CN117765509A

CN117765509A - Guardrail detection method and device, electronic equipment and storage medium

Info

Publication number: CN117765509A
Application number: CN202311753717.9A
Authority: CN
Inventors: 彭靖玥; 张丹
Original assignee: Uisee Technologies Beijing Co Ltd
Current assignee: Uisee Technologies Beijing Co Ltd
Priority date: 2023-12-19
Filing date: 2023-12-19
Publication date: 2024-03-26

Abstract

The embodiment of the disclosure discloses a guardrail detection method, device, electronic equipment and storage medium, the method is used for obtaining an initial detection guardrail by clustering point cloud data acquired by a vehicle, determining point guardrails to be complemented from the initial detection guardrail, determining nodes to be complemented in the direction of the complemented points of each point guardrail, judging whether real points corresponding to the nodes to be complemented exist in a grid map corresponding to the point cloud data, judging whether real points exist in a detection area corresponding to the nodes to be complemented in the grid map if the real points do not exist in the grid map, updating the transverse coordinates of the nodes to be complemented, adding the nodes to be complemented in the direction of the complemented points guardrails, updating the point guardrails to be complemented, and circulating the steps until the length of the point guardrails to be complemented reaches a preset length threshold value, so as to obtain the final detection guardrail, thereby realizing the detection of whether the guardrails are welted and the correction of the non-welted guardrails, improving the detection accuracy of the guardrails, and avoiding the condition of guardrail detection as far as possible.

Description

Guardrail detection method and device, electronic equipment and storage medium

Technical Field

The disclosure relates to the technical field of automatic driving environment sensing, in particular to a guardrail detection method, a guardrail detection device, electronic equipment and a storage medium.

Background

The existing guardrail detection scheme mainly comprises three types: the first is a traditional detection mode which depends on laser point cloud clustering and related filtering algorithms; the second is a detection mode which relies on deep learning and mass data training models; the last is a detection mode combining a traditional algorithm and deep learning.

The traditional target detection by utilizing laser point cloud data clustering is a method for comparing information based on points, and one of the defects of the method is low speed because all point data are used; secondly, the number of the point clouds depends on the number of lines of the radar, and the higher the number of the lines is, the more expensive the radar is; thirdly, the forming conditions and the filtering conditions of guardrails in various clustering algorithms are difficult to meet the requirement of various guardrail shapes in reality, so that a plurality of target edges are mistakenly detected to form guardrails, and the guardrails are likely to be missed in the case of being blocked.

The defects of detection schemes using deep learning are also obvious, and one is that various schemes for route selection have advantages and disadvantages; the camera is used for detecting the extremely good weather conditions of the guardrails, and in various cases, the guardrails displayed by the camera have similar characteristics to the lane lines, so that the picture detection can falsely detect the lane lines as the guardrails; secondly, the deep learning is utilized to relate to a large amount of data annotation, the problem of false detection caused by the omission of the guardrail still cannot be completely solved under the extremely high cost, and the model is difficult to update in a rapid iteration in a short period.

The detection scheme combining the traditional algorithm with the deep learning can also solve the problem of false detection of the guard rail missing detection.

Disclosure of Invention

To solve or at least partially solve the above technical problems, embodiments of the present disclosure provide a guardrail detection method, apparatus, electronic device, and storage medium.

In a first aspect, an embodiment of the present disclosure provides a guardrail detection method, including:

clustering point cloud data acquired by a vehicle, obtaining initial detection guardrails according to clustering results, and determining point guardrails to be complemented in all the initial detection guardrails;

determining nodes to be added in the point supplementing direction of each point to be supplemented guardrail, judging whether real points corresponding to the nodes to be added exist in a grid map corresponding to the point cloud data, and judging whether the real points exist in a detection area corresponding to the nodes to be added in the grid map if the real points do not exist;

if yes, updating the abscissa of the node to be added, adding the node to be added to the point supplementing direction of the point supplementing guardrail to update the point supplementing guardrail, and returning to execute the step of determining the node to be added in the point supplementing direction of the point supplementing guardrail until the length of the point supplementing guardrail reaches a preset length threshold value, so as to obtain the final detection guardrail.

In a second aspect, embodiments of the present disclosure also provide a guardrail detection device, the device comprising:

the point supplementing guardrail determining module is used for clustering point cloud data acquired by the vehicle, obtaining initial detection guardrails according to clustering results, and determining point guardrails to be supplemented in all the initial detection guardrails;

the node judging module is used for determining nodes to be added in the point supplementing direction of each point to be supplemented guardrail, judging whether real points corresponding to the nodes to be added exist in a grid map corresponding to the point cloud data, and judging whether the real points exist in a detection area corresponding to the nodes to be added in the grid map if the real points do not exist;

and the node adding module is used for updating the abscissa of the node to be added if the node to be added is the node to be added, adding the node to be added to the point supplementing direction of the point supplementing guardrail so as to update the point supplementing guardrail, and returning to execute the step of determining the node to be added in the point supplementing direction of the point supplementing guardrail until the length of the point supplementing guardrail reaches a preset length threshold value, so that the final detection guardrail is obtained.

In a third aspect, embodiments of the present disclosure further provide an electronic device, including: one or more processors; a storage means for storing one or more programs; the one or more programs, when executed by the one or more processors, cause the one or more processors to implement the guardrail detection method as described above.

In a fourth aspect, the disclosed embodiments also provide a computer readable storage medium having stored thereon a computer program which when executed by a processor implements the guardrail detection method as described above.

The guardrail detection method provided by the embodiment of the disclosure obtains initial detection guardrails by clustering point cloud data acquired by vehicles, determines point guardrails to be complemented from the initial detection guardrails, determines nodes to be complemented in the direction of the point to be complemented for each point guardrail to be complemented, judges whether real points corresponding to the nodes to be complemented exist in a grid map corresponding to the point cloud data, judges whether real points exist in a detection area corresponding to the nodes to be complemented in the grid map if the real points do not exist, updates the abscissa of the nodes to be complemented if the real points do not exist in the detection area corresponding to the nodes to be complemented in the grid map, adds the nodes to be complemented to the direction of the point guardrail to be complemented to update the point guardrail to be complemented, and the step of determining the node to be added in the direction of the point to be complemented is carried out again until the length of the guardrail at the point to be complemented reaches the preset length threshold value, so that the final detection guardrail is obtained, the detection of whether the guardrail is welted or not and the correction of the guardrail without the welted guardrail are realized, the problems that clustered guardrails are not welted and the guardrail extension line does not have points are solved, the detection error caused by adopting the straight line fitting guardrail is reduced, the detection accuracy of the guardrail is improved, the condition of false detection of the guardrail is avoided as much as possible, and the times that vehicles cannot be automatically driven and are forced to take over the vehicles due to the fact that no real points exist on the guardrail are effectively reduced.

