Road surface small target detection method based on laser radar

Abstract

The invention discloses a road surface small target detection method based on a laser radar, and belongs to the technical field of radars, and the method comprises the steps: carrying out the statistical rate space-time downsampling of point cloud data, carrying out the regional road surface detection of a road surface detection range, carrying out the plane fitting of a multi-subregion normal vector median, and carrying out the detection of a small target. After the road surface of each area is found, non-road-surface points are segmented by using a confidence coefficient combined threshold, combined threshold segmentation is carried out by integrating the height difference between the three-dimensional points and the ground and the reflectivity, European clustering is carried out on the non-road-surface points obtained through segmentation, and possible small targets are obtained; the technical problems of removing data redundancy in pavement small target detection and improving the small target detection precision are solved, the calculation efficiency of small target detection is improved, the pavement unevenness and outlier influence are effectively overcome, the pavement estimation precision is improved, the small target segmentation precision can be effectively improved, false detection and missing detection are avoided, and the detection efficiency is improved. And the reliability of detection results is ensured.

Description

Technical Field

[0001] The present invention belongs to the field of radar technology, and in particular relates to a method for detecting small road targets based on laser radar. Background Art

[0002] Small foreign objects on the road surface (such as metal fragments, stones, parts, tools, etc.) pose a serious threat to the safety of vehicles, especially aircraft, especially on airport runways. The presence of these small targets may lead to flight accidents, so timely detection of these small targets is the key to ensuring the safe operation of vehicles. Laser radar (LiDAR) technology has shown great advantages in the detection of small targets on the road surface. Compared with traditional cameras, LiDAR is not affected by lighting conditions and can work stably in harsh environments such as night and fog, ensuring that targets can be effectively detected under various conditions. The point cloud data collected by LiDAR can not only provide three-dimensional spatial information of the target, but also accurately identify targets of different sizes and shapes through the height data of the point cloud.

[0003] However, the road surface small target detection technology based on LiDAR also faces multiple challenges. First, small targets are small in size and have unclear features, so they are easily affected by background noise, resulting in false detection or missed detection. Second, there is a certain redundancy in the point cloud data of LiDAR, especially near the road surface, where the data volume is large, which increases the computational burden. Third, the road surface itself may be uneven, and due to the amplification effect of the estimation error, the error of the road surface will further increase during long-distance detection, thereby affecting the accuracy of target detection. In addition, when collecting point clouds, in order to ensure that small targets are detected, a longer collection time may be required, which leads to an increase in the amount of data, which in turn places higher requirements on the real-time performance of the processing algorithm. Summary of the invention

[0004] The purpose of the present invention is to provide a method for detecting small road targets based on laser radar, which solves the technical problems of removing data redundancy in detecting small road targets and improving the accuracy of small target detection.

[0005] To achieve the above object, the present invention adopts the following technical solution:

[0006] A method for detecting small road targets based on laser radar comprises the following steps:

[0007] Step 1: The data acquisition module collects the original three-dimensional point cloud data on the road surface through the laser radar, calculates the point cloud density statistics based on the collected point cloud data, and performs spatiotemporal downsampling on the point cloud data based on the statistical information, pre-processes the sampling results, and obtains the down-sampled point cloud data;

[0008] Step 2: The road surface detection module retrieves point cloud data, performs road surface fitting, and processes the area to improve the accuracy of road surface detection. Specifically, it performs multi-level robust plane fitting, obtains the road surface normal vector of each area through area segmentation and sub-area normal vector calculation, and constructs a road surface model.

[0009] Step 3: The non-road surface point segmentation module retrieves the road surface model, performs joint threshold segmentation based on height difference and reflectivity, and separates non-road surface points from road surface points. Specifically, each point cloud is judged, the height difference between it and the road surface model is calculated, and segmentation is performed in combination with the reflectivity information to obtain a set of non-road surface points.

