CN109186625B

CN109186625B - Method and system for accurately positioning intelligent vehicle by using hybrid sampling filtering

Info

Publication number: CN109186625B
Application number: CN201811243970.9A
Authority: CN
Inventors: 李明喜
Original assignee: Beijing Auto Brain Technology Co ltd
Current assignee: Beijing Auto Brain Technology Co ltd
Priority date: 2018-10-24
Filing date: 2018-10-24
Publication date: 2020-05-05
Anticipated expiration: 2038-10-24
Also published as: CN109186625A

Abstract

The invention discloses a method and a system for accurately positioning an intelligent vehicle by using mixed sampling filtering. A new probability positioning frame is adopted, a random sampling consistency algorithm and a particle filter/histogram filter algorithm are efficiently fused together by using a kernel density estimation method, the respective advantages are fully utilized to make up the deficiency of the other side, and then a probability grid map based on a Gaussian mixture model is used for more robust, accurate and rapid positioning.

Description

Method and system for accurately positioning intelligent vehicle by using hybrid sampling filtering

Technical Field

The invention belongs to the technical field of intelligent driving, and particularly relates to a method and a system for accurately positioning an intelligent vehicle by using hybrid sampling filtering.

Background

The intelligent driving technology relates to a plurality of disciplines such as cognitive engineering, vehicle engineering, electronic and electrical engineering, control science and engineering, artificial intelligence and the like, and the intelligent vehicle is an important mark for measuring national research strength and industrial level. The appearance of the unmanned vehicle fundamentally changes the vehicle driving mode in the traditional 'man-vehicle-road' closed loop system, frees a driver from fatigue driving, controls the vehicle to run by utilizing an advanced sensor and an information technology, automatically completes the conventional, lasting, low-grade and repeated operations in the driving activity, can greatly improve the efficiency and safety of a traffic system, improves the quality of human mobile life, and has wide social application value. Meanwhile, the research of the intelligent driving technology can greatly enhance the core competitiveness of China in the aspects of automobile active safety and the like, and has great strategic significance for improving the independent innovation capacity of automobile electronic products and automobile industry in China.

The intelligent vehicle further comprises parts such as environment perception, accurate positioning, information fusion, intelligent decision making, vehicle control and the like on the basis of vehicle dynamics. Among them, the positioning technology is a very critical basic work in the field of automatic driving. The positioning accuracy based on a Global Navigation Satellite System (GNSS) can only reach about 10m, and the automatic driving requirement cannot be met. Satellite positioning based on carrier phase differential (RTK) can maintain centimeter-level accuracy in an environment with good signals, but the signals are easily shielded by surrounding tall objects or generate multipath effect, so that the fluctuation of a positioning result and the reduction of the positioning accuracy are caused. A combined positioning technique combining RTK-GNSS with an Inertial Measurement Unit (IMU) or odometer can improve the problem to some extent. However, because inertial navigation and odometer have accumulated errors, maintaining high accuracy positioning in environments where satellite positioning is unstable or ineffective for a long time requires very high accuracy in inertial navigation, which is very expensive.

In order to realize high-precision positioning within a relatively low cost range, currently, the mainstream method is to control the positioning precision within a usable range by using LiDAR or vision in combination with a high-precision map constructed in advance and by matching real-time data with map data. Currently, vision-based costs are lower than LiDAR, but the ranging accuracy is low and susceptible to light. The accuracy and reliability of LiDAR are relatively superior to vision.

Lidar may provide three-dimensional position information and reflected intensity information of the surface of an object in the surrounding environment, both of which may be used for localization. The method has high positioning precision and is slightly influenced by dynamic targets such as vehicles/pedestrians, but the method has limited application scenes. The road surface marks become blurred in rainy and snowy weather, so that the positioning precision is influenced; on a non-intersection road section, the road surface lacks longitudinal characteristics, and the longitudinal positioning precision is difficult to ensure; in unstructured roads such as rural and off-road roads, the road surface may be completely free of road marking information, and the positioning method may be completely ineffective.