Drawings

The above and other features, advantages, and aspects of embodiments of the present disclosure will become more apparent by reference to the following detailed description when taken in conjunction with the accompanying drawings. The same or similar reference numbers will be used throughout the drawings to refer to the same or like elements. It should be understood that the figures are schematic and that elements and components are not necessarily drawn to scale.

FIG. 1 is a flow chart of a guardrail detection method in an embodiment of the present disclosure;

FIG. 2 is a schematic representation of a cluster in an embodiment of the disclosure;

FIG. 3 is a schematic illustration of an initial detection fence after fitting in an embodiment of the present disclosure;

FIG. 4 is a schematic structural view of a guardrail detection device according to an embodiment of the present disclosure;

fig. 5 is a schematic structural diagram of an electronic device in an embodiment of the disclosure.

Detailed Description

Embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. While certain embodiments of the present disclosure have been shown in the accompanying drawings, it is to be understood that the present disclosure may be embodied in various forms and should not be construed as limited to the embodiments set forth herein, but are provided to provide a more thorough and complete understanding of the present disclosure. It should be understood that the drawings and embodiments of the present disclosure are for illustration purposes only and are not intended to limit the scope of the present disclosure.

It should be noted that the terms "first," "second," and the like in this disclosure are merely used to distinguish between different devices, modules, or units and are not used to define an order or interdependence of functions performed by the devices, modules, or units.

The names of messages or information interacted between the various devices in the embodiments of the present disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

Before describing the guardrail detection method provided by the embodiment of the present disclosure in detail, a description is given of a technical problem solved by the method. Illustratively, the prior art is as follows:

patent 1 (CN 111881752 a): according to the scheme, the guard rail is detected by utilizing a target detection technology and a 3D point cloud segmentation technology in deep learning. Firstly, detecting a guardrail target on image data, and further projecting a detection coordinate result back into laser point cloud data; secondly, carrying out semantic segmentation on laser point cloud data by using a point cloud segmentation algorithm, and extracting the boundary of the guardrail; and finally, the detected result and the segmented result are in one-to-one correspondence, so that a high-precision guardrail target can be obtained.

Patent 2 (CN 109254289 a): the guardrail detection method comprises the steps of firstly carrying out continuous frame stationary obstacle clustering based on millimeter wave point clouds after filtering to obtain adjacent obstacles in a certain range around a vehicle, fitting the adjacent obstacles into straight lines, calculating the distance between the adjacent obstacles and the adjacent fitted straight lines in a subsequent circulation mode, setting a plurality of distance thresholds or clustering thresholds used for filtering, representing the obstacles smaller than the thresholds as fitted obstacles, namely suspected guardrail targets, and finally carrying out curve or straight line fitting on the fitted obstacle points to obtain point set coordinates of the guardrail obstacles, thereby realizing guardrail detection.

Patent 3 (KR 102423781B 1): the method is characterized in that a plane fitting algorithm or a ground detection algorithm based on a map is adopted to filter ground points in laser point clouds, a self-vehicle sensor is used as a coordinate system, residual point cloud data are mapped into the voxel clouds, non-empty voxels are sequenced according to coordinates, whether the voxels are perpendicular to the ground and parallel to vehicles are checked in sequence, whether the voxels are likely to be part of a guardrail or not is judged, and curve fitting is further carried out on the selected voxels according to preset thresholds such as direction consistency, distance and length, so that guardrail detection is realized.

Patent 4 (CN 116129359 a) is a method for detecting guardrails by combining an edge detection algorithm with a deep learning algorithm, firstly, edge image detection is performed in a video stream with guardrails by using a Canny algorithm, further, dimension stitching is performed between the edge image detection and RBG images, and fine tuning is performed on a network of Resnet50, so that a model for detecting the guardrails on the edges is constructed.

In summary, the existing guardrail detection schemes are mainly divided into three types; the first is the traditional detection mode which completely depends on laser point cloud clustering and related filtering algorithms, such as patent 2 and patent 3; the second is a detection mode which simply depends on deep learning and mass data training models, as in patent 1; the last is a detection scheme combining a traditional algorithm with deep learning, as in patent 4.

The traditional target detection by utilizing laser point cloud data clustering is a method for comparing information based on points, and one of the defects of the method is low speed because all point data are used, and although patent 3 divides the point data into voxel clouds, the division process takes a long time; secondly, the number of the point clouds depends on the line number of the radar, the higher the line number is, the more points are reflected by the target, the better the point cloud clustering effect is, the higher the accuracy of guardrail detection is, and the higher the line number is, the more expensive the radar price is; thirdly, the forming condition and the filtering condition of guardrails in various clustering algorithms are difficult to meet various guardrail shapes in reality, and as the generation requirement of the guardrails is that point clouds are parallel to vehicles in consideration of patent 3, the omission ratio of guardrail detection in the process of turning of vehicles can be greatly increased, and the distance between candidate point sets and the number of point set points are only considered in the process of judging whether the guardrails are guardrail objects or not in patent 2, the continuity of the point sets and the information height difference in the clustering process of obstacles are not utilized to filter non-guardrail data, so that a plurality of target edges are mistakenly detected to form guardrails, and the guardrails are likely to be omitted under the condition of shielding.

The defects of the detection scheme using deep learning are obvious, one is that the route is selected by using pure visual detection (such as patent 4) or three-dimensional point cloud detection, or combining the visual detection with the point cloud detection (such as patent 1), and various schemes have advantages; the camera is used for detecting the extremely good weather conditions of the guardrails, and because the camera is influenced by illumination changes, and in various cases, the guardrails displayed by the camera have similar characteristics with the lane lines, the picture detection can falsely detect the lane lines as the guardrails; in the detection method combining the vision and the point cloud in the patent 1, errors exist in the alignment mode of the two-dimensional data and the three-dimensional data, and how to integrate the view angle information of the radar and the view angle information of the camera, so that the information such as textures, colors and the like with accurate positioning and fine granularity is originally one of the difficult problems to overcome; secondly, the deep learning is utilized to necessarily involve a large amount of data annotation, under the condition that the model structure is unchanged, the quantity of training data is directly proportional to the accuracy of model detection and is also directly proportional to the training time length, the problem of false detection caused by the missing detection of the guardrail still cannot be completely solved under the condition of extremely high cost, and the model is difficult to update in a rapid iteration in a short period.

Therefore, the above-mentioned scheme all can bring the guardrail and leak and examine the false detection problem, in order to solve this technical problem, this disclosed embodiment provides a guardrail detection method, not rely on one of them scheme entirely, can realize whether to the guardrail to the detection of welting and to the correction of not welting guardrail, solve the problem that does not have the point on the guardrail of cluster is not welting and the guardrail extension line, reduce the detection error that adopts straight line fitting guardrail to bring, when improving guardrail detection accuracy, avoid appearing the condition of guardrail false detection as far as possible.