[0010] Step 4: The small target clustering module retrieves the non-road point set, clusters the non-road points in the non-road point set using the Euclidean clustering algorithm, and identifies the location of the small target. Specifically, the segmented non-road points are clustered and the centroid coordinates of each small target are output;

[0011] Step 5: The calculated center of mass coordinates of each small target will be output as the detection result, and the center of mass coordinates of each small target will be regarded as the position information of the small target and passed to the alarm or response module for subsequent alarm triggering or further processing.

[0012] Preferably, when executing step 1, the specific steps are as follows:

[0013] Step 1-1: The laser radar scans the road surface to obtain the original three-dimensional point cloud data, including the spatial coordinates and reflection intensity of each point;

[0014] Step 1-2: The data acquisition module acquires the 3D point cloud data collected by the laser radar in real time, and generates a point cloud density statistical rate by regularly counting the density changes of the point cloud. The point cloud density statistical rate is based on the distribution of the point cloud in different distance intervals during the point cloud acquisition process;

[0015] Step 1-3: Use the point cloud density statistics to slice the point cloud data collected in real time. Each slice corresponds to a different distance area. Further filter the valid data based on the timestamp and distance, discard the redundant points, and generate the downsampled point cloud data.

[0016] Preferably, when executing step 2, the following steps are specifically included:

[0017] Step 2-1: Divide the road surface into multiple areas according to the distance from near to far, and the size of each area is determined according to the point cloud density and sampling rules;

[0018] Step 2-2: Each region is further divided into multiple sub-regions, and each sub-region is independently fitted with a normal vector to obtain a local normal vector of the region;

[0019] Step 2-3: Calculate the median of the normal vectors of each sub-area to obtain the final normal vector of the area for subsequent pavement fitting.

[0020] Preferably, when executing step 3, the specific steps are as follows:

[0021] Step 3-1: For each point cloud data, calculate the height difference between it and the road surface model, that is, the vertical distance between the point and the road surface;

[0022] Step 3-2: Determine whether each point is a non-road surface point according to the set confidence threshold. If the height difference is large, it is considered to be a non-road surface point; if the height difference is small, it is judged in combination with the reflectivity information of the point;

[0023] Step 3-3: Based on the confidence level and threshold segmentation judgment, retain non-road surface points and discard road surface points.

[0024] Preferably, the Euclidean clustering algorithm specifically uses the Euclidean distance measurement method to calculate the distance between each non-road surface point and other points, and cluster similar points into one category; for each cluster, its centroid coordinates are calculated as the detection result of the small target.

[0025] The laser radar-based road surface small target detection method described in the present invention solves the technical problems of removing data redundancy in road surface small target detection and improving the accuracy of small target detection. The present invention eliminates redundant point clouds through statistical spatiotemporal downsampling to reduce the amount of data, thereby improving the computational efficiency of subsequent small target detection. A multi-level robust plane fitting method is adopted to effectively overcome the influence of road surface unevenness and outliers, and improve the accuracy of road surface estimation, especially the performance in long-distance detection. Through the confidence combined with threshold segmentation technology, the height difference and reflectivity information are integrated to effectively improve the accuracy of small target segmentation and avoid false detection and missed detection. The Euclidean clustering method is used to accurately locate small targets and ensure the reliability of detection results. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is the main flow chart of the present invention;

[0027] Figure 2 It is a flow chart of the confidence combined threshold segmentation of the present invention. DETAILED DESCRIPTION

[0028] Depend on Figure 1-Figure 2 A method for detecting small road targets based on laser radar is shown, comprising the following steps:

[0029] Step 1: The data acquisition module collects the original three-dimensional point cloud data on the road surface through the laser radar, calculates the point cloud density statistics based on the collected point cloud data, and performs spatiotemporal downsampling on the point cloud data based on the statistical information, pre-processes the sampling results, and obtains the down-sampled point cloud data; the specific steps are as follows:

[0030] Step 1-1: The laser radar scans the road surface to obtain the original three-dimensional point cloud data, including the spatial coordinates and reflection intensity of each point;

[0031] Step 1-2: The data acquisition module acquires the 3D point cloud data collected by the laser radar in real time, and generates a point cloud density statistical rate by regularly counting the density changes of the point cloud. The point cloud density statistical rate is based on the distribution of the point cloud in different distance intervals during the point cloud acquisition process;

[0032] Step 1-3: Use the point cloud density statistics to slice the point cloud data collected in real time. Each slice corresponds to a different distance area. Further filter the valid data based on the timestamp and distance, discard the redundant points, and generate the downsampled point cloud data.