The positioning application scene based on the height information is wider, is less influenced by weather, and is suitable for both structured roads and unstructured roads. One key problem faced, however, is overcoming the effects of moving objects (moving vehicles and pedestrians), varying static objects (city renovations, parked vehicles, growth of foliage) and the resulting occlusion in the environment on LiDAR positioning.

And the two kinds of information are fused for positioning, and the weights of the two kinds of information are self-adaptively adjusted according to the confidence coefficient of a positioning result. However, when the ground mark is polluted and a large amount of dynamic targets exist in the surrounding environment, the method is difficult to obtain a stable and accurate positioning structure.

The core problem of LiDAR positioning is the registration problem of real-time data and map data, and can be divided into the registration of a point cloud map and the registration of a grid map according to different map types.

Point cloud map-based registration is commonly used in scene reconstruction and SLAM, and two representative methods are iterative closest point ICP and gaussian distribution transformation NDT. Many researchers make contributions to the development of two types of registration methods, and provide a series of excellent improved algorithms, such as GICP, ML-NDT, IRON and the like, so that the robustness of the algorithms is improved.

Registration based on these methods usually requires defining a cost function for registration of two frames of point clouds and iteratively minimizing the registration cost by calculating a function gradient. Due to the nature of the gradient descent, these methods are highly dependent on the initial position and are affected by local minima. Under the conditions of large initial error and large environmental noise, the method is easy to fall into local optimum, and often cannot achieve ideal registration effect. Therefore, they are usually used as local precise registration algorithms, before which coarse registration is usually required to filter noise and provide a more ideal initial position for precise registration.

A Random Sample Consensus (RANSAC) based method is one of the most commonly used rough registration methods at present, many scholars improve the RANSAC algorithm from feature extraction, and four-Point matching (4-Point consistency Set, 4PCS) and the improved algorithm thereof carry out RANSAC search by using coplanar four points meeting affine invariant constraint, so that even if the coincidence area of two frames of Point clouds is small and the initial position deviation is large, a good registration result can still be obtained, and the method has wide application. However, in the urban environment faced by unmanned driving, a large number of dynamic targets often exist, and it is very difficult to use a random sampling mode to simultaneously find four static target points, so that the algorithm efficiency is greatly reduced. The ML-RANSAC algorithm solves the point cloud registration problem under a high dynamic scene through the established multi-layer architecture of sample-iteration-sample. The invention expands the algorithm and integrates the algorithm into the probability positioning frame based on the grid map.

In grid map based registration, particle filtering is one of the most widely applied algorithms in the field of mobile robots. The algorithm has excellent performance when the problem that prior distribution is very complex or unknown is processed, and under the condition that the particle number is sufficient and the distribution is reasonable, even if the search space is large, the optimal solution can be found with high probability after multiple iterations. With the iteration, the particles are gradually converged, so that the accuracy of the algorithm is gradually improved, but the diversity of the particles is also lost, so that the particle depletion is caused, the robustness of the algorithm is reduced, and the risk of falling into the local optimum is increased in a large-scale high-dynamic scene.

Considering the unmanned application scenario, a satellite positioning system can generally limit the positioning error within a range of several meters, and google proposes a Histogram Filter (HF) algorithm. This approach offers the significant advantage of not having to worry about the loss of particles near the correct position, and has also become one of the mainstream methods of LiDAR positioning in recent years. Both histogram filtering and particle filtering have the problem of low efficiency when the search space is large, and the problem is more prominent in high-dimensional situations. Aiming at the problem, the Lucas-kanade (LK) algorithm is firstly adopted to determine the heading of the vehicle, and then the histogram filtering is used for positioning two dimensions of the horizontal position, so that the positioning efficiency in a large search space can be improved, but the LK algorithm is easy to lose effectiveness when the deviation of the heading angle is overlarge.