Fig. 1 is a flowchart of a guardrail detection method in an embodiment of the present disclosure. The method may be performed by a guardrail detection device, which may be implemented in software and/or hardware, which may be configured in an electronic device. As shown in fig. 1, the method specifically may include the following steps:

s110, clustering the point cloud data acquired by the vehicle, obtaining initial detection guardrails according to a clustering result, and determining the guardrails to be complemented with points in all the initial detection guardrails.

In the disclosed embodiments, the point cloud data may be data collected by a laser sensor on the vehicle. Specifically, aiming at the point cloud data collected by the vehicle, voxel grid clustering processing can be carried out on the point cloud data so as to obtain all initial detection guardrails through clustering; wherein, the voxel grid clustering process can be understood as mapping the point cloud data into grids for clustering.

In a specific embodiment, clustering point cloud data collected by a vehicle, and obtaining an initial detection guardrail according to a clustering result, wherein the method comprises the following steps:

step 111, mapping the point cloud data acquired by the vehicle to each grid divided in advance to obtain a grid map corresponding to the point cloud data;

step 112, determining whether each grid in the grid map is valid, and determining a connected domain based on each valid grid in the grid map;

step 113, judging whether the number of real points in each effective grid forming the connected domain is larger than a preset number threshold value or not according to each connected domain, and whether the difference value between the highest heights of the effective grids is smaller than a preset height difference or not, if yes, determining a clustering target based on the connected domain;

and 114, determining guardrail clustering targets in all the clustering targets, and determining an initial detection guardrail based on all the guardrail clustering targets, wherein the highest height is the height average value of all the real points in the grid within the first height range.

In step 111, a vehicle body coordinate system may be previously constructed according to a vehicle center, and point cloud data acquired in a sensing range of the vehicle may be mapped to each grid to obtain a corresponding grid map. Wherein each grid may be a grid within a pre-divided area of size w×h.

Further, in step 112, it may be determined whether each grid is valid by the real point falling into each grid, and the grid is valid, which may be understood as the grid is in an unpermeable state, and the grid is invalid, which may be understood as the grid is in a passable state.

For each grid in the grid map, whether it is valid or not can be judged by the information recorded in the grid. Wherein each grid in the grid map can record information such as the number of real points falling into the grid, the height of the highest real point in the grid, the height of the lowest real point in the grid, the height difference between the highest height and the relative height, and the like.

For step 112, optionally, determining whether each grid in the grid map is valid includes:

for each grid in the grid map, if the height difference between the highest height and the relative height of the grid is greater than a height difference threshold value, determining the grid as an effective grid; wherein the relative height is the height average of all the real points in the grid that lie within the second height range.

Wherein the minimum value of the first height range is not less than the maximum value in the second height range. By way of example, the second height range may be [0,1.2], in m, the minimum value within the first height range may be 1.5m, or any value greater than 1.5m, e.g., the first height range may be 1.5, infinity ].

Specifically, the height average of all the real points in the first height range in each grid may be calculated as the highest height, and the height average of all the real points in the second height range in each grid may be calculated as the relative height. Further, if the height difference between the highest height and the relative height of the grid is greater than the height difference threshold, it indicates that the grid is valid, i.e., the grid is a valid grid.

In the embodiment of the disclosure, in order to improve the accuracy of judging whether the grid is effective, the secondary judgment can be further performed by combining other information. Optionally, for step 112, after determining whether each grid in the grid map is valid, further comprising:

judging whether the number of the real points in the effective grids is smaller than the set number or not according to each effective grid, if yes, updating the effective grids into invalid grids under the condition that the number of the first target points in the effective grids meets a first proportion;

otherwise, updating the effective grid into an ineffective grid under the condition that the number of second target points in the effective grid meets a second proportion; the first target point is a real point with intensity larger than a preset intensity threshold value, and the second target point is a real point with height larger than the preset guardrail height.

The set number may be a minimum number of real points preset to determine whether the grid is valid. The first ratio may be a ratio of a preset first target point in all real points of the grid, such as 1/2, and the second ratio may be a ratio of a preset second target point in all real points of the grid, such as 2/3. A true point may be understood as a point in the point cloud data.

Specifically, each effective grid can be traversed in a circulating manner, if the number of the real points in the effective grid is smaller than the set number, the number of the real points with the intensity larger than the preset intensity threshold in the effective grid can be counted, and the effective grid is updated to be an ineffective grid under the condition that the number meets the first proportion; if the number of the real points in the effective grid is not less than the set number, the number of the real points with the height greater than the preset guardrail height in the effective grid can be counted, and the effective grid is updated to be an ineffective grid under the condition that the number meets the second proportion. Through the mode, whether the grids are effective or not is judged secondarily, and the grids of the aerial noise blocks can be further updated to invalid grids, so that the clustering accuracy is further improved.

After determining each effective grid in the grid map, further, clustering may be performed in conjunction with all the effective grids. In the embodiment of the present disclosure, considering that if the difference between the highest heights of two effective grids is greater than a certain value, it means that the two effective grids may belong to two targets, respectively, should not be clustered into one target, in which case even if the two effective grids are connected, the two effective grids should not be taken as a part of the same target.

Therefore, in order to avoid the situation that the pedestrian and the guardrail are clustered into the same target, in the step 112, the connected domain may be determined according to each effective grid in the grid map, and then in step 113, it is determined whether the connected domain can be clustered into one target by combining the number of the real points of the effective grid in the connected domain and the difference value between the highest heights. The connected region is understood to be a connected region formed by adjacent real points.

Specifically, in step 113, for each connected domain, when the number of real points in each effective grid constituting the connected domain is greater than a preset number threshold, and the difference between the highest heights of the effective grids is smaller than a preset height difference, the connected domain may be determined as a clustering target.

For example, fig. 2 is a schematic clustering diagram in an embodiment of the present disclosure, where three grids on the right may form a connected domain, the number of real points of the three grids on the right meets a preset number threshold, and a difference between the highest heights of adjacent grids is smaller than a preset height difference, so that the three grids on the right may be clustered into one clustering target. Although the second grid and the third grid can form a connected domain, the second grid can be clustered into a cluster target independently because the difference between the highest heights of the second grid and the third grid is large.

In the embodiment, the objects with unobvious boundaries are split through the height difference between different grids, so that the detection rate of pedestrians at the side of the guardrail can be effectively improved, the accuracy of the guardrail is improved, and the risk in the automatic driving and driving process is reduced.

After each clustered object is obtained in step 113, in step 114, the guardrail clustered objects therein may be obtained by screening each clustered object. Wherein, the guardrail clustering targets can be clustering targets with the classification of guardrails. For example, the distance between each real point in the clustered target and the road network boundary can be calculated, the distance is compared with a preset distance threshold, and if the maximum distance is smaller than the distance threshold, the clustered target can be determined to be a guardrail clustered target.

