Vehicle position identification method and system based on point cloud descriptor extraction

Abstract

The invention discloses a vehicle position identification method and system based on point cloud descriptor extraction. The method comprises the following steps: carrying out denoising processing on obtained vehicle laser point cloud data; and carrying out point cloud global descriptor extraction on the laser point cloud data after de-noising processing by adopting a point cloud descriptor pre-training model, and obtaining a vehicle position according to the obtained point cloud global descriptor. The geometric information of the point cloud is reserved to a great extent while the use of the original point cloud data with huge data volume is avoided, the attention mechanism is introduced, the space and time dimension features of the data are extracted, and a global descriptor with rotation invariance is constructed. Compared with a traditional descriptor construction method, the method has the advantages that the constructed descriptor is more robust, the position re-identification effect is good, and the position re-identification can be well carried out under the condition of reverse re-visit or re-visit with a certain rotation angle.

Description

Technical Field

[0001] The present invention belongs to the field of high-precision maps and positioning technology, and relates to a vehicle position recognition method and system based on point cloud descriptor extraction. Background Art

[0002] High-precision maps and positioning technology are basic technologies in the field of traffic dispatching and management that integrate many high technologies such as sensor technology, computer technology, information processing technology, and communication technology. Among them, the Global Navigation Satellite System (GNSS) and the Inertial Navigation System (INS) are two commonly used systems for ground vehicle positioning. However, in some scenarios, the GNSS signal cannot be received by the corresponding signal receiver due to obstruction, and the INS has the defect of cumulative error. The combination of the two can improve the performance of the positioning system to a certain extent, but it still cannot meet the positioning needs in complex environments.

[0003] To solve the above defects, the simultaneous localization and mapping technology (SLAM) was born. SLAM technology is mainly divided into visual SLAM and laser SLAM according to the type of applied sensors. Visual SLAM technology uses the camera as the core sensor for map construction and positioning, while laser SLAM uses the lidar as the core sensor for map construction and positioning. The SLAM technology algorithm process mainly includes four parts: front-end odometer, back-end optimization, loop detection, and map construction. The loop detection algorithm detects whether there is a loop by performing position recognition, thereby eliminating the front-end accumulated error. Most of the existing loop detection or position recognition algorithms are based on the method of manually extracting descriptors, using few features, and it is difficult to use the rich geometric information of three-dimensional point clouds. The detected loops are few or not accurate enough, resulting in the inability to minimize the accumulated error, lack of consistency in mapping, and inability to accurately obtain vehicle location information. Summary of the invention

[0004] The purpose of the present invention is to provide a vehicle position recognition method and system based on point cloud descriptor extraction to overcome the shortcomings of the prior art.

[0005] A vehicle position recognition method based on point cloud descriptor extraction comprises the following steps:

[0006] S1, denoising the acquired vehicle laser point cloud data;

[0007] S2, use the point cloud descriptor pre-training model to extract the point cloud global descriptor of the denoised laser point cloud data, and obtain the vehicle position based on the obtained point cloud global descriptor and the corresponding screening method.

[0008] Preferably, specifically, laser radar is used to acquire laser point cloud data.

[0009] Preferably, point cloud key frames are extracted from the laser point cloud data of the vehicle to be positioned, and several frames with the highest similarity are obtained as candidate frames by comparing the similarities of global descriptors corresponding to the acquired point cloud key frames, and the candidate frames are used to realize position recognition or repositioning.

[0010] Preferably, the point cloud key frames are selected by taking one frame every ten frames of point cloud data or taking one frame every meter or when the heading angle changes by more than ten degrees.

[0011] Preferably, the point cloud descriptor pre-training model is obtained by training the model using a point cloud depth map and a point cloud bird's-eye view map.

[0012] Preferably, the laser point cloud is converted into a two-dimensional range image by spherical projection to obtain a point cloud depth map.

[0013] Preferably, a frame of laser point cloud is divided according to the ScanContext method, and the maximum height value of the point cloud is selected in each curved-edge trapezoidal cylinder for filling to generate a point cloud bird's-eye view.