Although in most cases the error in the satellite positioning results will not be particularly large, there are some more extreme cases. For example, in a tunnel, the occlusion of the tunnel can cause the satellite positioning failure, while the feature in the longitudinal direction of the tunnel can cause the LiDAR positioning failure, the relative positioning can be performed only by dead reckoning, and after the vehicle travels a certain distance and time in the environment, a large accumulated error can be generated, and at the moment, if the vehicle travels to a fork in the tunnel, the LiDAR and the map are required to be used for positioning under a large initial deviation. Furthermore, unmanned vehicles also encounter problems like "kidnapping", such as starting up in underground garages, where the initial position error is usually very large, sometimes hundreds of meters, and the heading of the vehicle is unknown. The positioning in this case is also to be solved by the autopilot.

Disclosure of Invention

The invention provides a method and a system for accurately positioning an intelligent vehicle by using mixed sampling filtering, which realize the accurate positioning of the vehicle in a non/weak satellite signal environment by using the mixed sampling filtering and combining high-precision map data constructed in advance and real-time point cloud data acquired by a laser radar or a camera, and can be used for the auxiliary driving and unmanned driving of the vehicle.

In order to achieve the purpose, the invention adopts the following technical scheme: firstly, a probability grid map is constructed by using a Gaussian mixture model, then a new probability positioning frame is adopted, a method based on random sampling consistency and a method based on Bayesian filtering are efficiently fused together by using a kernel density estimation method, the respective advantages are fully utilized to make up for the deficiency of the other side, and more robust, accurate and rapid positioning is carried out.

And performing kernel density estimation on all sampling results in the random sampling consistency method instead of only the result with the maximum ratio of the interior points, updating Bayesian filtering by taking the kernel density estimation as prior distribution, and then taking the prediction distribution of the Bayesian filtering as prior of a new round of random sampling at the next moment.

The method comprises the following specific steps:

a. the method comprises the steps of projecting point cloud data collected in advance into two-dimensional grids according to corresponding accurate world coordinates, wherein each grid comprises height information and reflectivity information, the two kinds of information are represented by a Gaussian mixture model, the number of components in the Gaussian mixture model is not less than 1, the grids are squares, and the side length is not less than 5 cm.

b. Denoising, segmenting and feature extracting are respectively carried out on the collected real-time point cloud and the map data, and then registration based on RANSAC is carried out on the extracted feature points.

c. And (3) taking the proportion of the interior points obtained by sampling each time meeting the constraint condition in RANSAC as the weight or confidence coefficient of the position, and clustering and fitting all the results in the RANSAC by using a kernel density estimation method to obtain the parameter of each kernel. Wherein the threshold value d of the distance of the inner point_inlierDetermines the error range of the coarse registration and the bandwidth h of the kernel density estimation_KDEAnd the sampling space size w of the filtering algorithm in the accurate positioning_spaceAre determined by the error range. Therefore, through the framework, we can naturally combine the key parameters of the three parts: an inlier distance threshold value in RANSAC, a bandwidth in Gaussian kernel density estimation and an initial sampling space in filtering positioning are unified together.

h_KDE＝k₁·d_inlier

w_space＝k₂·d_inlier

d. And (3) obtaining a plurality of accurate high-probability positions by using a Bayes-based filtering method such as histogram filtering near the distributed peak values, when the number of the peak values exceeds 1, using the histogram filtering for each peak value, and calculating the posterior probability of the positions at the current moment according to the kernel density function and the prior probability at the moment. And using the position with the maximum posterior probability as the final position of the positioning, and calculating the position variances according to the probability to obtain the confidence coefficient of the positioning.

e. And respectively calculating prior distribution of the next moment positioning corresponding to the positions by utilizing the motion state information output by the inertial measurement unit or the odometer according to the Markov principle and the Bayes rule.

Wherein x is_tIndicating the position of the vehicle at time t,

indicates the predicted value, u, of the vehicle at time t +1_t+1Representing the motion state of the vehicle from the time t to the time t +1, and the vehicle state information obeying parameter at the time t +1 is (mu)_t+1,σ_t+1) Gaussian distribution of (Δ x)_t+1,tTo represent

And x_tThe amount of change in (c).