Furthermore, the initial detection guardrails can be obtained by fitting all the guardrail clustering targets. In embodiments of the present disclosure, the fitted initial detection fence may exist in the form of a line; the initial detection fence may be comprised of at least one fence segment, each fence segment may include at least one node.

Illustratively, fig. 3 is a schematic diagram of an initial detection fence after fitting in an embodiment of the present disclosure. For the guard bar line in fig. 3 (a close fitting straight line, i.e. an initial detection guard bar), it may be made up of a plurality of point sets, each of which may be understood as a guard bar segment.

Considering that in the initial detection fence after fitting, there may be invalid fence segments, i.e. fence segments that are not present in the actual point cloud data, or that in the initial detection fence after fitting, the distance between adjacent fence segments may be too long. Therefore, after the initial detection guardrails are obtained, invalid guardrail section removal and splitting treatment can be performed on each initial detection guardrail.

For step 114, optionally, after obtaining the initial detection fence according to the clustering result, the method further includes the following steps:

step 115, judging whether each guardrail section in the initial detection guardrails is effective or not according to each initial detection guardrail, and removing the ineffective guardrail section from the initial detection guardrails;

116, eliminating the initial detection guardrail under the condition that the number of guardrail sections in the initial detection guardrail is zero, and judging whether the initial detection guardrail meets the splitting condition or not based on the distance between adjacent guardrail sections under the condition that the number of guardrail sections in the initial detection guardrail is more than one;

and 117, if so, splitting the initial detection guardrails, and obtaining a plurality of new initial detection guardrails based on the splitting result.

In step 115, it may be determined whether each guardrail segment is valid in combination with the nodes in each guardrail segment. Wherein nodes in a guardrail segment can be understood as points constituting the guardrail segment, and mapping to a grid map can correspond to a grid. In a specific embodiment, determining whether each of the initial detection guard rail segments is valid comprises:

in the case that the number of nodes in the guardrail section is one, determining that the guardrail section is invalid;

judging whether all nodes in the guardrail section are effective under the condition that the number of the nodes in the guardrail section is two, if so, determining that the guardrail section is ineffective under the condition that the distance between the nodes in the guardrail section is larger than a preset distance;

judging whether the first node and the tail node in the guardrail section are valid or not under the condition that the number of nodes in the guardrail section is greater than two, if so, determining whether a new node between any two adjacent nodes in the guardrail section is valid or not to obtain the number of effective nodes corresponding to the guardrail section, and if the number of the effective nodes does not meet the preset node proportion, determining that the guardrail section is invalid; wherein, the node effectively means that corresponding real points exist in the grid map.

That is, if there is only one node in the guardrail section, it can be determined directly that the guardrail section is invalid. If the guardrail section comprises two nodes, whether each node is effective or not can be judged firstly, namely whether each node has corresponding real points in the grid or not is judged, if the number of the effective nodes is smaller than two, the guardrail section is determined to be ineffective, if the number of the effective nodes is equal to two, whether the distance between the two effective nodes is larger than a preset distance is further judged, if the distance between the two effective nodes is larger than the preset distance, the guardrail section can be determined to be ineffective, otherwise, the guardrail section can be determined to be effective.

If the number of nodes contained in the guardrail section is greater than two, it can be determined whether the head node and the tail node are valid. If the first node and the tail node are both valid, judging whether a new node between two adjacent nodes (namely, the middle point between the two nodes) is valid from the first node or not until the tail node is judged, counting the number of the valid nodes at the moment, if the number of the valid nodes meets the preset node proportion, determining that the guardrail section is valid, otherwise, determining that the guardrail section is invalid. Wherein the preset node ratio may be 1/3.

In one example, determining whether each of the initial detection guard rail segments is valid further comprises:

If the number of nodes in the guardrail section is greater than two, if the head node and the tail node in the guardrail section are effective, judging whether each node is effective from the head node to the rear, obtaining the number of effective nodes, and if the number of the effective nodes does not meet the preset node proportion, determining that the guardrail section is ineffective;

if the tail nodes in the guardrail section are effective and the head nodes are not effective, starting from the tail nodes, judging whether each node is effective or not, obtaining the number of effective nodes, and determining that the guardrail section is not effective under the condition that the number of the effective nodes is smaller than the number of preset nodes.

If the head node is effective and the tail node is ineffective, judging whether each node is effective or not in sequence from the head node to the back, wherein the time complexity is O (n), n represents the number of the nodes so as to find the effective tail node in the guardrail section, judging whether the effective node number between the effective head node and the effective tail node meets the preset node proportion or not, if yes, determining that the guardrail section is effective, and if not, determining that the guardrail section is ineffective.

If the head node is invalid and the tail node is valid, sequentially judging whether each node is valid from the tail node, wherein the time complexity is O (n), n represents the number of the nodes so as to find the effective head node in the guardrail section, further judging whether the effective node number between the effective head node and the tail node meets the preset node proportion, if yes, determining that the guardrail section is valid, and if not, determining that the guardrail section is invalid.

For each initial detection fence, the fence segments in which it is determined to be invalid may be culled. After the invalid guard rail segment in the initial detection guard rail is removed, further, in the step 116, in the case that the number of guard rail segments in the initial detection guard rail is zero, the initial detection guard rail is indeed invalid and the initial detection guard rail is removed.

Under the condition that the number of guardrail sections in the initial detection guardrail is a plurality of, judging whether the distance between adjacent guardrail sections is greater than a preset splitting distance, if so, splitting the initial detection guardrail, and taking the split guardrail as a new initial detection guardrail.

It can be understood that if the number of guardrail segments in the initial detection guardrail is m, the maximum splitting frequency is m-1, that is, the initial detection guardrail formed by m guardrail segments can be split into m new initial detection guardrails at most. After the initial detection guardrails are split, the number of the guardrail sections in each new initial detection guardrail, the head node, the tail node and other information of the guardrail sections can be updated.

Furthermore, the guardrail with the points to be complemented can be screened out from all the initial detection guardrails. The guardrail to be complemented can be an initial detection guardrail of which the detection range does not meet the preset range. The guardrail detection range may be a guardrail range detected in the traveling direction under a vehicle body coordinate system, and the preset range may be a preset standard range of the guardrail detected in the traveling direction, for example, -20m to 100m.

In a specific embodiment, determining the point to be complemented guard bar in all initial detection guard bars comprises:

among all initial detection guardrails, taking the initial detection guardrails of which the detection range of the guardrails does not meet the preset range as guardrails of points to be complemented; judging whether the transverse coordinate difference between any two point guardrails to be complemented exceeds a preset transverse threshold value, if so, eliminating one point guardrail to be complemented; the preset transverse threshold value is half of the preset guardrail width.

