Road obstacle perception system for unmanned vehicles based on lidar

Abstract

The present invention relates to the field of laser radar technology, and specifically to a road obstacle perception system for unmanned vehicles based on laser radar. The system includes: a laser radar data acquisition module, which adjusts the laser radar scanning frequency, collects point cloud data from static and dynamic obstacles, and obtains preliminary scanning data; and filters the preliminary scanning data to remove noise data. In the present invention, by calculating the dynamic characteristics of the point cloud data, the key points and movement trajectory of each obstacle are extracted, thereby optimizing the perception of obstacles in a dynamic environment. By adopting the method of time series analysis, the movement state and trajectory of the obstacle can be directly deduced according to the change of the point cloud, thereby improving the tracking ability of moving obstacles in complex dynamic scenes. In the process of spatial reconstruction, the point cloud data is reconstructed by geometric subdivision, and the dynamic information is combined with the static point cloud structure, thereby enhancing the parsing ability of spatial data.

Description

Technical Field

[0001] The present invention relates to the field of laser radar technology, and in particular to a road obstacle perception system for an unmanned vehicle based on laser radar. Background Art

[0002] Laser radar (LiDAR) is a sensor system based on light detection and ranging technology. It measures the distance, position and shape of target objects by emitting laser pulses and receiving reflected signals. LiDAR can generate high-precision three-dimensional point cloud data and is widely used in unmanned driving, terrain mapping, environmental monitoring, industrial automation and urban planning. However, the existing LiDAR technology has limitations in its ability to process dynamic information of obstacles. It is difficult to extract the trajectory and status of dynamic obstacles from point cloud data. It mainly relies on a fixed scanning mechanism to analyze static scenes and lacks in-depth mining of dynamic information, resulting in misjudgment or missed judgment when dealing with complex dynamic scenes. Therefore, improvements are needed. Summary of the invention

[0003] The purpose of the present invention is to solve the shortcomings of the prior art and propose a road obstacle perception system for unmanned vehicles based on laser radar.

[0004] In order to achieve the above-mentioned purpose, the present invention adopts the following technical solution: The road obstacle perception system of an unmanned vehicle based on laser radar includes:

[0005] The laser radar data acquisition module adjusts the laser radar scanning frequency, collects point cloud data from static and dynamic obstacles, and obtains preliminary scanning data; filters the preliminary scanning data, removes noise data, retains shape information, and generates optimized scanning data;

[0006] A dynamic obstacle analysis module calculates the gradient strength of each point in the point cloud based on the optimized scanning data, identifies key obstacle points, performs time series analysis on the key obstacle points, determines the movement state and trajectory of the obstacle, and generates moving obstacle trajectory information;

[0007] The laser radar data optimization module receives the moving obstacle trajectory information, uses Delaunay triangulation to spatially reconstruct key obstacle points to obtain spatially reconstructed data; compares the spatially reconstructed data with existing map data, optimizes the point cloud data structure, and generates optimized point cloud data;

[0008] The real-time perception data output module generates the position and size information of the obstacle in real time based on the optimized point cloud data to obtain the obstacle positioning data.

[0009] Preferably, the steps of acquiring the preliminary scanning data are:

[0010] Set the scanning frequency of the laser radar to collect raw point cloud data from static and dynamic obstacles to obtain raw point cloud data;

[0011] Based on the original point cloud data, signal attenuation compensation is performed to remove irrelevant information to obtain cleaned point cloud data;

[0012] Based on the cleaned point cloud data, the spatial association and structural link between the data points are calculated, the spatial position and structure of the object are mapped, the three-dimensional shape of the object is reconstructed, and preliminary scanning data is obtained.

[0013] Preferably, the step of acquiring the optimized scan data is:

[0014] Based on the preliminary scan data, low-frequency noise is removed by a high-pass filter, and high-frequency noise is removed by a low-pass filter to obtain scan data after preliminary noise reduction;

[0015] Based on the scan data that has undergone preliminary noise reduction, edge detection is applied to retain the edges and corner points of the object to obtain optimized scan data.

[0016] Preferably, the steps of acquiring the key obstacle points are:

[0017] Based on the optimized scan data, a deep learning model is trained to identify and calculate local geometric features in the point cloud data to obtain a spatial variation index for each point;

[0018] Based on the spatial variation index of each point, the gradient strength of each point is calculated, and the calculation formula is:

[0019]

[0020] in, Represents the reflection intensity function of the point cloud, It is the first The three-dimensional coordinates of the point, represents the partial derivative in the corresponding coordinate direction, It is The gradient strength of a point, is the adjustment coefficient, which indicates the intensity of neighborhood influence, Yes and The spatial weight between Yes The neighborhood point set of and They are points and The reflection intensity;

[0021] According to the gradient strength of each point, a threshold is set, and the key obstacle points are identified through threshold comparison and screening.

[0022] Preferably, the steps of acquiring the moving obstacle trajectory information are:

[0023] Based on the key obstacle points, time series analysis is performed, and the data points of the continuous time frame are processed using a Kalman filter to estimate the state and speed of each point to obtain time-dependent speed and position data;

[0024] Based on the time-dependent speed and position data, the speed of each point is calculated using the following formula:

[0025]

[0026] in, Indicates time At point speed, and The time and Interval exist and The displacement difference in direction, is the time interval between two frames, is a decay factor;

[0027] According to the speed of each point, a threshold is set to determine whether it is a moving obstacle, and the moving obstacle trajectory information is obtained.

[0028] Preferably, the steps of acquiring the spatial reconstruction data are:

[0029] Based on the moving obstacle trajectory information, Delaunay triangulation is performed, and the formula is:

[0030]

[0031] in, Indicated by point is the sum of the areas of the triangles at the vertices, and They are points and neighboring points The two-dimensional coordinates of Yes The set of adjacent points of ;

[0032] The sum of the triangle areas is used to perform spatial reconstruction, and the sum of the triangle areas is mapped into a three-dimensional spatial structure to obtain spatial reconstruction data.