[0014] Preferably, a spatial feature is generated using the spatial information of a single-frame point cloud depth map and a point cloud bird's-eye view, and feature extraction is performed again through a layer of convolutional neural network to obtain a point cloud global descriptor.

[0015] A vehicle position recognition system based on point cloud descriptor extraction, data acquisition preprocessing module and recognition module;

[0016] The data acquisition preprocessing module performs denoising on the acquired vehicle laser point cloud data;

[0017] The recognition module uses the point cloud descriptor pre-training model to extract the point cloud global descriptor from the denoised laser point cloud data, and obtains the vehicle position based on the obtained point cloud global descriptor.

[0018] Preferably, the data acquisition preprocessing module specifically adopts a laser radar; point cloud key frames are extracted from the laser point cloud data of the vehicle to be located, and several frames with the highest similarity are obtained as candidate frames by comparing the similarities of global descriptors corresponding to the acquired point cloud key frames, and position recognition or relocation is achieved using the candidate frames.

[0019] Compared with the prior art, the present invention has the following beneficial technical effects:

[0020] The present invention provides a vehicle position recognition method based on point cloud descriptor extraction, which performs denoising on the acquired vehicle laser point cloud data; uses a point cloud descriptor pre-training model to extract a point cloud global descriptor from the denoised laser point cloud data, and obtains the vehicle position based on the acquired point cloud global descriptor. The present invention uses both a point cloud depth map and a point cloud bird's-eye view map as inputs, avoiding the use of a large amount of original point cloud data while retaining the geometric information of the point cloud to a great extent. The attention mechanism is introduced to extract the spatial and temporal dimensional features of the data at the same time, and a global descriptor with rotation invariance is constructed. Compared with the traditional descriptor construction method, the descriptor constructed by the present invention is more robust and has a good position re-identification effect. In the case of reverse revisit or revisit with a certain rotation angle, it can perform position re-identification well. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Schematic diagram of a laser radar in an embodiment of the present invention.

[0022] Figure 2 Schematic diagram of point cloud depth map coordinate conversion in an embodiment of the present invention.

[0023] Figure 3 It is a point cloud depth map coordinate conversion diagram in an embodiment of the present invention.

[0024] Figure 4 It is a coordinate conversion diagram of the point cloud bird's-eye view in an embodiment of the present invention.

[0025] Figure 5 It is a network model diagram in an embodiment of the present invention.

[0026] Figure 6 It is a feature pre-extraction model diagram in an embodiment of the present invention.

[0027] Figure 7 is a diagram of a single-frame attention mechanism in an embodiment of the present invention.

[0028] Figure 8 is a diagram of a multi-frame attention mechanism in an embodiment of the present invention.

[0029] Fig. 9 It is a flow chart of a vehicle position recognition method by extracting point cloud descriptors in an embodiment of the present invention. DETAILED DESCRIPTION

[0030] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0031] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0032] like Figure 8 As shown, the present invention provides a vehicle position recognition method based on point cloud descriptor extraction, the steps are as follows:

[0033] S1, using laser radar to collect laser point cloud data and perform denoising;

[0034] Specifically, it is to filter dynamic objects from laser point cloud data;

[0035] S2, processing the laser point cloud data to obtain a point cloud depth map and a point cloud bird's-eye view map, using the point cloud depth map and the point cloud bird's-eye view map to train the model to obtain a point cloud descriptor pre-training model, using the point cloud descriptor pre-training model to extract the point cloud global descriptor from the laser point cloud data of the to-be-identified position, and obtaining the vehicle position according to the extracted point cloud global descriptor.

[0036] The vehicle position is obtained according to the extracted point cloud global descriptor, which specifically includes the following steps: extracting point cloud key frames from the laser point cloud data of the vehicle to be positioned, selecting the point cloud key frames by taking one frame every ten frames of point cloud data or taking one frame every meter or when the heading angle changes by more than ten degrees, and obtaining several frames with the highest similarity as candidate frames (loop frames) by comparing the similarities of the global descriptors corresponding to the obtained point cloud key frames to achieve position recognition or repositioning.