That is, an initial detection guardrail whose guardrail detection range does not satisfy a preset range may be used as the guardrail to be complemented, for example, an initial detection guardrail whose guardrail detection range is-20 m to 80m, 0m to 100m, 20m to 80m, etc. may be used as the guardrail to be complemented.

Furthermore, in order to avoid repeated point filling of the initial detection guardrails with relatively short transverse distance, one of the guardrails to be filled with points can be removed under the condition that the transverse coordinate difference between the two guardrails to be filled with points does not exceed a preset transverse threshold value, so that repeated point filling is avoided.

S120, determining nodes to be added in the complementary point direction of each point to be added guardrail, judging whether real points corresponding to the nodes to be added exist in the grid map corresponding to the point cloud data, and judging whether the real points exist in the detection area corresponding to the nodes to be added in the grid map if the real points do not exist.

Specifically, for each point to be complemented guardrail, the node to be added can be determined in the direction of the complement point of the point to be complemented guardrail. The point-supplementing direction may be a tail of the point-supplementing guardrail, specifically may be a tangent line at a tail node of the point-supplementing guardrail, or the point-supplementing direction may be a head of the point-supplementing guardrail, specifically may be a tangent line at a head node of the point-supplementing guardrail.

In the embodiment of the disclosure, the point compensating direction of the point to be compensated guardrail can be determined according to the difference between the guardrail detection range and the preset range of the point to be compensated guardrail.

Specifically, the node to be added may be determined in the direction of the complement point of the guardrail to be complemented, where the node to be added may be a node to be added to the guardrail to be complemented, i.e. a node increased along the direction of the complement point of the guardrail to be complemented.

Further, it may be determined whether there are real points in the grid map corresponding to the nodes to be added. The real points corresponding to the nodes to be added can be real points in the grid corresponding to the nodes to be added in the grid map. If true points in the grid corresponding to the nodes to be added exist in the grid map, the nodes to be added can be added to the point supplementing direction of the point supplementing guardrail to update the point supplementing guardrail, and then the step of determining the nodes to be added in the point supplementing direction of the point supplementing guardrail is performed again.

If the real points corresponding to the nodes to be added do not exist in the grid map, further judging whether the real points exist in the detection area corresponding to the nodes to be added in the grid map. The detection area corresponding to the node to be added may be a neighborhood of a grid corresponding to the node to be added in the grid map, for example, an area determined by extending a preset length (for example, 20 m) to a complementary point direction at a grid corresponding to the node to be added and combining a preset width (for example, a width of 5 grids).

And S130, if so, updating the abscissa of the node to be added, adding the node to be added to the point supplementing direction of the point supplementing guardrail to update the point supplementing guardrail, and returning to execute the step of determining the node to be added in the point supplementing direction of the point supplementing guardrail until the length of the point supplementing guardrail reaches a preset length threshold value, thereby obtaining the final detection guardrail.

Specifically, if there is a real point in the detection area corresponding to the node to be added in the grid map, the abscissa of the node to be added may be updated to change the abscissa of the node to be added, and the abscissa of the node to be added is bound to a position according to the real point in the detection area, where the real condition in the point cloud data is met.

In a specific embodiment, updating the abscissa of the node to be added includes:

determining a target grid closest to the node to be added in all grids with the true points in the detection area, or taking the grid with the maximum number of the true points in the detection area as the target grid;

updating the abscissa of the node to be added based on the center of the target grid.

That is, the abscissa of the node to be added may be bound to the center of the target mesh closest to the node to be added within the detection area, or the abscissa of the node to be added may be bound to the center of the mesh having the largest number of real points within the detection area.

The incremental guardrail with points to be repaired can be more welted by updating the abscissa of the nodes to be added, so that the condition of the guardrail that the guardrail is missed in detection is reduced.

After updating the abscissa of the node to be added, further, the node to be added may be added to the direction of the point to be added guardrail to update the point to be added guardrail, so as to realize the point to be added of the point to be added guardrail, and the step S120 is returned to continue until the length of the point to be added guardrail reaches the preset length threshold, and the point to be added guardrail at this time is used as the final detection guardrail.

Of course, if no real point exists in the detection area corresponding to the node to be added in the grid map, the node to be added does not need to be added to the point supplementing direction of the point supplementing guardrail, and the point supplementing of the point supplementing guardrail can be stopped.

In the embodiment, through the two-dimensional grid map, the clustered point guardrails to be complemented are further checked and adjusted, the problem that the guardrails are not welted is optimized, the detection error caused by fitting the guardrails is reduced, the rebinding of the node positions can be carried out in the checking process, the accuracy of the guardrails is improved, meanwhile, the situation that the false detection of the guardrails still occurs near the real points corresponding to the guardrails in the point cloud data is avoided as much as possible, and the frequency of forced manual taking over of the vehicles caused by the fact that no real points exist on the guardrails but the self-vehicle cannot automatically drive can be effectively reduced.

In the embodiment of the invention, since the initial detection guard rail is obtained based on the guard rail clustering target fitting, the situation that the clustering target actually being the guard rail is judged to be a non-guard rail clustering target possibly exists in the screening process of the guard rail clustering target, for example, the pedestrian and the guard rail are separated into one clustering target due to the fact that the pedestrian and the guard rail are close to each other, and therefore the pedestrian and the guard rail are filtered out due to the fact that the width of the guard rail is not met in the subsequent screening process. Therefore, the guard bars which are missed in the clustering process can be determined by combining a pre-trained detection model aiming at the non-guard bar clustering targets obtained in the clustering process.

In a specific implementation manner, the method provided by the embodiment of the disclosure further includes: determining classification detection results and classification confidence of all real points in the point cloud data based on a pre-trained detection model, wherein the classification detection results are pedestrians, vehicles and other objects;

after the guardrail clustering targets are determined in all the clustering targets, the method provided by the embodiment of the disclosure further comprises the following steps:

step 141, judging whether the classification detection results of the head node and the tail node of each non-guardrail clustering target are other objects or not according to each non-guardrail clustering target;

step 142, if yes, taking the first node of the non-guardrail clustering target as the current node, determining a new node in the neighborhood range of the current node according to the grid with the maximum number of real points and the highest classification confidence, taking the new node as the current node again, and returning to the step of determining the new node in the neighborhood range of the current node until the current node is the tail node of the non-guardrail clustering target, so as to obtain the guardrail missing detection target;

and step 143, determining a final detection guardrail based on all guardrail detection omission targets.

To reduce the dependency of guardrail detection on the depth model, it may be desirable to choose whether the depth model needs to be trained to have the ability to identify the type of guardrail. In order to improve the decoupling performance of the traditional clustering and the model, in the embodiment of the disclosure, the depth model does not need to be trained to have the capability of identifying the type of the guardrail, the result output by the depth model is used as auxiliary information, and the non-guardrail clustering targets determined in the clustering process are combined to detect whether the guardrail is missed.