[0033] Preferably, the step of obtaining the optimized point cloud data is:

[0034] Based on the spatial reconstruction data, the spatial reconstruction data is compared with the existing map data, including checking the geographical location and height information of each data point, to obtain a preliminary comparison result;

[0035] Based on the preliminary comparison results, the distortion or deviation caused by the limitation of the scanner device is corrected to obtain optimized point cloud data.

[0036] Preferably, the step of acquiring the obstacle positioning data is:

[0037] Based on the optimized point cloud data, the boundary points of all obstacles are extracted, the geometric characteristics of spatial distribution are calculated point by point, the point sets constituting the obstacles are identified, and the point sets are classified and aggregated to obtain preliminary boundary information of the obstacles;

[0038] Based on the preliminary boundary information of the obstacle, a curve is reconstructed for the boundary points of each obstacle through contour fitting, and the length, width and height of the reconstructed curve are extracted to obtain the size characteristics of the obstacle;

[0039] Obstacle location data is generated based on the size characteristics of the obstacle and combined with the spatial distribution information of the boundary points.

[0040] Compared with the prior art, the advantages and positive effects of the present invention are:

[0041] In the present invention, by calculating the dynamic characteristics of the point cloud data, the key points and movement trajectory of each obstacle are extracted, thereby optimizing the perception of obstacles in a dynamic environment. By using the time series analysis method, the movement state and trajectory of the obstacle can be directly derived according to the point cloud changes, thereby improving the tracking ability of moving obstacles in complex dynamic scenes. In the process of spatial reconstruction, the point cloud data is reconstructed using geometric subdivision, and the dynamic information is combined with the static point cloud structure, thereby enhancing the parsing ability of spatial data. By comparing with existing map data, the point cloud data structure is further optimized, solving the problem of error accumulation in point cloud data or static map errors, making the generated data more suitable for subsequent perception and decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a system flow chart of the present invention. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0044] See also Figure 1The present invention provides a technical solution: a laser radar-based unmanned vehicle road obstacle perception system comprising:

[0045] The laser radar data acquisition module adjusts the laser radar scanning frequency, collects point cloud data from static and dynamic obstacles, and obtains preliminary scanning data; filters the preliminary scanning data, removes noise data, retains shape information, and generates optimized scanning data;

[0046] The dynamic obstacle analysis module calculates the gradient strength of each point in the point cloud based on the optimized scanning data and identifies key obstacle points. It also performs time series analysis on key obstacle points to determine the movement state and trajectory of the obstacles and generate moving obstacle trajectory information.

[0047] The LiDAR data optimization module receives the moving obstacle trajectory information, uses Delaunay triangulation to spatially reconstruct key obstacle points, and obtains spatially reconstructed data; compares the spatially reconstructed data with the existing map data, optimizes the point cloud data structure, and generates optimized point cloud data;

[0048] The real-time perception data output module generates the position and size information of obstacles in real time based on the optimized point cloud data to obtain obstacle positioning data.

[0049] The steps to obtain the preliminary scan data are:

[0050] Set the scanning frequency of the laser radar to collect raw point cloud data from static and dynamic obstacles to obtain raw point cloud data;

[0051] Based on the original point cloud data, signal attenuation compensation is performed to remove irrelevant information and obtain cleaned point cloud data;

[0052] Based on the cleaned point cloud data, the spatial associations and structural links between data points are calculated, the spatial position and structure of the object are mapped, the three-dimensional shape of the object is reconstructed, and preliminary scanning data is obtained.

[0053] Specifically, after multiple tests on vehicle speed, it was determined that the speed range is usually between 0km / h and 120km / h. Through statistics on existing test data, the scanning frequency range of the lidar is set to 10 to 20 times per second, and this range corresponds to conventional urban and suburban driving scenarios. If the vehicle speed exceeds 100km / h, the scanning frequency can be adjusted to 25 times per second when recording historical comparisons, and faster-moving obstacle reflection points can be included in the observation. The frequency adjustment value involved in this process is derived from the layered speed file data collected in the past, and is adjusted based on the vehicle's average speed. The average acceleration and road condition difference distribution are corrected, and then the lidar sensor is placed on the movable platform and the continuous scanning function is activated. The original information such as azimuth, elevation and echo intensity at the scanning time is recorded. At each time step after the scanning is completed, the collected reflected light beam information is compared with the blank area of ​​the surrounding scene, and the angular range and detection distance that may contain obstacle data are recorded. In this way, static obstacles and dynamic obstacles are preliminarily identified, and then the beam data at all times are summarized and the relative positions of obstacles in different time periods are marked, and these records are integrated into original point cloud data.

[0054] When using the original point cloud data, first set a reference range for judging the degree of signal attenuation based on the laser radar's transmission power and echo energy distribution records. The range is calculated by accumulating the peak and valley values ​​under multiple scans of the same environment, taking the average value and then adding or subtracting an empirical deviation. If the actual detected echo signal is lower than the lower limit of the reference range, it is determined to be a significant attenuation value, and compared with the distance measurement record of the external reflective surface during the vehicle's driving process to identify signal distortion caused by too far measurement or low reflectivity materials, and then perform amplitude correction on all data information in the attenuation range, and compare the corrected echo energy with the pre-defined noise threshold. This noise threshold is based on the same model. After multiple sampling of the detection results of the laser radar under different temperature and humidity conditions, the average value is selected according to the signal stability and a standard deviation is added to obtain it. If the echo energy exceeds the upper limit of the noise threshold, it is regarded as abnormal data and eliminated. In addition, when judging irrelevant information, it will detect whether the point cloud coordinates exceed the recorded driving area around the vehicle. For example, the horizontal coordinates and the vertical coordinates are limited to between -50 meters to 50 meters and -100 meters to 100 meters. If some point clouds exceed this range and do not have observable dynamic properties during subsequent vehicle driving, they will be classified as irrelevant information and eliminated. After the above processing, a relatively coherent point cloud mapping is formed, and finally the cleaned point cloud data is obtained after aggregation.