Correspondingly, a system for accurately positioning the intelligent vehicle by using mixed sampling filtering is also provided, which comprises a mixed sampling filtering unit, a high-precision map unit constructed in advance and a real-time point cloud data unit collected by a radar or a camera,

and the hybrid sampling filtering unit is utilized to combine the pre-constructed high-precision map unit data and a real-time point cloud data unit acquired by a radar or a camera, so as to realize the accurate positioning of the vehicle in the environment without/with weak satellite signals.

The mixed sampling filtering unit is a probability positioning frame which adopts a kernel density estimation method to fuse a random sampling consistency method and a Bayes filtering method, carries out kernel density estimation on all sampling results in the random sampling consistency method instead of only the result with the maximum proportion of interior points, updates the Bayes filtering by using the result as prior distribution, and uses the prediction distribution of the Bayes filtering as the prior of a new round of random sampling at the next moment.

The point cloud data are projected into two-dimensional grids by the aid of the pre-constructed high-precision map unit, each grid comprises height information and reflectivity information, the height information and the reflectivity information are both represented by a Gaussian mixture model, the number of components in the Gaussian mixture model is not less than 1, the grids are squares, and the side length is not less than 5 cm.

The real-time point cloud data unit acquired by the radar or the camera comprises an original point cloud directly obtained by analysis from a sensor and a characteristic point cloud extracted by algorithm processing.

Compared with the prior art, the method has the advantages that the requirement of automatic driving positioning on expensive high-precision inertial navigation equipment is eliminated, and high-precision real-time positioning meeting the automatic driving requirement can be realized under the condition of greatly reducing the cost;

the method has wider application range and scene, and can be used for urban structured roads, rural semi-structured roads and off-road unstructured roads; the method can be used for static or low-dynamic scenes and can also be used for high-dynamic scenes with various vehicles, pedestrians and large environmental changes; the solar energy electric heating water heater can work under good weather conditions and can be used in rain, snow and fog days;

the tolerable initial pose deviation can reach hundreds of meters, the accurate positioning in extreme scenes such as vehicle underground garage starting, long-term tunnel driving and the like can be realized, and important technical support is provided for the realization and mass production of automatic driving at the L3, L4 and L5 levels.

Drawings

FIG. 1 is a satellite map of an experimental section of a preferred embodiment of the present application;

FIG. 2 is a schematic diagram of the motion trajectory before and after positioning using the method of the present invention;

FIG. 3 is a schematic diagram comparing positioning results of three different methods;

in fig. 3: (a) a lateral error, (b) a longitudinal error, (c) a course angle error;

FIG. 4 is a schematic diagram of the positioning result of the intersection;

in fig. 4: (a) the method comprises the following steps of (a) positioning a top view of the example on a high-precision map, (b) positioning a 3D view of the example on the high-precision map, (c) positioning a front-view camera view of a scene, (D) positioning distribution of histogram filtering, (e) positioning distribution of multilevel random sampling consistency, (f) positioning distribution of the example;

FIG. 5 is a schematic diagram of the positioning result of the straight road;

in fig. 5: (a) the method comprises the following steps of (a) positioning a top view of the example on a high-precision map, (b) positioning a 3D view of the example on the high-precision map, (c) positioning a front-view camera view of a scene, (D) positioning distribution of histogram filtering, (e) positioning distribution of multilevel random sampling consistency, (f) positioning distribution of the example;

FIG. 6 horizontal position error distributions under different initial deviations;

in fig. 6: (a) a scatter plot, (b) a histogram;

FIG. 7 is a flow chart of method steps of the present application.

Detailed Description

The invention is described in further detail below with reference to the figures and specific examples. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

It should be noted that the embodiments and features of the embodiments in the present application may be combined with each other without conflict.