Thus, in the embodiment of the disclosure, the depth model has the capability of identifying two categories of pedestrians and vehicles, and for other categories besides pedestrians and vehicles (including motor vehicles and non-motor vehicles), the depth model does not need to be concerned about the specifically subdivided category, and the description of other objects is unified.

Specifically, the pre-trained detection model may be a 3D semantic segmentation model or a 3D point cloud target detection model. Illustratively, taking a 3D semantic segmentation model as an example, a pre-trained detection model may determine classification detection results (i.e., class information) and classification confidence of each real point in the point cloud data. The semantic map corresponding to the point cloud data can be constructed according to the classification detection result and the classification confidence coefficient of each real point, and then the classification confidence coefficient in the semantic map is removed from being larger than the confidence coefficient threshold value and is the classification detection result of pedestrians and vehicles through the preset confidence coefficient threshold value.

Further, in step 141, for each non-guardrail clustering target, according to the related information (such as the coordinates of the start line and the end line, the number of grids, the height information of the target, etc.) recorded in the clustering process, the non-guardrail clustering targets may be matched with the semantic map, so as to correspond the head node and the tail node of the non-guardrail clustering targets to the semantic map, and obtain the classification detection results of the head node and the tail node.

Specifically, if the classification detection results of the first node and the tail node of the non-guardrail clustering target are other objects, the first node can be used as the current node from the first node of the non-guardrail clustering target, according to the static and continuous characteristics of the guardrails, the grid with the highest number of real points and highest classification confidence is searched in the neighborhood range (such as the range of about 30 degrees) of the current node and used as the new node, the new node is further used as the current node again, the steps are executed again, and the cycle is performed until the tail node of the non-guardrail clustering target is reached, so that a guardrail missing detection target can be selected from the non-guardrail clustering target which is missed.

Furthermore, fitting treatment can be further carried out on all the guard bar missed detection targets to obtain at least one final detection guard bar so as to reduce the possibility of guard bar missed detection.

Through the embodiment, the information of the point cloud data is fully utilized under the assistance of the depth model, the accuracy of guardrail detection is effectively improved, the omission rate is reduced, the accuracy of guardrail detection can be ensured, and due to the fact that pedestrians and vehicles are not included in new nodes searched based on the semantic map, even if certain errors exist in guardrail detection, automatic driving decision cannot be influenced.

In addition, a model with the guardrail detection capability does not need to be specially trained, and the accuracy of guardrail detection is improved under the condition that the multiplexing rate of other task models is improved; the traditional cluster detection and depth model detection schemes are decoupled, the optimization can be performed based on any detection method of other tasks, and the occupied amount of calculation resources is small.

The guardrail detection method provided by the embodiment obtains the initial detection guardrails by clustering the point cloud data collected by the vehicle, determines the point guardrails to be complemented from the initial detection guardrails, determines the nodes to be complemented in the direction of the point to be complemented for each point guardrail, judges whether the real points corresponding to the nodes to be complemented exist in the grid map corresponding to the point cloud data, judges whether the real points exist in the detection area corresponding to the nodes to be complemented in the grid map if the real points do not exist in the grid map, updates the abscissa of the nodes to be complemented, adds the nodes to be complemented in the direction of the point to be complemented in order to update the point guardrails to be complemented, and the step of determining the node to be added in the direction of the point to be complemented is carried out again until the length of the guardrail at the point to be complemented reaches the preset length threshold value, so that the final detection guardrail is obtained, the detection of whether the guardrail is welted or not and the correction of the guardrail without the welted guardrail are realized, the problems that clustered guardrails are not welted and the guardrail extension line does not have points are solved, the detection error caused by adopting the straight line fitting guardrail is reduced, the detection accuracy of the guardrail is improved, the condition of false detection of the guardrail is avoided as much as possible, and the times that vehicles cannot be automatically driven and are forced to take over the vehicles due to the fact that no real points exist on the guardrail are effectively reduced.

Fig. 4 is a schematic structural view of a guardrail detecting device according to an embodiment of the present disclosure. As shown in fig. 4: the device comprises: the point supplement guardrail determination module 410, the node judgment module 420, and the node addition module 430.

The point-supplementing guardrail determining module 410 is configured to cluster point cloud data collected by a vehicle, obtain initial detection guardrails according to a clustering result, and determine point-supplementing guardrails to be complemented in all the initial detection guardrails;

the node judging module 420 is configured to determine, for each point guardrail to be complemented, a node to be added in a point complementing direction of the point guardrail to be complemented, judge whether a real point corresponding to the node to be added exists in a grid map corresponding to the point cloud data, and if not, judge whether a real point exists in a detection area corresponding to the node to be added in the grid map;

and the node adding module 430 is configured to update the abscissa of the node to be added if yes, add the node to be added to the point-supplementing direction of the point-supplementing guardrail to update the point-supplementing guardrail, and return to perform the step of determining the node to be added in the point-supplementing direction of the point-supplementing guardrail until the length of the point-supplementing guardrail reaches the preset length threshold, thereby obtaining the final detection guardrail.

Optionally, the node adding module 430 is further configured to determine, among all grids with real points in the detection area, a target grid closest to the node to be added, or use a grid with the largest number of real points in the detection area as the target grid; updating the abscissa of the node to be added based on the center of the target grid.

Optionally, the node adding module 430 is further configured to use, as the point to be complemented guard bars, the initial detection guard bars whose guard bar detection range does not meet the preset range in all the initial detection guard bars; judging whether the transverse coordinate difference between any two point guardrails to be complemented exceeds a preset transverse threshold value, if so, eliminating one point guardrail to be complemented; the preset transverse threshold value is half of the preset guardrail width.

Optionally, the point-complement guardrail determination module 410 includes a guardrail processing unit, configured to determine, for each initial detection guardrail, whether each guardrail segment in the initial detection guardrail is valid, and reject invalid guardrail segments from the initial detection guardrails; rejecting the initial detection guardrail under the condition that the number of guardrail sections in the initial detection guardrail is zero, and judging whether the initial detection guardrail meets a splitting condition or not based on the distance between adjacent guardrail sections under the condition that the number of guardrail sections in the initial detection guardrail is more than one; if yes, splitting the initial detection guardrails, and obtaining a plurality of new initial detection guardrails based on the splitting result.

Optionally, the guardrail processing unit is further configured to determine that the guardrail section is invalid when the number of nodes in the guardrail section is one; judging whether all nodes in the guardrail section are effective under the condition that the number of the nodes in the guardrail section is two, if so, determining that the guardrail section is ineffective under the condition that the distance between the nodes in the guardrail section is greater than a preset distance; judging whether the first node and the tail node in the guardrail section are valid or not under the condition that the number of nodes in the guardrail section is larger than two, if so, determining whether a new node between any two adjacent nodes in the guardrail section is valid or not to obtain the number of the effective nodes corresponding to the guardrail section, and if the number of the effective nodes does not meet the preset node proportion, determining that the guardrail section is invalid; wherein, the node effectively means that corresponding real points exist in the grid map.