[0037] The present invention is based on the attention mechanism and neural network, uses laser radar to collect laser point cloud data of vehicles, collects data in different environments, and trains the model after data preprocessing to obtain a point cloud descriptor pre-training model. The present invention mainly processes the laser radar data, converts the original acquired laser point cloud data into a point cloud depth map and a point cloud bird's-eye view map as the input of the neural network model. These two different types of images can utilize the significant features of the laser point cloud, which have the characteristics of high generalization, and restore the geometric information of the laser point cloud from different perspectives, so that the extracted global descriptor is more robust.

[0038] The acquisition of point cloud depth map is to convert the laser point cloud into a two-dimensional distance image through spherical projection. The two-dimensional distance image is a single channel, and the channel value of each channel is the distance from the corresponding laser point to the laser radar in a frame of laser point cloud. This is also a common point cloud data processing method in autonomous driving scenarios. After the laser point cloud is converted into a depth image, it is usually input into a two-dimensional convolutional neural network to achieve feature extraction, target detection, and semantic segmentation tasks.

[0039] The specific spherical projection process is as follows, and the laser radar schematic diagram is as follows Figure 1 As shown, its vertical field of view angle f consists of two parts, namely f up and f down .

[0040] A frame of laser point cloud obtained by scanning a laser radar can be regarded as a hollow cylinder centered on the laser radar. Each point p of a frame of laser point cloud P is projected into i =(x,y,z) through Converting to spherical coordinates converts Cartesian coordinates to polar coordinates.

[0041] like Figure 2 As shown, the three-dimensional laser point cloud in the Cartesian coordinate system is converted into the angle and distance in the polar coordinate system. Their conversion relationship is as follows:

[0042]

[0043] Where r is point p i The distance to the LiDAR, Φ is the LiDAR pitch angle (the vertical field of view of each line of the LiDAR), and θ is the LiDAR yaw angle (rotation around the Z axis).

[0044] The range of the yaw angle obtained by rotating the laser radar around the Z axis is [-π, +π], and the range of the pitch angle is [down, up]. Expanding a frame of laser point cloud with the laser radar as the center can obtain a 3D cylinder. Figure 3As shown, to convert a three-dimensional cylinder into a two-dimensional image, you need to first move the origin to the upper left vertex of the image. The range of the converted yaw angle and pitch angle is as follows:

[0045]

[0046] Then normalization is performed to facilitate finding the position of each laser point on the image. As shown in the following formula:

[0047]

[0048] Assuming the width and height of the projected image are w and h respectively, the position of each laser point in the image can be obtained by the following formula:

[0049]

[0050] Finally, the following formula is simplified to convert the three-dimensional point cloud into image coordinates (,v), that is, the point cloud depth map R, as follows:

[0051]

[0052] Where r is point P i The distance to the lidar is the depth, f = f up + down is the vertical field of view of the laser radar, w and h are the width and height of the projected two-dimensional image. In practical applications, w represents the number of points per line of the laser radar, and h represents the number of scanning lines of the laser radar.

[0053] The point cloud bird's-eye view is also a top-down view of the laser point cloud. Unlike the point cloud depth map, the bird's-eye view is essentially a cylindrical projection. The point cloud bird's-eye view divides a frame of laser point cloud into several layers from low to high, and each layer uses the ScanContext method to generate a point cloud bird's-eye view. Finally, a frame of point cloud is extracted into several bird's-eye views. The bird's-eye view of the present invention is acquired using the method proposed by ScanContext. Another reason for using the ScanContext method to generate a bird's-eye view is that the two frames of ScanContext bird's-eye views generated by revisiting a certain location under rotation are also relatively column-shifted, and thus have the same characteristics as the depth map. The channel value of each channel of the generated point cloud bird's-eye view is the maximum height value of all laser point clouds within the divided curved trapezoid. The specific process is as follows:

[0054] like Figure 4 As shown, let P ij is the set of points in the curved trapezoidal cylinder where the i-th ring and the j-th sector overlap, then a frame of laser point cloud contains all the points which can be expressed as follows:

[0055]

[0056] After dividing a frame of laser point cloud in the above manner, the maximum height value of all points p in each curved trapezoidal cylinder is assigned to the curved trapezoidal cylinder. Therefore, the above curved trapezoidal cylinder is finally encoded as follows:

[0057]

[0058] Where Z(p) is the height value of the returned point p. If there is no point inside the divided curved trapezoidal cylinder, the value is 0.

[0059] A frame of laser point cloud is finally described as a bird's-eye view B of size N s ×N r The matrix can be expressed as follows:

[0060]

[0061] where a ij =φ(Ρ ij ).

[0062] like Figure 5 As shown in FIG. , this figure is a global descriptor extraction network model of the present invention. Figure 6 As shown, the present invention uses a point cloud depth map and a point cloud bird's-eye view map with a channel height and width of 1×32×900 as network input, and first uses a convolutional neural network to pre-extract spatial features as the input of a single-frame attention module. The convolutional neural network module mainly includes nine fully convolutional layers.

[0063] A point cloud depth map or point cloud bird's-eye view map with a height and width of 32×900 is used as input. After nine convolutional layers, the point cloud depth map or point cloud bird's-eye view map is compressed into a feature F of size 256×1×900. The channel dimension is expanded, and the number of feature channels is expanded to 256 dimensions. The width of the feature volume is kept consistent with the width of the original depth map and the bird's-eye view map, which is still 900. At the same time, the height is compressed to 1. In this way, the coarse feature F of the point cloud depth map or point cloud bird's-eye view map is extracted, while the width is guaranteed to remain unchanged. In view of the translation and rotation invariance of the convolutional neural network, the coarse feature F still has rotation equivariance. Taking the point cloud depth map as an example, Con9 is used to represent the feature pre-extraction module. Then the depth map FC obtained after a frame of point cloud is rotated by an angle of θ s =on9(∏R θ Ρ) is just a column translation, and the same is true for the bird's-eye view.

[0064] The single-frame attention module is as follows Figure 6As shown in Figure 1, the coarse feature F generated by the feature pre-extraction module is used as input to further extract spatial features. The single-frame attention module mainly uses the Transformer attention mechanism, which consists of a multi-head attention mechanism module MHSA, a feedforward network module FFN, and a layer normalization module LN.

[0065] The input coarse feature F of the multi-head attention mechanism is divided into several sub-features along the channel dimension according to the number of heads, that is, the query Q, key K and value V of the multi-head attention mechanism are the segmentation of the feature F generated by the feature pre-extraction module along the channel dimension, that is, F = Conc (Q, K, V), where Conc (*) is the channel dimension splicing operation in Transformer. Multi-head attention can learn the relationship between features captured by the self-attention mechanism. This module enhances the association between different features within the single-frame laser point cloud depth map and bird's-eye view, and obtains high-order features at the single-frame spatial level. The feature extracted by the multi-head attention mechanism is represented by Α, then the feature A extracted by the Transformer's attention mechanism can be expressed as follows:

[0066]

[0067] where {Q, K, V} is the segmentation along the channel dimension along the feature volume generated by the depth map or bird's-eye view in the feature pre-extraction module, and d k Indicates the dimension of the split.

[0068] Then input into the feedforward network FFN and the layer normalization module LN to extract the final spatial feature S, whose channel, height and width are 256×1×900, which can be expressed as follows:

[0069] S=LN(FFN(LN(Conc(F,A)))+LN(Conc(F,A)))

[0070] Among them, Conc(*) is the channel dimension concatenation operation in Transformer, which connects the input feature F and the attention mechanism output A, which is a common operation in the Transformer architecture.