[0033] In order to ensure that there are a certain number of points on small targets far away from the road surface, the 3D point cloud collection must be maintained for a sufficiently long time during a detection. This results in many redundant points near the road surface, increasing the amount of data and affecting the calculation efficiency of the algorithm.

[0034] It can be seen that the redundancy of the point cloud is related to the distance. This correlation is affected by the environment in addition to the difference in equipment. In this embodiment, the three-dimensional point cloud of the road surface is counted regularly (when the machine is turned on, every day, or every 12 hours) to form a point cloud density statistical rate, that is, the law of how the point cloud density changes with distance.

[0035] At each detection, the point cloud density statistics are used to perform spatiotemporal downsampling of the real-time point cloud, specifically:

[0036] Let the total acquisition time be T.

[0037] Slice the point cloud in terms of distance, from near to far, into S slices, and let the distances of each slice be x1, x2, ..., x s For the i-th slice, only the slices from the start of acquisition T i Data in time, Discard the point cloud data for the rest of the slice.

[0038] This embodiment uses statistical spatiotemporal downsampling to perform spatiotemporal downsampling of laser point cloud data according to the regularly counted point cloud density statistics, eliminating data redundancy at each distance, reducing the amount of data, and facilitating the efficiency of subsequent small target detection. The spatiotemporal downsampling operation does not require additional calculations, and only performs three-dimensional point filtering based on timestamps and distances, with high filtering efficiency. The regularly counted point cloud density statistics can dynamically adapt to various laser radar devices and various environments.

[0039] Step 2: The road surface detection module retrieves point cloud data, performs road surface fitting, and processes the area to improve the accuracy of road surface detection. Specifically, it performs multi-level robust plane fitting, obtains the road surface normal vector of each area through area segmentation and sub-area normal vector calculation, and constructs a road surface model. The specific steps include the following:

[0040] Step 2-1: Divide the road surface into multiple areas according to the distance from near to far, and the size of each area is determined according to the point cloud density and sampling rules;

[0041] Step 2-2: Each region is further divided into multiple sub-regions, and each sub-region is independently fitted with a normal vector to obtain a local normal vector of the region;

[0042] Step 2-3: Calculate the median of the normal vectors of each sub-area to obtain the final normal vector of the area for subsequent pavement fitting.

[0043] Road surface fitting will be affected by road surface unevenness, outliers, and estimation error amplification. In this embodiment, multi-level robust plane fitting is performed, with three different scale levels: road surface, region, and sub-region. First, the road surface is divided into N regions from near to far, and each region is individually fitted with a road surface plane. For each region, multiple sub-regions are selected to perform multi-sub-region normal vector median plane fitting. The specific steps are as follows:

[0044] For each region P i , divided horizontally, J sub-areas are obtained, and plane fitting is performed on each sub-area to obtain the normal vector n j =(a j , b j , c j ). For these J normal vectors n j , j = 1, ..., J performs median calculation to obtain the median normal vector n median , as the normal vector of the region.

[0045] J normal vectors n j =(a j , b j , c j ), j = 1, ..., J is calculated as follows: calculate a j , bj , cj The median value of each median , b median , c median Calculate n j and vector (a median , b median , c median ). Then the normal vector n with the smallest Euclidean distance is j That is the median normal vector n median .

[0046] This embodiment uses multi-level robust plane fitting to perform plane fitting at three different scale levels: road surface, region, and sub-region, overcoming problems such as uneven road surface, outliers, and amplified estimation errors. Since the road surface estimation error amplifies with distance in the direction of road surface extension, dividing the road surface into several regions can solve the problems of amplified estimation errors and uneven road surface. The road surface in an area may be locally uneven, and there may be outliers such as small targets and system noise, which may affect road surface estimation. To address these problems, this embodiment selects multiple sub-regions of the area, performs plane fitting on each sub-region, and then calculates the median of the normal vectors of the multiple sub-regions as the regional normal vector.