[0055] By performing neighborhood search on the cleaned point cloud data point by point, the Euclidean distance between each point and several neighboring points around it is recorded. When determining whether they belong to the same local structure, the distance threshold is set to 0.2 meters. This threshold is derived from a rough measurement of the shape of common road obstacles and can be adjusted appropriately according to the difference in vehicle type and road width. If the Euclidean distance between a point and its neighbors is less than the threshold, these points are grouped in the same spatial connection group. When performing topological analysis on each connection group, the coordinate distribution and height difference of the points in the group are calculated one by one. In this way, the contour features of the object boundary are obtained, and the coordinate difference between the points is used to calculate the height difference. The preliminary morphological data is recorded, and then when mapping the spatial position and structure of the object, a three-dimensional coordinate system reference is assigned to each connection group, the coordinate range of the Z axis is limited to between -3 meters and 3 meters, and the range of the X axis and Y axis is limited to between -100 meters and 100 meters. If a continuous point cloud distribution is detected within these ranges, it is regarded as a complete object and surface interpolation calculation is performed. During the surface interpolation process, the boundary direction between points is summarized by using the triangular mesh superposition method, and the summary results are recorded in the three-dimensional point coordinate table. Then, the interpolation calculation results of all local connection groups are integrated one by one to generate the corresponding three-dimensional shape data, and finally the preliminary scanning data is obtained.

[0056] The steps to obtain the optimized scan data are:

[0057] Based on the preliminary scan data, low-frequency noise is removed by a high-pass filter, and high-frequency noise is removed by a low-pass filter to obtain scan data with preliminary noise reduction;

[0058] Based on the scan data that has undergone preliminary denoising, edge detection is applied to retain the edges and corners of the object to obtain optimized scan data.

[0059] Specifically, based on the preliminary scanning data obtained earlier, the signal distribution is first calculated point by point for the data information therein and the signal frequency range is recorded. The cutoff frequency in the high-pass filter is set at about 10 Hz and compared with the environmental noise curve collected in the previous experiment. If a low-frequency component is found to exceed this range, it will be removed. When determining the value of 10 Hz, the noise test records of different indoor and outdoor areas are referred to, and the error of 1 Hz is added or subtracted according to the average level of multiple groups of data to obtain the most realistic value. At the same time, when implementing the low-pass filter, the cutoff frequency is set to 500 Hz and the energy level of each frequency band in the screening process is recorded to determine whether there is a high-frequency component that exceeds 500 Hz but still affects the subsequent calculations. When it is confirmed that the signal distribution amount above 500 Hz is less than 5%, it is considered to be negligible. Otherwise, it is necessary to readjust the cutoff frequency to 600Hz to continue comparing the above energy distribution results. When setting this part of the cutoff value, the speed and acceleration information collected under different vehicle driving scenarios are mainly referred to. For example, the vehicle speed is compared between 0km / h and 120km / h and the energy peak of the corresponding frequency component is observed before the specific value is determined comprehensively. After all frequency bands are processed, the amplitude and phase distribution of the residual signal after each filtering are recorded in frequency order, and then a new time domain signal sequence is constructed and compared with the original data for differential comparison. When the differential signal amplitude fluctuates within the range of -0.1 to 0.1, it is determined that the filtering effect of this section is stable. If the differential amplitude is found to be greater than 0.1, it means that the filtering parameters need to be rechecked and multiple corrections need to be made in combination with the acquisition noise range to finally obtain the scanning data after preliminary noise reduction.

[0060] After obtaining the scan data after preliminary noise reduction, the spatial coordinates and corresponding intensity information of each point are first recorded, and the gradient change of adjacent points in the three-dimensional coordinates is calculated to identify the potential edge position. When determining whether it belongs to the edge, the gradient threshold is set to 0.5 and compared with the gradient value of each point. This 0.5 is obtained by looking at the gradient feature statistics of the transition area of ​​the surface of common objects in the previous measurement, and can float up or down by 0.1 to 0.3 when the on-site environment changes greatly. If the gradient value of a point exceeds the threshold, it will be marked as a high gradient point, and such high gradient points must be continuously detected. The horizontal and vertical differences between the point and the surrounding low-gradient points are used to locate the precise boundary direction through the differential value. On this basis, the connection mode of several adjacent high-gradient points is analyzed and the corner parts are marked as corner points. In addition to the reference gradient threshold, the judgment of corner points also needs to detect the curvature mutation in more than two consecutive directions in the neighborhood. At this time, the curvature threshold can be set to 10 and observe whether the curvature increment exceeds 10 and then judged in combination with the coordinate difference. After all edge points and corner points are marked, these marking information are unified into the spatial index system of the point cloud object for comprehensive summary, and finally the optimized scan data is obtained.

[0061] The steps to obtain key obstacle points are:

[0062] Based on the optimized scan data, the deep learning model is trained to identify and calculate the local geometric features in the point cloud data and obtain the spatial variation index of each point;

[0063] Based on the spatial variation index of each point, the gradient strength of each point is calculated. The calculation formula is:

[0064]

[0065] in, Represents the reflection intensity function of the point cloud, It is the first The three-dimensional coordinates of the point, represents the partial derivative in the corresponding coordinate direction, It is The gradient strength of a point, is the adjustment coefficient, which indicates the intensity of neighborhood influence, Yes and The spatial weight between Yes The neighborhood point set of and They are points and The reflection intensity;

[0066] According to the gradient strength of each point, a threshold is set, and the key obstacle points are identified through threshold comparison and screening.