[0071] The single-frame attention module is also rotationally equivariant. Since (Q, K, V) is the segmentation of the feature F generated by the feature pre-extraction module along the channel dimension, taking the depth map as an example, we have FC s =onc(QC s ,C s ,C s ), so we have the following derivation:

[0072]

[0073] In Transformer, it is explained that LN and FFN are applied to each element in the feature quantity separately and identically, so it is also rotationally equivariant to the features, so there is the following formula:

[0074] LN(FFN(LN(Conc(FC s ,AC s )))+LN(Conc(FC s ,AC s )))

[0075] =LN(FFN(LN(Conc(F,A)))+LN(Conc(F,A)))C s

[0076] =SC s

[0077] like Figure 7 As shown in the figure, the structure of the multi-frame attention module is the same as that of the single-frame attention module. In this module, the spatial feature S of size 256×1×900 extracted by the single-frame attention module and the coarse feature F of size 256×1×900 extracted by the feature pre-extraction module are first concatenated in the channel dimension to obtain a feature M of size 512×1×900. The feature M combines the spatial features extracted by the attention mechanism with the original coarse features, so it has stronger descriptiveness and can better reflect the characteristics of the original depth map and bird's-eye view. The feature M is then input into a convolutional layer for feature fusion to obtain Then, three consecutive frames The long feature D with a size of 512×1×2700 is obtained by splicing in the width dimension. and D can be expressed as follows:

[0078]

[0079]

[0080] Conw(*) is used to represent the connection in the width dimension. represents the i-th feature

[0081] Finally, the long feature D is used as the input of the multi-frame attention module. In the width dimension splicing, the multi-frame attention module can extract the temporal relationship contained in the continuous frames under the action of the attention mechanism, so that the multi-frame temporal features can be extracted by using the multi-frame temporal association, and finally a spatiotemporal feature K with temporal and spatial characteristics is formed, and its size is 512×1×2700. Since the structure of the multi-frame attention module is the same as that of the single-frame attention module, the above formula can still be used to express it. The feature extracted by the attention mechanism is expressed as A′, as shown in the following formula:

[0082]

[0083] Then the feature K can be expressed as follows:

[0084]

[0085] In summary, the single-frame attention module uses the spatial information of the single-frame point cloud depth map and the point cloud bird's-eye view to generate a spatial feature and extracts the feature again through a layer of convolutional neural network; the multi-frame attention module cascades the features of the single-frame attention module of three consecutive frames in the width dimension, so that the multi-frame attention module can extract the time information to generate a higher-order enhanced spatiotemporal feature K.

[0086] Since the structure of the multi-frame attention module is the same as that of the single-frame attention module, the input is only the concatenation of the features of three consecutive frames in the width dimension, and the concatenation of the width dimension does not affect the rotational equivariance, so the spatiotemporal feature K is still rotationally equivariant and is still equivalent to column translation.

[0087] The feature aggregation module uses the classic NetVLAD network for feature aggregation, and takes the spatiotemporal features K generated by the multi-frame attention module as input for feature aggregation to obtain a global descriptor. Compared with the traditional pooling method, NetVLAD retains more local feature information while having better discrimination and robustness. The NetVLAD network has an important influence in the field of loop detection or location recognition. It is also a network that uses CNN for location recognition. One of its important features is that it can be directly implemented with a neural network, and can aggregate local features into a global descriptor. It can also be trained and optimized end-to-end. Specifically, a convolutional neural network can be used to extract local features of an image, and then these features are passed to the NetVLAD layer for encoding, and finally a global descriptor is output to improve the performance of feature representation. A lot of work in the field of loop detection or location recognition is improved based on this. For example, PointNetVLAD combines PointNet and NetVLAD for descriptor extraction. It first uses PointNet to directly extract features from the original point cloud, and then uses NetVLAD for feature aggregation, and finally generates a global descriptor for loop detection.