[0047] Step 3: The non-road surface point segmentation module retrieves the road surface model, performs joint threshold segmentation based on height difference and reflectivity, and separates non-road surface points from road surface points. Specifically, each point cloud is judged, the height difference between it and the road surface model is calculated, and segmentation is performed in combination with the reflectivity information to obtain a set of non-road surface points. The specific steps are as follows:

[0048] Step 3-1: For each point cloud data, calculate the height difference between it and the road surface model, that is, the vertical distance between the point and the road surface;

[0049] Step 3-2: Determine whether each point is a non-road surface point according to the set confidence threshold. If the height difference is large, it is considered to be a non-road surface point; if the height difference is small, it is judged in combination with the reflectivity information of the point;

[0050] Step 3-3: Based on the confidence level and threshold segmentation judgment, retain non-road surface points and discard road surface points.

[0051] After the previous step is completed, a road surface model composed of planes can be obtained. For each point of the real-time point cloud, the height difference from the road surface can be calculated according to the area where it is located. Since small targets are small in size and have unclear features, they account for a small proportion in the point cloud and are easily submerged by background noise. If they are used to segment non-road surface points, the confidence level is not high enough, which can easily lead to false detection and missed detection. Confidence combined with threshold segmentation, the height difference h of the comprehensive three-dimensional point i and reflectivity r i to split.

[0052] First determine whether the condition "high confidence in height difference" is met: h i >∈, ∈ is a set tolerance parameter. If it is satisfied, it is a non-surface point and is retained; if it is not satisfied, enter the "medium height difference confidence" judgment. Judge whether the condition "medium height difference confidence" is met: |h i |≤∈, if satisfied, further judgment is made by combining reflectivity confidence; if not satisfied, it is not a non-road surface point and is discarded. Determine whether the condition "reflectivity confidence meets the standard" is met: r i >T (T is the threshold value obtained by applying the OTSU algorithm to the reflectivity values ​​of all points in the area). If it is satisfied, it is a non-road surface point and is retained; if it is not satisfied, it is not a non-road surface point and is discarded.

[0053] Since the target to be inspected is a small target with a small size and a small proportion in the point cloud, it is easily submerged by background noise. The confidence level is not high enough when the height difference from the road surface is used to segment non-road surface points. To address this problem, this embodiment adopts confidence combined with threshold segmentation to integrate the height difference from the road surface and reflectivity. When the height difference from the ground is large, the confidence level of the small target point is high and reliable segmentation can be performed. When the height difference from the ground is small, the confidence level of the small target point is not high. At this time, the reflectivity confidence level is combined for joint segmentation.

[0054] Step 4: The small target clustering module retrieves the non-road point set, clusters the non-road points in the non-road point set using the Euclidean clustering algorithm, and identifies the location of the small target. Specifically, the segmented non-road points are clustered and the centroid coordinates of each small target are output;

[0055] The Euclidean clustering algorithm specifically uses the Euclidean distance measurement method to calculate the distance between each non-road surface point and other points, and clusters similar points into one category; for each cluster, its centroid coordinates are calculated as the detection result of the small target.

[0056] Euclidean clustering is a classic unsupervised learning algorithm that measures similarity by calculating the Euclidean distance between data points and clusters similar data points into one category. Each cluster output by Euclidean clustering corresponds to each small target. This method reports the centroid coordinates of each cluster as the location of the small target.

[0057] Step 5: The calculated center of mass coordinates of each small target will be output as the detection result, and the center of mass coordinates of each small target will be regarded as the position information of the small target and passed to the alarm or response module for subsequent alarm triggering or further processing.