Optionally, the guardrail processing unit is further configured to determine, if the number of nodes in the guardrail section is greater than two, that the first node and the last node in the guardrail section are valid, and if the first node and the last node are not valid, determine whether each node is valid from the first node to obtain the number of valid nodes, and determine that the guardrail section is not valid if the number of valid nodes does not meet a preset node ratio; if the tail nodes in the guardrail section are effective and the head nodes are ineffective, starting from the tail nodes, judging whether each node is effective or not, obtaining the number of effective nodes, and determining that the guardrail section is ineffective under the condition that the number of the effective nodes is smaller than the number of preset nodes.

Optionally, the point-compensating guardrail determining module 410 includes a clustering unit, where the clustering unit is configured to map point cloud data collected by a vehicle to each grid divided in advance to obtain a grid map corresponding to the point cloud data; determining whether each grid in the grid map is valid, and determining a connected domain based on each valid grid in the grid map; judging whether the number of real points in each effective grid forming the connected domain is larger than a preset number threshold value or not according to each connected domain, and judging whether the difference value between the highest heights of the effective grids is smaller than a preset height difference or not, if so, determining a clustering target based on the connected domain; and determining guardrail clustering targets in all the clustering targets, and determining an initial detection guardrail based on all the guardrail clustering targets, wherein the highest height is the height average value of all the real points in the grid, which are located in the first height range.

Optionally, the clustering unit is further configured to determine, for each grid in the grid map, that the grid is an effective grid if a height difference between a highest height and a relative height of the grid is greater than a height difference threshold; wherein the relative height is a height average of all the real points in the grid that lie within the second height range.

Optionally, the clustering unit is further configured to determine, for each effective grid, whether the number of real points in the effective grid is smaller than a set number, and if yes, update the effective grid to an ineffective grid if the number of first target points in the effective grid meets a first proportion; otherwise, updating the effective grid into an ineffective grid under the condition that the number of second target points in the effective grid meets a second proportion; the first target point is a real point with intensity larger than a preset intensity threshold value, and the second target point is a real point with height larger than the preset guardrail height.

Optionally, the device further includes a missing detection processing module, which is configured to determine a classification detection result and a classification confidence level of each real point in the point cloud data based on a pre-trained detection model, where the classification detection result is a pedestrian, a vehicle, and other objects; judging whether classification detection results of a head node and a tail node of each non-guardrail clustering target are other objects or not according to each non-guardrail clustering target; if yes, taking the first node of the non-guardrail clustering target as a current node, determining a new node in the neighborhood range of the current node according to the grid with the maximum number of real points and the highest classification confidence, taking the new node as the current node again, and returning to the step of determining the new node in the neighborhood range of the current node until the current node is the tail node of the non-guardrail clustering target, so as to obtain a guardrail missing detection target; and determining the final detection guardrails based on all the guardrails missing detection targets.

The guardrail detection device provided by the embodiment of the disclosure can execute steps in the guardrail detection method provided by the embodiment of the disclosure, and has the execution steps and beneficial effects, which are not repeated here.

Fig. 5 is a schematic structural diagram of an electronic device in an embodiment of the disclosure. Referring now in particular to fig. 5, a schematic diagram of an electronic device 500 suitable for use in implementing embodiments of the present disclosure is shown. The electronic device shown in fig. 5 is merely an example and should not be construed to limit the functionality and scope of use of the disclosed embodiments.

As shown in fig. 5, the electronic device 500 may include a processing means 501, a ROM502RAM503, a bus 504, an input/output (I/0) interface 505, an input means 506, an output means 507, a storage means 508, and a communication means 509. A processing device (e.g., central processing unit, graphics processor, etc.) 501, which may perform various suitable actions and processes to implement the methods of embodiments as described in the present disclosure, in accordance with programs stored in a Read Only Memory (ROM) 502 or loaded from a storage device 508 into a Random Access Memory (RAM) 503. In the RAM503, various programs and data required for the operation of the electronic apparatus 500 are also stored. The processing device 501, the ROM502, and the RAM503 are connected to each other via a bus 504. An input/output (I/O) interface 505 is also connected to bus 504.

In particular, according to embodiments of the present disclosure, the processes described above with reference to flowcharts may be implemented as computer software programs. For example, embodiments of the present disclosure include a computer program product comprising a computer program embodied on a non-transitory computer readable medium, the computer program comprising program code for performing the method shown in the flowchart, thereby implementing the guardrail detection method as described above. In such an embodiment, the computer program may be downloaded and installed from a network via the communication means 509, or from the storage means 508, or from the ROM 502. The above-described functions defined in the methods of the embodiments of the present disclosure are performed when the computer program is executed by the processing device 501.

It should be noted that the computer readable medium described in the present disclosure may be a computer readable signal medium or a computer readable storage medium, or any combination of the two. The computer readable storage medium can be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or a combination of any of the foregoing. More specific examples of the computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer diskette, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing. In the context of this disclosure, a computer-readable storage medium may be any tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device. In the present disclosure, however, the computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, with the computer-readable program code embodied therein. Such a propagated data signal may take any of a variety of forms, including, but not limited to, electro-magnetic, optical, or any suitable combination of the foregoing. A computer readable signal medium may also be any computer readable medium that is not a computer readable storage medium and that can communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device. Program code embodied on a computer readable medium may be transmitted using any appropriate medium, including but not limited to: electrical wires, fiber optic cables, RF (radio frequency), and the like, or any suitable combination of the foregoing.

The computer readable medium may be contained in the electronic device; or may exist alone without being incorporated into the electronic device. The computer readable medium carries one or more programs which, when executed by the electronic device, cause the electronic device to perform the guardrail detection method provided by any of the embodiments of the present disclosure.

Alternatively, the electronic device may perform other steps described in the above embodiments when the above one or more programs are executed by the electronic device.

In the context of this disclosure, a machine-readable medium may be a tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device. The machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. The machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of a machine-readable storage medium would include an electrical connection based on one or more wires, a portable computer diskette, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

Scheme 1, a guardrail detection method, the method includes:

Scheme 2, the method according to scheme 1,

the updating of the abscissa of the node to be added comprises the following steps:

Solution 3, the method according to solution 1, wherein determining the point to be complemented guard bar in all initial detection guard bars includes:

among all initial detection guardrails, the initial detection guardrails with the detection range of the guardrails not meeting the preset range are used as guardrails with points to be complemented:

judging whether the transverse coordinate difference between any two point guardrails to be complemented exceeds a preset transverse threshold value, if so, eliminating one point guardrail to be complemented;

the preset transverse threshold value is half of the preset guardrail width.