[0067] Specifically, based on the optimized scan data obtained above, tens of thousands of points with complete three-dimensional coordinates and reflection intensity information are selected as input samples of the deep learning model in the training stage. The local neighborhood distribution of each point is recorded and these neighborhood features are paired with the corresponding manually labeled categories. The network weights are adjusted item by item through multiple rounds of iterations to reduce the loss value. The loss value can be calculated by combining the difference between the actual labeled category in the sample and the network output. If it is detected that the loss value has not dropped by more than 0.01 after dozens of consecutive iterations, the learning rate is updated and training continues. The specific setting of the learning rate is usually between 0.001 and 0.0001. The loss trend is observed in a fine-tuning manner. If it is found during the training process that the reflection intensity distribution of some points deviates from the average value by more than 20%, the corresponding neighborhood will be classified into a separate high-difference interval and analyzed in detail. In the later stage, a verification stage will be carried out and the comparison between the output results in the verification set and the known labels will be recorded. When the verification accuracy reaches more than 98%, the model is considered to have converged and stabilized. The trained network is then used to process the entire point cloud data and output geometric eigenvalues ​​in the local space divided by each grid. These eigenvalues ​​will be used as important parameters to measure local differences and will be saved in a list for subsequent analysis. Finally, the spatial variation index of each point is obtained after comprehensive evaluation.

[0068] The benefit of the formula is that it can incorporate the intensity difference between the point and its neighbors into the gradient calculation process, using the neighborhood influence coefficient With spatial weight Comprehensively measure changes in multiple dimensions;

[0069] The acquisition steps are as follows: first, the reflection intensity values ​​of the point cloud data obtained above are recorded between 0 and 255, and the abnormal values ​​beyond this range are compared and corrected according to the historical distribution records of the same type of laser radar, so as to determine the true reflection intensity of each point;

[0070] The acquisition step is to record the specific position coordinates point by point based on the three-dimensional coordinate system collected previously, and the value range can be within -100 meters to 100 meters. If some points exceed this range, the trajectory data of the vehicle navigation system is compared to find out whether it is indeed abnormal, and finally all confirmed valid coordinates are retained;

[0071] The steps of obtaining are: select the f value under the known coordinates and calculate the partial derivative by differentiating the f values ​​at adjacent coordinates. If the interval between the adjacent points is large, interpolation can be performed to obtain a continuous f distribution, and then the specific partial derivative result is obtained by calculating the ratio of the coordinate increment to the intensity increment;

[0072] The acquisition steps are as follows: combine the long-term observations of the changes in illumination and reflectivity on different test roads, record the intensity fluctuation curve at each observation, superimpose the curves of multiple observations to determine the reference value, and then correct it according to the additional dispersion caused by the driving speed from 0km / h to 120km / h to form a variable curve, and select fixed points from the curve as If the road environment is relatively simple, a smaller value of 0.3 to 0.5 can be selected. If the terrain is undulating, it can be increased to 0.7 to 0.9.

[0073] The acquisition steps are as follows: record the positional relationship between the i-th point and the j-th point in the actual road environment. If the distance between the two points is within 0.2 meters, a higher weight such as 0.8 to 1.0 is given; if the distance is above 0.5 meters, a lower weight such as 0.2 to 0.4 is given. The intermediate range is linearly interpolated according to the distance. Then, the weight of each adjacent point pair is stored and called in the calculation according to the complete point cloud distribution;

[0074] The steps of obtaining are as follows: search for other points within 1 meter for each point according to the neighborhood index established previously. If the height difference of obstacles exceeds 1 meter, reduce the neighborhood radius to exclude points that are too far away, thereby obtaining N(i);

[0075] The acquisition steps of are the same as f, except that different point coordinates are indexed and their reflection intensities at the same scanning time are recorded respectively;

[0076] Calculation process:

[0077] Select a specific point i and determine its coordinates , obtained by extracting partial derivatives of the adjacent coordinates before and after ,at this time The reflection intensity is 180;

[0078] Retrieve the neighborhood set N(i) of point i. If it contains point j1 and point j2 and their actual distances to point i are 0.3 meters and 0.15 meters respectively, then Set to 0.6, Set to 0.85, the reflection intensities corresponding to points j1 and j2 are 160 and 200 respectively. Recorded as 0.5;

[0079] Substitute the parameters into In, get ;

[0080] Substitute all partial derivatives and neighborhood terms into the main formula:

[0081]

[0082]

[0083]

[0084] The result shows that point i has a gradient strength value of 17.21 in the corresponding scene. If a gradient strength greater than 12 is found in subsequent comparisons, it will be classified as a significant area, indicating that the reflection intensity at the location of point i varies greatly. Combined with this result, it can be further evaluated whether point i belongs to a complex obstacle structure or an edge area.

[0085] According to the gradient intensity of each point obtained previously, a threshold range needs to be determined in the subsequent processing and the screening is completed. First, when recording the distribution of all gradient intensities, the average value and standard deviation are observed. Those gradient intensities exceeding the average value plus two standard deviations are determined as strong change areas, and those gradient intensities below the average value minus two standard deviations are classified as stable areas. The middle part is further subdivided and graded according to the vehicle driving environment, such as 0 to 10, 10 to 15, and 15 to 30. The numerical ranges of these gradient thresholds are collected from a large number of urban roads and various obstacle data. If the number of points corresponding to the strong change area is detected to account for more than 30%, it can be checked whether there is a large-scale structural mutation. Otherwise, the established screening strategy is maintained unchanged. At this time, in the point-by-point comparison process, if the gradient intensity of a point is in the strong change interval, it is identified as a key obstacle point, and if it is only in the stable area, it is classified as an ordinary background point. Finally, the marked results are summarized to identify the key obstacle points.