[0058] The laser radar-based road surface small target detection method described in the present invention solves the technical problems of removing data redundancy in road surface small target detection and improving the accuracy of small target detection. The present invention eliminates redundant point clouds through statistical spatiotemporal downsampling to reduce the amount of data, thereby improving the computational efficiency of subsequent small target detection. A multi-level robust plane fitting method is adopted to effectively overcome the influence of road surface unevenness and outliers, and improve the accuracy of road surface estimation, especially the performance in long-distance detection. Through the confidence combined with threshold segmentation technology, the height difference and reflectivity information are integrated to effectively improve the accuracy of small target segmentation and avoid false detection and missed detection. The Euclidean clustering method is used to accurately locate small targets and ensure the reliability of detection results.

Claims

1. A method for detecting small road targets based on laser radar, characterized in that: The steps include: Step 1: The data acquisition module collects the original three-dimensional point cloud data on the road surface through the laser radar, calculates the point cloud density statistics based on the collected point cloud data, and performs spatiotemporal downsampling on the point cloud data based on the statistical information, pre-processes the sampling results, and obtains the down-sampled point cloud data; Step 2: The road surface detection module retrieves point cloud data, performs road surface fitting, and processes the area to improve the accuracy of road surface detection. Specifically, it performs multi-level robust plane fitting, obtains the road surface normal vector of each area through area segmentation and sub-area normal vector calculation, and constructs a road surface model. Step 3: The non-road surface point segmentation module retrieves the road surface model, performs joint threshold segmentation based on height difference and reflectivity, and separates non-road surface points from road surface points. Specifically, each point cloud is judged, the height difference between it and the road surface model is calculated, and segmentation is performed in combination with the reflectivity information to obtain a set of non-road surface points. Step 4: The small target clustering module retrieves the non-road point set, clusters the non-road points in the non-road point set using the Euclidean clustering algorithm, and identifies the location of the small target. Specifically, the segmented non-road points are clustered and the centroid coordinates of each small target are output; Step 5: The calculated center of mass coordinates of each small target will be output as the detection result, and the center of mass coordinates of each small target will be regarded as the position information of the small target and passed to the alarm or response module for subsequent alarm triggering or further processing.

2. A method for detecting small road targets based on laser radar as claimed in claim 1, characterized in that: When executing step 1, the specific steps are as follows: Step 1-1: The laser radar scans the road surface to obtain the original three-dimensional point cloud data, including the spatial coordinates and reflection intensity of each point; Step 1-2: The data acquisition module acquires the 3D point cloud data collected by the laser radar in real time, and generates a point cloud density statistical rate by regularly counting the density changes of the point cloud. The point cloud density statistical rate is based on the distribution of the point cloud in different distance intervals during the point cloud acquisition process; Step 1-3: Use the point cloud density statistics to slice the point cloud data collected in real time. Each slice corresponds to a different distance area. Further filter the valid data based on the timestamp and distance, discard the redundant points, and generate the downsampled point cloud data.

3. The method for detecting small road targets based on laser radar as claimed in claim 1, characterized in that: When executing step 2, the specific steps include: Step 2-1: Divide the road surface into multiple areas according to the distance from near to far, and the size of each area is determined according to the point cloud density and sampling rules; Step 2-2: Each region is further divided into multiple sub-regions, and each sub-region is independently fitted with a normal vector to obtain a local normal vector of the region; Step 2-3: Calculate the median of the normal vectors of each sub-area to obtain the final normal vector of the area for subsequent pavement fitting.

4. The method for detecting small road targets based on laser radar according to claim 1, characterized in that: When executing step 3, the specific steps are as follows: Step 3-1: For each point cloud data, calculate the height difference between it and the road surface model, that is, the vertical distance between the point and the road surface; Step 3-2: Determine whether each point is a non-road surface point according to the set confidence threshold. If the height difference is large, it is considered to be a non-road surface point; if the height difference is small, it is judged in combination with the reflectivity information of the point; Step 3-3: Based on the confidence level and threshold segmentation judgment, retain non-road surface points and discard road surface points.

5. The method for detecting small road targets based on laser radar as claimed in claim 1, characterized in that: The Euclidean clustering algorithm specifically uses the Euclidean distance measurement method to calculate the distance between each non-road surface point and other points, and clusters similar points into one category; for each cluster, its centroid coordinates are calculated as the detection result of the small target.

Images