In the embodiment, a velodyne HDL-64E laser radar is used for point cloud data acquisition, the algorithm runs on a 4-core industrial personal computer with the main frequency of 2.40GHz, the internal memory is 8GB, and the system adopts ubuntu 16.04.

The present example uses a gaussian model with two components to model the reflectivity and height distribution of each grid separately. For the height information, two Gaussian components are used for respectively describing the height distribution of the ground and the obstacles; for the reflectivity information, two components are used to describe the distribution of high reflectivity data and low reflectivity data, respectively, on the ground. The grid size was 20 cm.

In the embodiment, a region growing algorithm based on multivariate geometric constraint is adopted to synchronously de-noising and segmenting the laser radar point cloud and the grid map respectively, and the gravity center point of each segmented target is extracted as a feature point for registration to carry out ML-RANSAC based on 'sample-iteration-sample'.

Kernel density estimation is a nonparametric method for estimating probability density functions，x₁,x₂,...,x_nThe probability density function of n independent and identically distributed sample points is set as f, and the kernel density is estimated as follows:

where K (.) is a kernel function, a gaussian kernel function is used in this example. h >0 is the smooth bandwidth (bandwidth).

The example estimates three parameters (x, y, θ) of the vehicle, in which model only the horizontal position is modeled and estimated, i.e. x is (x, y), and the heading angle of the vehicle uses the heading angle corresponding to the pose with the largest weight in each gaussian kernel.

The Bayes filtering is a filter based on Bayes principle, and comprises Kalman filtering, particle filtering, histogram filtering and the like, and the histogram filtering of the current mainstream is adopted in the embodiment. In order to improve the positioning speed, the histogram filtering is improved by adopting a branch-and-bound method.

In the example, a real urban road scene is selected for testing, and the testing road section is shown in fig. 1. The total length of the road section is 7.3km, and the road section comprises urban internal main roads, large-scale intersections, urban expressways, viaduct gates and the like. The time interval between the acquisition of the high-precision map of the road section and the acquisition of the experimental data is more than 1 year, one is data in winter, and the other is data in summer. The environment of many places has changed to some extent, and fig. 1 shows two sets of changing scenes, one set is the change of road facilities at the intersection, the other set is the change of vegetation on two sides of the road in winter and summer, and the vegetation is provided with the anti-freezing fence in winter but not in summer. These static scene changes can increase the difficulty of LiDAR positioning. We compare the algorithm proposed by the present invention with histogram filtering HF based positioning methods (hereinafter HF) and ML-RANSAC based positioning methods (hereinafter ML-RANSAC).

According to different types of initial position deviation, the experiment is divided into 2 parts which are all carried out in a road scene containing a large number of dynamic targets.

1 continuously variable position deviation

Experiment one is to test the positioning accuracy and stability of the algorithm in a high dynamic city environment with a large number of vehicles and pedestrians by using only one IMU with lower accuracy to calculate the vehicle pose without using GNSS and using the calculation result of inertial navigation as the input of LiDAR positioning. The accumulated error generated by pure estimation is continuously and slowly changed, and no position jump exists.

Meanwhile, the vehicle running track based on the differential GPS is collected, and the offline LiDAR SLAM technology is used as real data of the vehicle position on the basis of the vehicle running track. The experiment included two phases, AB and CD. FIG. 2 shows real data of a vehicle travel trajectory, an IMU derived trajectory, and a LiDAR located trajectory based on the method of the present invention, respectively.

The positioning errors of the three methods in the experiment are shown in figure 3, and it can be seen that the average error and the maximum error of the method are the minimum of the three. The HF precision is high, the average error is close to the error of the method, but the algorithm robustness is slightly poor, in some dynamic scenes, due to the shielding and interference of dynamic targets, the registration error is easy to occur, the ML-RANSAC transverse positioning stability is superior to that of HF, but in some scenes with less longitudinal features, because a filtering positioning algorithm and a probability framework of the method are not combined, reasonable confidence degree evaluation is lacked, longitudinal fluctuation is easy to occur in the positioning result, and the overall positioning precision is poor.

