sampling time of the last point at the current moment; t k-1 It is the sampling time of the last point in the previous moment.

[0123] S55: Based on the proportion of the current resampling time, linear interpolation is performed on the total sampling time t to obtain the extended sampling time. At this time:

[0124]

[0125] in, Let θ represent the rotation matrix between the current time step and the previous time step; e represents the natural constant, and θ represents the rotation matrix between the current time step and the previous time step. The magnitude of the rotation axis, w represents the unit vector of the rotation axis. Describes the skew-symmetric matrix of w; It is represented as the translation matrix of the linear interpolation of the previous frame.

[0126] S56: Obtain the current pose of the radar, expressed as:

[0127]

[0128] Among them, R t Represents the rotation matrix at the current moment; T represents the rotation matrix at the previous time step; t This represents the translation matrix for the current frame; This represents the translation matrix for the linear interpolation of the previous frame;

[0129] S57: Project the radar's current pose onto the predicted point cloud in S53 to achieve intra-frame correction of the predicted point cloud, and obtain and publish the current position of the inspection vehicle.

[0130] In order to eliminate point cloud deformation caused by the movement of the inspection vehicle, this invention requires intra-frame correction to eliminate the effects of motion.

[0131] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A positioning method for a track inspection vehicle based on the fusion of lidar and vision, characterized in that, Includes the following steps: S1: Collect images of damaged sections of the track and construct a damaged track dataset; collect image data of all parts of the tunnel, extract semantic information of markers in the image data, and construct a tunnel visual feature dataset; S2: Equip light inspection vehicles with data acquisition hardware; S3: Use acquisition hardware to acquire images of the inspection vehicle during operation; extract image features from each frame of the images; compare the image features of each frame with the damaged track dataset and tunnel visual feature dataset from S1 to obtain global and local similarity; identify the damaged track based on global and local similarity, the specific steps are as follows: S31: Acquire operational images of the forward field of view during patrol vehicle operations; S32: Scale each frame of the task image to a preset size and use a convolutional neural network to extract global image features of the task image; S33: Extract ORB local image features from the job image; S34: Obtain the Euclidean distance between the global feature vectors of the images in the tunnel visual feature dataset and the global image features of the task images, and obtain the global similarity based on the Euclidean distance; S35: Use the bag-of-words model to perform image retrieval on local image features of ORB to obtain local similarity; S36: Identify damaged orbits based on global and local similarity; S4: The inspection vehicle continues to move forward and continues to acquire global and local similarity data as in S3 until the damaged track is identified and the corresponding marker is matched; the inspection vehicle stops. Set global feature similarity threshold and local feature similarity threshold. If the global similarity and local similarity of the current frame image in the operation image are both within the set threshold range, it is considered that the current frame image matches the target in the damaged track dataset and visual feature dataset in S1, and the track damage and markers in the field of view of the inspection vehicle in the current frame are identified; the inspection vehicle stops. S5: After the inspection vehicle stops, it acquires tunnel point cloud data through the data acquisition hardware on the inspection vehicle, obtains the current pose of the inspection vehicle relative to the marker through the point cloud data, and publishes the current position of the inspection vehicle.

2. The positioning method for a track inspection vehicle according to claim 1, characterized in that, Global image features of the task images are extracted using the MobileNetV3 convolutional neural network.

3. The positioning method for a track inspection vehicle according to claim 2, characterized in that, The specific steps for extracting global image features from the task image using the MobileNetV3 convolutional neural network are as follows: The inverse residual structure is used to upscale each frame of the image through 1×1 convolution; Feature extraction is performed using 3×3 convolutions; Dimensionality reduction is performed using 1×1 convolution to obtain the feature vectors of each frame of the image; The feature vectors of each frame are processed to obtain the global image features of the target image.

4. The positioning method for a track inspection vehicle according to claim 1, characterized in that, The specific steps for extracting ORB local image features from the job image in S33 are as follows: The FAST algorithm is used to obtain the pixel values ​​of multiple pixels surrounding the candidate key point in the task image. If the pixel values ​​of several or more consecutive pixels surrounding the candidate key point are greater than or less than the pixel value of the candidate key point, then the candidate key point is selected as the key point. ORB local image features are obtained based on selected key points.

5. The positioning method for a track inspection vehicle according to claim 1, characterized in that, In S34, the normalized global features of the global image features of the task image are obtained, and the Euclidean distance between the global feature vector of the image in the tunnel visual feature dataset and the normalized features of the task image is obtained.

6. The positioning method for a track inspection vehicle according to claim 5, characterized in that, The specific steps of step 35 are as follows: S351: Generating a feature image dictionary based on a tunnel visual feature dataset; S352: Based on the feature image dictionary, obtain the low-influence feature weights and high-influence feature weights of words in the images of the tunnel visual feature dataset; S353: Obtain the word weight of the corresponding word based on the low-impact feature weight and the high-impact feature weight; S354: Obtain the bag-of-words vectors of the words in the image; S355: Obtain the local similarity between a single frame image in the assignment image and a single frame image in the tunnel visual feature dataset based on the bag-of-words vectors of words in the image.

7. The positioning method for a track inspection vehicle according to claim 1, characterized in that, The specific steps of S5 are as follows: S51: After the inspection vehicle stops, it acquires tunnel point cloud data using the lidar on the inspection vehicle; S52: Obtain the predicted pose of the inspection vehicle relative to the marker; S53: Obtain the predicted point cloud of the inspection vehicle based on the predicted pose of the inspection vehicle and the observed point cloud data; S54: Obtain the pose change relationship at the current moment and the proportion of time in the current resampling time; S55: Based on the proportion of the current resampling time, perform linear interpolation on the total sampling time t to obtain the extended sampling time; S56: Obtain the current pose of the radar; S57: Project the radar's current pose onto the predicted point cloud in S53 to achieve intra-frame correction of the predicted point cloud, and obtain and publish the current position of the inspection vehicle.

8. The positioning method for a track inspection vehicle according to claim 1, characterized in that, The hardware for data acquisition in light inspection vehicles includes a power supply, lidar, and cameras.

Images