[0086] The steps to obtain the trajectory information of the moving obstacle are:

[0087] Based on the key obstacle points, time series analysis is performed, and the data points of continuous time frames are processed using the Kalman filter to estimate the state and speed of each point and obtain time-dependent speed and position data;

[0088] Based on the time-dependent velocity and position data, the velocity of each point is calculated using the following formula:

[0089]

[0090] in, Indicates time At point speed, and The time and Interval exist and The displacement difference in direction, is the time interval between two frames, is a decay factor;

[0091] According to the speed of each point, a threshold is set to determine whether it is a moving obstacle, and the moving obstacle trajectory information is obtained.

[0092] Specifically, based on the key obstacle points obtained above, the position information of these points in different time frames is first arranged in chronological order and the time interval between frames is recorded. Time series analysis is performed through the continuously acquired point position sequence. When reading the sampling time interval between adjacent frames, the sensor refresh frequency and the obstacle movement speed during the vehicle driving process need to be synchronously recorded. If the sensor refresh frequency is found to be greater than 20Hz, the time interval is set to about 0.05 seconds. If the refresh frequency is between 10Hz and 15Hz, the time interval is stabilized in the range of 0.07 seconds to 0.1 seconds, and the coordinate positions of the key obstacle points in two adjacent time frames are obtained one by one. In this way, the coordinate changes are obtained. After that, the predicted position and the measured position of the obstacle point are compared in each iteration according to the Kalman filter principle. The error in the comparison result is divided into two parts: process noise and measurement noise, and the corresponding noise covariance matrix is ​​set. The value of the process noise is taken with reference to the obstacle motion trajectory records collected on urban roads and relatively open roads. Statistics are performed, and the speed jitter range is analyzed in the real motion data of obstacles recorded multiple times. If the average speed around the vehicle is between 20km / h and 40km / h, the process noise can be set to a lower value to represent a higher degree of stability in the obstacle movement. If the average speed around the vehicle is between 60km / h and 80km / h, the process noise value is appropriately increased. The measurement noise is related to the laser radar echo accuracy and the obstacle reflection characteristics. It is necessary to distinguish whether there is reflection intensity fluctuation or occlusion in the same time period. If occlusion occurs, the relative increase in the measurement noise of the frame is recorded and corresponding trade-offs are made in subsequent calculations. As the number of iterations increases, the difference between the prior state and the posterior state of the filter can be used to evaluate the accuracy of the new noise matrix, and the overall Kalman gain change is recorded. When the Kalman gain no longer fluctuates significantly, it means that the current obstacle point movement trend is relatively consistent. Finally, the position information and corresponding velocity component records of each key obstacle point in multiple time frames are integrated to output time-dependent velocity and position data.

[0093] The formula is useful in that it can be combined with an attenuation factor for the displacement distance of a point in a plane Perform weighted processing to measure the velocity change within the observation interval of different time frames and take into account the short-term motion characteristics;

[0094] The steps to obtain are as follows: At the moment With the moment The horizontal coordinates record their differences. For example, in the coordinate distribution obtained above, if the point At the moment The x-coordinate is 16.3 meters at time The x coordinate is 15.8 meters, then rice;

[0095] The steps to obtain are as follows: The difference of the longitudinal coordinates is recorded in the same way. For example, if the point is in the y direction at the moment and The coordinates of are 30.6 meters and 29.6 meters respectively, then rice;

[0096] The acquisition step is to read the timestamp difference of adjacent frames. If the operating frequency of the laser radar is 10Hz, seconds, if at 20Hz then seconds. If the actual working frequency fluctuates slightly, it will be set to 0.07 or 0.08 seconds according to the actual record;

[0097] The acquisition steps are as follows: combine the obstacle motion smoothness records obtained previously and analyze the motion jitter amplitude under different vehicle speeds or road environments. If the obstacle moves relatively smoothly in the current environment, the value can be selected in the range of 0.2 to 0.3. If the surrounding vehicles are moving fast or obstacles are frequent, select the value between 0.4 and 0.6. ;

[0098] Calculation process:

[0099] Select current time point And record rice, rice, Seconds and ;

[0100] calculate ,get rice 2 ;

[0101] For the attenuation factor Perform calculations, ,thereby ;

[0102] Substitute the above results into the main formula:

[0103]

[0104]

[0105] The results show that at this moment The speed is 11.0139 meters per second. If subsequent statistics show that the value exceeds 10 meters per second, it can be regarded as a medium or higher moving speed and can be classified into the high dynamic obstacle range during classification. If the speed value is below 3 meters per second, it is regarded as a low-speed obstacle movement mode.

[0106] According to the speed of each point and combined with the time-dependent speed and position data, it is necessary to statistically summarize the speed distribution. When more than 25% of the points in a certain area have a speed greater than 5 meters per second, they will be given special attention. The specific method is to sort the speed values ​​of each point and find the quantile position. The list of points in the higher speed ranking interval is stored and checked whether their spatial positions are concentrated in the same area. If most points in the same area have similar speed values ​​in several consecutive frames and are all higher than 3 meters per second, the area is identified as the range of mobile obstacles. Then, the horizontal coordinate distribution of each point in the range is compared and whether there is a significant change, such as the displacement increment of the horizontal coordinate exceeding 2 meters or the displacement increment of the vertical coordinate exceeding 5 meters in a short period of time. If these situations occur and the increment level remains similar in the next two or three time frames, these points are classified into the mobile obstacle set. Finally, after completing the speed and spatial aggregation analysis, the trajectory information of the mobile obstacle is obtained.