Solution 4, the method according to solution 1, after the obtaining an initial detection fence according to the clustering result, the method further includes:

judging whether each guardrail section in the initial detection guardrails is effective or not according to each initial detection guardrail, and removing ineffective guardrail sections from the initial detection guardrails;

rejecting the initial detection guardrail under the condition that the number of guardrail sections in the initial detection guardrail is zero, and judging whether the initial detection guardrail meets a splitting condition or not based on the distance between adjacent guardrail sections under the condition that the number of guardrail sections in the initial detection guardrail is more than one;

If yes, splitting the initial detection guardrails, and obtaining a plurality of new initial detection guardrails based on the splitting result.

The method of claim 5, according to claim 4, the determining whether each of the initial detection fence segments is valid, comprising:

determining that the guardrail section is invalid in the case that the number of nodes in the guardrail section is one;

judging whether all nodes in the guardrail section are effective under the condition that the number of the nodes in the guardrail section is two, if so, determining that the guardrail section is ineffective under the condition that the distance between the nodes in the guardrail section is greater than a preset distance;

judging whether the first node and the tail node in the guardrail section are valid or not under the condition that the number of nodes in the guardrail section is larger than two, if so, determining whether a new node between any two adjacent nodes in the guardrail section is valid or not to obtain the number of the effective nodes corresponding to the guardrail section, and if the number of the effective nodes does not meet the preset node proportion, determining that the guardrail section is invalid;

wherein, the node effectively means that corresponding real points exist in the grid map.

The method according to claim 6, wherein the determining whether each of the initial detection fence segments is valid further comprises:

If the number of nodes in the guardrail section is greater than two, if the head node and the tail node in the guardrail section are effective, judging whether each node is effective from the head node to the rear, and obtaining the number of effective nodes, and if the number of the effective nodes does not meet the preset node proportion, determining that the guardrail section is ineffective;

if the tail nodes in the guardrail section are effective and the head nodes are ineffective, starting from the tail nodes, judging whether each node is effective or not, obtaining the number of effective nodes, and determining that the guardrail section is ineffective under the condition that the number of the effective nodes is smaller than the number of preset nodes.

According to the scheme 7, according to the method of the scheme 1, the clustering is performed on the point cloud data collected by the vehicle, and the initial detection guardrail is obtained according to the clustering result, including:

mapping point cloud data acquired by a vehicle into grids which are divided in advance to obtain a grid map corresponding to the point cloud data;

determining whether each grid in the grid map is valid, and determining a connected domain based on each valid grid in the grid map;

judging whether the number of real points in each effective grid forming the connected domain is larger than a preset number threshold value or not according to each connected domain, and judging whether the difference value between the highest heights of the effective grids is smaller than a preset height difference or not, if so, determining a clustering target based on the connected domain;

Determining guardrail clustering targets in all the clustering targets, and determining an initial detection guardrail based on all the guardrail clustering targets;

wherein the highest height is the height average of all the real points in the grid within the first height range.

Solution 8, the method of solution 7, the determining whether each grid in the grid map is valid, comprising:

for each grid in the grid map, if the height difference between the highest height and the relative height of the grid is greater than a height difference threshold value, determining the grid as an effective grid;

wherein the relative height is a height average of all the real points in the grid that lie within the second height range.

Solution 9, the method according to solution 8, after said determining whether each grid in the grid map is valid, further comprising:

judging whether the number of the real points in each effective grid is smaller than the set number or not according to each effective grid, if yes, updating the effective grid into an ineffective grid under the condition that the number of first target points in the effective grid meets a first proportion;

otherwise, updating the effective grid into an ineffective grid under the condition that the number of second target points in the effective grid meets a second proportion;

The first target point is a real point with intensity larger than a preset intensity threshold value, and the second target point is a real point with height larger than the preset guardrail height.

Solution 10, the method of solution 7, further comprising:

determining classification detection results and classification confidence of all real points in the point cloud data based on a pre-trained detection model, wherein the classification detection results are pedestrians, vehicles and other objects;

after determining the guardrail cluster target from all the cluster targets, the method further comprises:

judging whether classification detection results of a head node and a tail node of each non-guardrail clustering target are other objects or not according to each non-guardrail clustering target;

if yes, taking the first node of the non-guardrail clustering target as a current node, determining a new node in the neighborhood range of the current node according to the grid with the maximum number of real points and the highest classification confidence, taking the new node as the current node again, and returning to the step of determining the new node in the neighborhood range of the current node until the current node is the tail node of the non-guardrail clustering target, so as to obtain a guardrail missing detection target;

And determining the final detection guardrails based on all the guardrails missing detection targets.

Scheme 11, a guardrail detection device includes:

Scheme 12, an electronic device, the electronic device comprising:

One or more processors;

a storage means for storing one or more programs;

the one or more programs, when executed by the one or more processors, cause the one or more processors to implement the methods of any of aspects 1-10.

Aspect 13, a computer readable storage medium, having stored thereon a computer program, which when executed by a processor, implements the method according to any of aspects 1-10.

The foregoing description is only of the preferred embodiments of the present disclosure and description of the principles of the technology being employed. It will be appreciated by persons skilled in the art that the scope of the disclosure referred to in this disclosure is not limited to the specific combinations of features described above, but also covers other embodiments which may be formed by any combination of features described above or equivalents thereof without departing from the spirit of the disclosure. Such as those described above, are mutually substituted with the technical features having similar functions disclosed in the present disclosure (but not limited thereto).

Claims

1. A method of guardrail detection, the method comprising:

2. The method of claim 1, wherein the updating the abscissa of the node to be added comprises:

3. The method of claim 1, wherein said determining the point to complement guard rail among all initial detection guard rails comprises:

among all initial detection guardrails, taking the initial detection guardrails of which the detection range of the guardrails does not meet the preset range as guardrails of points to be complemented;

the preset transverse threshold value is half of the preset guardrail width.

4. The method of claim 1, wherein after the initial detection fence is derived from the clustering result, the method further comprises:

5. The method of claim 4, wherein said determining whether each of said initial detection fence segments is valid comprises:

6. The method of claim 5, wherein said determining whether each of said initial detection fence segments is valid further comprises:

7. The method of claim 1, wherein clustering the point cloud data collected by the vehicle to obtain an initial detection fence according to the clustering result comprises:

8. A guardrail detection device, comprising:

9. An electronic device, the electronic device comprising:

one or more processors;

a storage means for storing one or more programs;

the one or more programs, when executed by the one or more processors, cause the one or more processors to implement the method of any of claims 1-7.

10. A computer readable storage medium, on which a computer program is stored, characterized in that the program, when being executed by a processor, implements the method according to any of claims 1-7.