[0088] The basic idea of ​​NetVLAD is to cluster each local feature in the image, assign each feature to the nearest cluster center, and calculate the weight of each cluster center based on the assignment result, and finally weight the features of all cluster centers to obtain a global descriptor. Specifically, the process of NetVLAD can be divided into three steps: 1. Local feature clustering, assign each local feature vector of the image to the nearest cluster center. A clustering algorithm is usually used here, in which the number of cluster centers is a hyperparameter. 2. Cluster center weight calculation, for each cluster center, calculate its matching degree with all local features, and use the matching degree as a weight. The softmax function is usually used here to calculate the weight. 3. Feature weighted summation, weighted sum of the feature vectors of all cluster centers to obtain the final global descriptor.

[0089] The present invention uses NetVLAD to perform feature aggregation to generate a global descriptor G. Through the aggregation of NetVLAD, the spatiotemporal feature K is compressed into a global descriptor G of size 1×256, which can be expressed by the following formula:

[0090] G = NetVLAD(K)

[0091] NetVLAD indicates the use of the NetVLAD network for descriptor aggregation.

[0092] At the same time, it is also a permutation-invariant network model. The output features are independent of the order of the input features. Therefore, regardless of whether there is column translation in the spatiotemporal features K of point clouds of different frames, the final global descriptor output after NetVLAD is the same, making the aggregated descriptor rotationally invariant, which can be expressed as follows:

[0093] G = NetVLAD(K) = NetVLAD(KC s )

[0094] KC S Represents the column translation of the spatial feature K.

[0095] The global descriptor obtained by the point cloud depth map branch is represented as G R , the global descriptor obtained by the point cloud bird's eye view branch is represented as G B In the designed network structure, the values ​​of the two width dimensions are directly added to obtain the final global descriptor G. Then the final global descriptor G can be expressed as follows:

[0096] G = addinwidth(G R ,G B )

[0097] Where addinwidth means adding directly in the width dimension.

[0098] The present invention adopts the Faiss fast retrieval algorithm to perform global descriptor similarity calculation and loop retrieval to achieve position identification. Faiss has the characteristics of fast speed, low memory usage and easy use, and is often used to process large-scale data sets. Since the data collected by sensors such as laser radar during the driving process of autonomous driving vehicles or mobile robots will increase, resulting in more and more laser point cloud frames in the data set, it takes too long to use ordinary methods for similarity calculation and loop retrieval. Even if the loop is detected, the real-time requirements of laser SLAM cannot be guaranteed, which affects the driving safety of autonomous vehicles. Therefore, the present invention introduces the Faiss fast retrieval algorithm for similarity calculation and loop retrieval, and uses Euclidean distance to calculate descriptor similarity.

[0099] In order to meet the requirements of SLAM loop detection, the present invention does not consider the previous σ frames of laser point cloud when searching for loops (i.e., searching for the descriptor most similar to the current global descriptor), because the cumulative pose error estimated by the SLAM algorithm in the first few frames of the data set is not large, and loop retrieval is not of practical significance for optimizing this small local map, that is, the first 1 to σ global descriptors are not searched for loops, and the σ value is adjusted by the specific scene. The current global descriptor G is also not considered. i The previous frames and the current global descriptor correspond to the frames. Because in the actual environment, the lidar collects data at a high frequency, and the environment may not change significantly in a short period of time, resulting in a very high similarity of consecutive τ frames during this period, and the current frame is completely similar to the current frame. If these frames are considered, many false loops may be detected, that is, G is not retrieved. i To G i-τ These τ+1 global descriptors, τ is adjusted by the specific scene. At the same time, only the laser point cloud of the global descriptor corresponding frame that is within α meters from the laser point cloud of the frame corresponding to the current global descriptor is considered. The purpose of this is to avoid detecting false loops. If this is not considered, there may be a situation where two global descriptors have high similarity, but in fact the two locations they correspond to are very far away and are only similar in environment. α is adjusted by the specific scene. The present invention adds the above three conditions to the Faiss library to perform global descriptor similarity calculation and loop retrieval. In this way, the global descriptor corresponding frame with the closest Euclidean distance to the current global descriptor is selected as the loop candidate frame to avoid wasting computing resources and retrieving false loops. In summary, the G i The index I of the laser point cloud frame with the highest similarity is expressed as follows:

[0100] I = Faiss(min(d(G i ,G j )))

[0101] Among them G jis a set of descriptors subject to the above three conditions.