[0107] The steps to obtain spatial reconstruction data are:

[0108] Based on the trajectory information of the moving obstacle, Delaunay triangulation is performed, and the formula is:

[0109]

[0110] in, Indicated by point is the sum of the areas of the triangles at the vertices, and They are points and neighboring points The two-dimensional coordinates of Yes The set of adjacent points of ;

[0111] The sum of the triangle areas is used to perform spatial reconstruction, and the sum of the triangle areas is mapped into a three-dimensional spatial structure to obtain spatial reconstruction data.

[0112] Specifically, the formula is beneficial in that each point Its adjacent point set The two-dimensional coordinate relationship between them is quantified in the form of triangle area, so that the neighborhood plane layout can be fully utilized to reflect the real spatial distribution when constructing the three-dimensional reconstruction model in the subsequent construction.

[0113] The steps to obtain are: first, The horizontal and vertical measurement coordinates in the plane are recorded. The horizontal value is between 0m and 100m, and the vertical value is between 0m and 50m. If the coordinates of some points fall below 0m or exceed the corresponding range, it is necessary to check with the actual measurement results of road distribution or obstacles to determine the final valid coordinates;

[0114] The steps to obtain The same way to get, just index to the neighboring point , by comparing the distances between adjacent points in the same area, if the distance is between 1 meter and 10 meters, it is included in the adjacent point set, and if it exceeds 10 meters, it is considered not in the neighborhood where a triangle can be directly constructed;

[0115] The acquisition steps are as follows: Combined with the previously obtained key obstacle point projection distribution records, search for the same point in the same area Other points within 10 meters, if a large number of points are detected within this range, then the points The nearest points are , and in the subsequent steps for each pair Coordinates for area calculation;

[0116] Calculation process:

[0117] Pick a point And record its two-dimensional coordinates , retrieve the adjacent points and The corresponding coordinates are and , including these two points ;

[0118] Calculate the points first and The area of ​​the triangle formed is:

[0119]

[0120] Recalculate the point and The area of ​​the triangle formed is:

[0121]

[0122] Add the areas of these two parts:

[0123]

[0124] The results show that The sum of the corresponding triangle areas in the current neighborhood is 30.555 square units. If the result is greater than 30, it means that the neighboring points are The plane distribution of is relatively divergent. If it is less than 5, it means that the neighboring points around the point are relatively concentrated. These area values ​​can be combined in the subsequent three-dimensional mapping to further evaluate the morphological elements of the overall structure.

[0125] Using the sum of the triangle areas obtained earlier, first calculate all the points The corresponding area cumulative values ​​are recorded one by one and sorted in sequence. Points with cumulative values ​​between 5 and 15 are classified as small area distribution, points between 15 and 30 are classified as medium area distribution, and points with cumulative values ​​greater than 30 are classified as large area distribution. If there are many points with large area distribution in the same area, the coordinates of these points are summarized and compared with the subsequent position change records of the key obstacle points. Then, the lateral and vertical expansion degrees of the triangle area are further calibrated. The upper and lower connection of adjacent triangles is determined by comparing the continuous changes of each point in the X-axis and Y-axis directions. For example, the interval from 0 to 100 meters in the X-axis direction can be compared and the boundary junctions of the adjacent triangle projections can be observed. If the cumulative area of ​​the boundary junction is greater than 25 and is maintained in the same area, then the boundary junction is considered to be a small area. If similar values ​​are maintained, it means that the local plane structure is relatively open, and in the Y-axis direction, the area continuity can be compared in the range of 0 meters to 50 meters. If the difference in the cumulative area values ​​of five consecutive points is within 5, it is considered that the adjacent triangles are relatively smooth in the vertical direction. Then, these distribution features are supplemented according to the originally stored three-dimensional height data, and the corresponding vertical coordinate records are added to the plane coordinates and combined to form a three-dimensional structure point set. Then these point sets are sequentially constructed into a spatial surface grid. When confirming the required interpolation accuracy, a resolution of 0.1 meters to 0.5 meters can be selected according to the vehicle driving environment. If the resolution exceeds 0.5 meters, it is easy to cause local corners to be too rough. The generation of spatial reconstruction data can be completed by comparing the above range and the comprehensive distribution results of adjacent points.

[0126] The steps to obtain optimized point cloud data are:

[0127] Based on the spatial reconstruction data, the spatial reconstruction data is compared with the existing map data, including the verification of the geographical location and height information of each data point, to obtain preliminary comparison results;

[0128] Based on the preliminary comparison results, the distortion or deviation caused by the limitations of the scanner equipment is corrected to obtain optimized point cloud data.

[0129] Specifically, based on the spatial reconstruction data obtained previously, the longitude and latitude coordinates of each data point in the map coordinate system are recorded and archived one by one. Then, when comparing these records with the corresponding coordinates in the existing map data, it is necessary to first match the same or similar geographical locations and check whether their height information is within the pre-established allowable deviation range. For example, the height deviation range can be set to -2 meters to 2 meters. This value is obtained based on the average error statistically calculated after multiple terrain measurements plus a standard deviation. If it is found that the height value of a point is greater than 2 meters different from the height at the same position in the existing map, it is classified as a significant deviation point and its coordinates are marked for subsequent processing. At the same time, the cumulative number of deviation points and their distribution positions should be recorded together. Through such a comparison process, it is possible to check one by one whether multiple roads, mountains or other terrain features on the map are consistent with the reconstructed data. During the inspection process, if the road coordinates are obviously misaligned, the coordinate deviation here will be registered separately in the east-west and north-south directions, and observe whether the deviation is related to the lidar scanning angle or the inclination of the vehicle-mounted equipment. If similar deviations appear in certain consecutive positions, it means that there is an overall tendency of coordinate system shift. It is necessary to further compare the track records of the vehicle's route and compare them with the rotation angle during the scan. If the deviation increases when the rotation angle exceeds 5 degrees, the rotation angle is classified as a key investigation factor. Finally, a preliminary comparison result is obtained through the continuous comparison and deviation summary of the above-mentioned geographic location and altitude information.