Fig. 4 and 5 show the positioning results for two different scenarios. In both sets of scenarios, the positioning results for HF and ML-RANSAC are biased to different degrees. Wherein (a) (b) respectively represent the top view and 3D view of the location results of LiDAR using the method of the present invention in a high-precision map, wherein the high-precision map is displayed with a grayscale map, red shows real-time point cloud data for location, blue represents the static object obtained by registration, and yellow line is the locus of LiDAR location. (c) The scene graph is shot by the vehicle-mounted front-view camera. (d) (e) (f) is probability distribution fitted by the positioning results of the three methods in each scene, and it can be seen that the peak value of the positioning distribution of HF deviates from the correct position obviously due to the interference of dynamic targets in the crossing scene, and ML-RANSAC is closer to the true position; in a straight road scene, due to the lack of the longitudinal features of the environment, the positioning distribution of HF has a large variance in the longitudinal direction, while the positioning distribution of ML-RANSAC has a plurality of peaks in the vehicle longitudinal direction. By fusing the two methods, the invention obtains positioning distribution with accurate peak value and smaller variance in the two scenes.

2 position deviation of discrete random jumps

Experiment two further promotes the degree of difficulty on the basis of experiment one, and increases a random offset as the input of LiDAR location on the calculation result that has accumulated error, simulates the location sudden change situation that GPS produced because of incomplete shielding and multipath effect, tests the precision and stability of the algorithm under the condition that jump error exists. The experiment is divided into three groups, the course angle offset randomly fluctuates within +/-5 degrees, and the horizontal position offset randomly fluctuates within the radius ranges of 2.5m, 5m and 10m respectively.

Fig. 6 shows a localization result horizontal distance error distribution scatter diagram and an error distribution histogram of the three localization methods at different initial deviations on the test road section. In order to more clearly present the distribution of the positioning errors, the ordinate of fig. 6(a) uses logarithmic coordinates, and it can be seen that the positioning accuracy of the method of the present invention is insensitive to the initial deviation, the positioning accuracy is always kept at a higher level, and the overall errors and fluctuations of the positioning results of HF and ML-RANSAC are larger. In this experiment, we only counted the results of LiDAR positioning, and rejected the extrapolated results of IMU between two LiDAR positioning. Because the time consumption for completing one-time positioning under different initial errors is different in the three methods, the total time is the same on the same section of test data, the positioning times in different initial error ranges are different, and the positioning times of the three methods have a decreasing trend with obvious positioning time effect along with the increase of the initial error.

As can be seen from fig. 6(b), the number of times that the horizontal error of the algorithm is less than 0.1m is much larger than that of other methods, and the error of the positioning result is centrally distributed within 0.5 m.

Table 1 shows the statistical results of the positioning of the three methods under different initial deviations, including the root-mean-square errors of the horizontal, longitudinal, course angle and horizontal position, the horizontal errors being less than the percentages of 0.1m and 0.4m, the mean and standard deviation of the algorithm time consumption. When the initial deviation is 10m, the positioning times of HF within the error of 0.4m only account for 9.09% of the total positioning times, and the method still accounts for 93.18%.

TABLE 1 comparison of positioning statistics under different initial position deviations

The experimental result shows that the algorithm maintains higher positioning precision and faster positioning speed under the condition of continuously accumulated initial errors or discrete random jitter initial errors. Compared with the histogram filtering algorithm and the multilevel random sampling consistency algorithm before improvement, the method further improves the positioning robustness on the premise of ensuring the positioning precision and speed, and can adapt to more complex positioning scenes; and the fast and accurate positioning can still be kept under the condition that the initial position has large deviation, which cannot be realized by the algorithm before the improvement.

The foregoing is only a preferred embodiment of the present invention, and it should be noted that, for those skilled in the art, various modifications and decorations can be made without departing from the principle of the present invention, and these modifications and decorations should also be regarded as the protection scope of the present invention.