[0102] The above method of the present invention provides a vehicle position recognition method based on point cloud descriptor extraction, and uses point cloud depth map and point cloud bird's-eye view map as input, avoiding the use of large amounts of original point cloud data while retaining the geometric information of the point cloud to a great extent. The attention mechanism is introduced to extract the spatial and temporal dimension features of the data at the same time, and a global descriptor with rotation invariance is constructed. Compared with the traditional descriptor construction method, the descriptor constructed by the present invention is more robust and has a good position re-identification effect. In the case of reverse revisit or revisit with a certain rotation angle, it can perform position re-identification well.

Claims

1. A vehicle position recognition method based on point cloud descriptor extraction, characterized in that: The following steps are involved: S1, denoising the acquired vehicle laser point cloud data; S2, use the point cloud descriptor pre-training model to extract the point cloud global descriptor of the denoised laser point cloud data, and obtain the vehicle position based on the obtained point cloud global descriptor and the corresponding screening strategy.

2. The vehicle position recognition method based on point cloud descriptor extraction according to claim 1 is characterized in that: Specifically, laser radar is used to acquire laser point cloud data.

3. The vehicle position recognition method based on point cloud descriptor extraction according to claim 1 is characterized in that: Point cloud key frames are extracted from the laser point cloud data of the vehicle to be positioned. By comparing the similarities of the global descriptors corresponding to the acquired point cloud key frames, several frames with the highest similarity are obtained as candidate frames, and the candidate frames are used to realize position recognition or repositioning.

4. The vehicle position recognition method based on point cloud descriptor extraction according to claim 3 is characterized in that: The point cloud keyframes are selected by taking one frame every ten frames of point cloud data or one frame every meter or when the heading angle changes by more than ten degrees.

5. The vehicle position recognition method based on point cloud descriptor extraction according to claim 1 is characterized in that: The point cloud descriptor pre-training model is obtained by training the model through the point cloud depth map and the point cloud bird's-eye view map.

6. The vehicle position recognition method based on point cloud descriptor extraction according to claim 5 is characterized in that: The laser point cloud is converted into a two-dimensional range image by spherical projection to obtain a point cloud depth map.

7. The vehicle position recognition method based on point cloud descriptor extraction according to claim 5 is characterized in that: Divide a frame of laser point cloud according to ScanContext, select the maximum height value of the point cloud in each curved trapezoidal cylinder for filling, and generate a bird's-eye view of the point cloud.

8. The vehicle position recognition method based on point cloud descriptor extraction according to claim 5 is characterized in that: A spatial feature is generated using the spatial information of a single-frame point cloud depth map and a point cloud bird's-eye view, and a layer of convolutional neural network is used to perform feature extraction again to obtain a point cloud global descriptor.

9. A vehicle position recognition system based on point cloud descriptor extraction, characterized in that: Data acquisition preprocessing module and recognition module; The data acquisition preprocessing module performs denoising on the acquired vehicle laser point cloud data; The recognition module uses the point cloud descriptor pre-training model to extract the point cloud global descriptor from the denoised laser point cloud data, and obtains the vehicle position based on the obtained point cloud global descriptor.

10. The vehicle position recognition method based on point cloud descriptor extraction according to claim 9, characterized in that: Specifically, the data acquisition preprocessing module specifically uses laser radar; point cloud key frames are extracted from the laser point cloud data of the vehicle to be positioned, and several frames with the highest similarity are obtained as candidate frames by comparing the similarities of global descriptors corresponding to the acquired point cloud key frames, and the candidate frames are used to achieve position recognition or repositioning.

Images