[0130] After completing the comparison results, it is necessary to correct the distortion or deviation that may be caused by the scanner equipment. First, all the coordinates marked as significant deviation points are separated from the overall point cloud list and compared in detail with the distribution of adjacent coordinates. If the jump of elevation data in the same area exceeds 3 meters and there are similar jumps in four or five consecutive points, the area is listed as a key correction area. The possible source of scanning error is found by recording the tilt angle of the vehicle during driving and the vertical axis correction curve in the reconstructed data. Then, the difference distribution of these tilt angles and the nearby point cloud distribution is compared one by one. If a positive correlation trend is found between the tilt angle and the difference distribution, the tilt correction is adjusted according to the trend. Positive coefficient, the specific value of the tilt correction coefficient is usually between 0.01 and 0.02 when the vehicle roll does not exceed 10 degrees, and it can be increased to between 0.03 and 0.05 when the roll is between 10 and 20 degrees. Record whether the vertical difference of each coordinate after correction has dropped significantly. If the difference is still above 2 meters after one correction, increase or decrease the correction coefficient again until the difference in most key correction areas drops to within 2 meters or connects smoothly with adjacent areas. If there are still very few points with excessive height differences, check separately whether there is equipment jitter or laser beam obstruction and confirm it in the relevant records. Finally, all coordinate points are corrected in batches and summarized to obtain optimized point cloud data.

[0131] The steps to obtain obstacle positioning data are as follows:

[0132] Based on the optimized point cloud data, the boundary points of all obstacles are extracted. By calculating the geometric characteristics of spatial distribution point by point, the point sets that constitute the obstacles are identified, and the point sets are classified and aggregated to obtain the preliminary boundary information of the obstacles.

[0133] Based on the preliminary boundary information of the obstacle, the boundary points of each obstacle are reconstructed by contour fitting, and the length, width and height of the reconstructed curve are extracted to obtain the size characteristics of the obstacle;

[0134] Obstacle location data is generated based on the size characteristics of the obstacle and combined with the spatial distribution information of the boundary points.

[0135] Specifically, based on the optimized point cloud data obtained above, firstly, all the peripheral points of the suspected obstacles are selected and listed in the order of three-dimensional coordinates. Then, for each point, the distance and direction are compared with several adjacent points in the X, Y, and Z directions. If the spatial distribution difference between a point and an adjacent point in a certain area exceeds 2 meters and the value of more than 2 meters is the average position deviation obtained by combining the actual measurement results of the road surface and the driving conditions of the vehicle plus a standard deviation, then this difference is marked as a significant dividing line, and the positions of these dividing lines are recorded and accumulated. If multiple significant dividing lines appear continuously and the connection directions of these dividing lines are roughly the same, then this direction is defined as the possible outer boundary direction, and then other potential abnormal difference points in the entire data set are searched and the repeated coordinate combinations are summarized. If the spatial aggregation degree is within the range of 0 to 2 meters and is densely distributed, it is regarded as the same outer edge part. If the spatial aggregation degree exceeds 2 meters, it is checked whether there are multiple independent obstacle boundary points. After preliminary positioning of all possible boundary points, the aspect ratio of the minimum enclosing area corresponding to each boundary point is calculated. For example, the X-axis direction is limited to the range of -50 meters to 50 meters, and the Y-axis direction is limited to the range of -100 meters to 100 meters. When a concentrated length of more than 5 meters is detected in one direction and the distribution in the other direction does not exceed 1 meter, it may be a narrow and long outer edge. This method can gradually aggregate boundary features of different forms into the same obstacle group. Finally, the aggregated boundary point set is uniformly labeled and the degree of difference in coordinates of adjacent boundary points is recorded to determine the preliminary outline of each obstacle in the plane dimension, thereby obtaining the preliminary boundary information of the obstacle.

[0136] Based on the preliminary boundary information of the obstacle obtained earlier, the outer boundary points of each obstacle are first arranged in order of distance and the direction differences of adjacent points are compared one by one. If the direction difference between two or three consecutive points is less than 10 degrees, they are divided into the same continuous segment. If it exceeds 10 degrees, it is divided into separate segments. Then, when performing contour fitting for each segment, the residual distribution of the points in the segment in the curve fitting is first calculated. If the mean value of the residual distribution exceeds 0.5 meters, it is necessary to search whether there is an arc jump or data outlier in the segment and record them one by one. When the mean value of the residual distribution is less than 0.5 meters, it means that the data in this segment can be fitted with a relatively smooth curve. The line fitting method is described. The fitting process can first record the position information of each key inflection point on the curve from the perspective of coordinate offset, and then make up the curve height difference between the inflection points through numerical interpolation. After completing the contour fitting of each segment, the corresponding curve segment length, width and corresponding height difference can be obtained and accumulated. If the height of some segments exceeds 2 meters, they can be marked as high obstacles during classification. If the width is less than 1 meter, it is marked as a narrow obstacle. Through this contour reconstruction-based method, the data of all segments are summarized and the length, width and height of each obstacle are finally obtained to obtain the size characteristics of the obstacle.