Claims

1. A method for accurately positioning an intelligent vehicle by using mixed sampling filtering is characterized by comprising the following steps: the method comprises the following steps of adopting mixed sampling filtering, combining high-precision map data constructed in advance and real-time point cloud data collected by a radar or a camera, and realizing accurate positioning of a vehicle in a non/weak satellite signal environment;

the mixed sampling filtering is a probability positioning frame which adopts a kernel density estimation method to fuse a random sampling consistency-based method and a Bayes filtering method, carries out kernel density estimation on all sampling results in the random sampling consistency method instead of only the result with the maximum proportion of interior points, updates the Bayes filtering by taking the result as prior distribution, and takes the prediction distribution of the Bayes filtering as the prior of a new round of random sampling at the next moment;

the method comprises the following specific steps:

a. projecting point cloud data collected in advance into two-dimensional grids according to corresponding accurate world coordinates, wherein each grid comprises height information and reflectivity information, the two kinds of information are represented by a Gaussian mixture model, the number of components in the Gaussian mixture model is not less than 1, the grids are squares, and the side length is not less than 5 cm;

b. denoising, segmenting and feature extracting are respectively carried out on the collected real-time point cloud and the map data, and then RANSAC-based registration is carried out on the extracted feature points;

c. using the proportion of the interior points obtained by sampling each time satisfying the constraint condition in RANSAC as the weight or confidence of the position, clustering and fitting all the results in the RANSAC by using a kernel density estimation method to obtain the parameter of each kernel, wherein the distance threshold value d of the interior points is_inlierDetermines the error range of the coarse registration and the bandwidth h of the kernel density estimation_KDEAnd the sampling space size w of the filtering algorithm in the accurate positioning_spaceAll are determined by the error range, so by the framework, the key parameters of the three parts: an interior point distance threshold value in RANSAC, a bandwidth in Gaussian kernel density estimation and an initial sampling space in filtering positioning are unified together;

h_KDE＝k₁·d_inlier

w_space＝k₂·d_inlier

d. using a Bayesian filtering method to perform histogram filtering near the distributed peak values to obtain a plurality of accurate high-probability positions, when the number of the peak values exceeds 1, using the histogram filtering for each peak value, calculating the posterior probability of the positions at the current moment according to the kernel density function and the prior probability at the current moment, using the position with the maximum posterior probability as the final position of the current positioning, and calculating the position variance according to the probability to obtain the confidence coefficient of the current positioning;

e. according to Markov principle and Bayes rule, the motion state information output by inertial measurement unit or milemeter is used to calculate the prior distribution of next time positioning corresponding to these positions,

wherein x is_tIndicating the position of the vehicle at time t,

And x_tThe amount of change in (c).

2. The method for smart vehicle fine positioning using hybrid sampling filtering as claimed in claim 1, wherein: the real-time point cloud data collected by the radar or the camera comprises an original point cloud directly analyzed from a sensor and a characteristic point cloud extracted through algorithm processing.

3. The utility model provides a system that intelligent vehicle utilized mixed sampling filtering to carry out accurate positioning which characterized by: comprises a mixed sampling filtering unit, a high-precision map unit constructed in advance and a real-time point cloud data unit collected by a radar or a camera,

the hybrid sampling filtering unit is combined with the high-precision map data unit constructed in advance and a real-time point cloud data unit acquired by a radar or a camera to realize the accurate positioning of the vehicle in the environment without/with weak satellite signals;

the method comprises the following specific steps:

h_KDE＝k₁·d_inlier

w_space＝k₂·d_inlier

wherein x is_tIndicating the position of the vehicle at time t,

And x_tThe amount of change in (c).

4. The system for smart vehicle fine positioning using hybrid sampling filtering of claim 3, wherein: the real-time point cloud data unit acquired by the radar or the camera comprises original point cloud directly analyzed from a sensor and characteristic point cloud extracted through algorithm processing.