[0137] According to the obstacle size characteristics obtained above, the size value corresponding to each obstacle is combined with its position distribution information in the coordinate system. For example, the X direction is limited to -100 meters to 100 meters, the Y direction is limited to -150 meters to 150 meters, and the Z direction is limited to -5 meters to 5 meters. The position offset of each obstacle point set is continuously compared according to the measured trajectory of the vehicle. If there is a misalignment of more than 5 meters in the X direction and similar results appear multiple times in this area, this belt is listed as a high-position offset belt. If there is a height difference of more than 1 meter in the Z direction, the corresponding obstacle is marked as a significant vertical variation. After the size and coordinate characteristics of all obstacles are summarized, they are compared one by one with map marks or known scene records to eliminate the repeated parts that are misidentified. Finally, their final position relationship is established in the plane and height coordinates. These position information and size characteristics are packaged, integrated and recorded into a complete positioning information sequence, so that the obstacle positioning data can be generated.

Claims

1. The road obstacle perception system for unmanned vehicles based on laser radar is characterized by: The system comprises: The laser radar data acquisition module adjusts the laser radar scanning frequency, collects point cloud data from static and dynamic obstacles, and obtains preliminary scanning data; filters the preliminary scanning data, removes noise data, retains shape information, and generates optimized scanning data; A dynamic obstacle analysis module calculates the gradient strength of each point in the point cloud based on the optimized scanning data, identifies key obstacle points, performs time series analysis on the key obstacle points, determines the movement state and trajectory of the obstacle, and generates moving obstacle trajectory information; The laser radar data optimization module receives the moving obstacle trajectory information, uses Delaunay triangulation to spatially reconstruct key obstacle points to obtain spatially reconstructed data; compares the spatially reconstructed data with existing map data, optimizes the point cloud data structure, and generates optimized point cloud data; A real-time perception data output module generates the position and size information of obstacles in real time based on the optimized point cloud data to obtain obstacle positioning data; The steps for obtaining the key obstacle points are as follows: Based on the optimized scan data, a deep learning model is trained to identify and calculate local geometric features in the point cloud data to obtain a spatial variation index for each point; Based on the spatial variation index of each point, the gradient strength of each point is calculated, and the calculation formula is: in, Represents the reflection intensity function of the point cloud, It is the first The three-dimensional coordinates of the point, represents the partial derivative in the corresponding coordinate direction, It is The gradient strength of a point, is the adjustment coefficient, which indicates the intensity of neighborhood influence, Yes and The spatial weight between Yes The neighborhood point set of and They are points and The reflection intensity; According to the gradient strength of each point, a threshold is set, and the key obstacle points are identified through threshold comparison and screening; The steps for obtaining the moving obstacle trajectory information are as follows: Based on the key obstacle points, time series analysis is performed, and the data points of the continuous time frame are processed using a Kalman filter to estimate the state and speed of each point to obtain time-dependent speed and position data; Based on the time-dependent speed and position data, the speed of each point is calculated using the following formula: in, Indicates time At point speed, and The time and Interval exist and The displacement difference in direction, is the time interval between two frames, is a decay factor; According to the speed of each point, a threshold is set to determine whether it is a moving obstacle, and the moving obstacle trajectory information is obtained.

2. The laser radar-based road obstacle perception system for unmanned vehicles according to claim 1 is characterized in that: The steps for obtaining the preliminary scanning data are: Set the scanning frequency of the laser radar to collect raw point cloud data from static and dynamic obstacles to obtain raw point cloud data; Based on the original point cloud data, signal attenuation compensation is performed to remove irrelevant information to obtain cleaned point cloud data; Based on the cleaned point cloud data, the spatial association and structural link between the data points are calculated, the spatial position and structure of the object are mapped, the three-dimensional shape of the object is reconstructed, and preliminary scanning data is obtained.

3. The laser radar-based road obstacle perception system for unmanned vehicles according to claim 1 is characterized in that: The steps for obtaining the optimized scan data are as follows: Based on the preliminary scan data, low-frequency noise is removed by a high-pass filter, and high-frequency noise is removed by a low-pass filter to obtain scan data after preliminary noise reduction; Based on the scan data that has undergone preliminary noise reduction, edge detection is applied to retain the edges and corner points of the object to obtain optimized scan data.

4. The laser radar-based road obstacle perception system for unmanned vehicles according to claim 1, characterized in that: The steps for obtaining the spatial reconstruction data are: Based on the moving obstacle trajectory information, Delaunay triangulation is performed, and the formula is: in, Indicated by point is the sum of the areas of the triangles at the vertices, and They are points and neighboring points The two-dimensional coordinates of Yes The set of adjacent points of ; The sum of the triangle areas is used to perform spatial reconstruction, and the sum of the triangle areas is mapped into a three-dimensional spatial structure to obtain spatial reconstruction data.

5. The laser radar-based road obstacle perception system for unmanned vehicles according to claim 1, characterized in that: The steps for obtaining the optimized point cloud data are as follows: Based on the spatial reconstruction data, the spatial reconstruction data is compared with the existing map data, including checking the geographical location and height information of each data point, to obtain a preliminary comparison result; Based on the preliminary comparison results, the distortion or deviation caused by the limitation of the scanner device is corrected to obtain optimized point cloud data.

6. The laser radar-based road obstacle perception system for unmanned vehicles according to claim 1, characterized in that: The steps for obtaining the obstacle positioning data are as follows: Based on the optimized point cloud data, the boundary points of all obstacles are extracted, the geometric characteristics of spatial distribution are calculated point by point, the point sets constituting the obstacles are identified, and the point sets are classified and aggregated to obtain preliminary boundary information of the obstacles; Based on the preliminary boundary information of the obstacle, a curve is reconstructed for the boundary points of each obstacle through contour fitting, and the length, width and height of the reconstructed curve are extracted to obtain the size characteristics of the obstacle; Obstacle location data is generated based on the size characteristics of the obstacle and combined with the spatial distribution information of the boundary points.

Images