Study on Obstacle Detection Method Based on Point Cloud Registration

Abstract

An efficient obstacle detection system is one of the most important guarantees for improving the active safety performance of autonomous vehicles. This paper proposes an obstacle detection method based on high-precision positioning applied to blocked zones to solve the problems of the high complexity of detection results, low computational efficiency, and high load in traditional obstacle detection methods. Firstly, an NDT registration method which uses the likelihood function as the optimal value of the registration score function to calculate the registration parameters is designed to match the scanning point cloud and the target point cloud. Secondly, a target reduction method combined with threshold judgment and the binary tree search algorithm is designed to filter the point cloud of non-road obstacles to improve the processing speed of the computing platform. Meanwhile, KD-tree is used to speed up the clustering process. Finally, a vehicle remote control simulation platform with the combination of a cloud platform and mobile terminal is designed to verify the effectiveness of the strategy in practical application. The results prove that the proposed obstacle detection method can improve the efficiency and accuracy of detection.

1. Introduction

It is essential for intelligent driving technology to determine the position of autonomous vehicles through precise positioning. However, applying GPS merely cannot be accurate to the centimeter level required by intelligent vehicles [1]. Although Simultaneous Localization and Mapping (SLAM) can achieve high-precision navigation and positioning, it is difficult to achieve online mapping and positioning [2,3]. Therefore, registration positioning based on high-precision point cloud maps and perception sensors is often applied to intelligent vehicles. The positioning method with LIDAR basically meets the positioning needs of intelligent vehicles and has the advantages of high accuracy and strong reliability [4,5].

The Normal Distribution Transform (NDT) algorithm [6] and Iterative Closest Point (ICP) algorithm are more applied in registration positioning [7,8,9]. Although the ICP algorithm works with high accuracy, it is challenging to eliminate uncertainty problems in dealing with local environmental differences [10]. Additionally, registrations like ICP take too much time to return information to intelligent vehicles, and it is hard to meet the real-time requirements of autonomous driving [11,12]. NDT is a rough registration algorithm that relies on the normal distribution of the point cloud to determine the transformation relationship between the registration point and the target point [13,14]. This algorithm first converts the target point into a normal distribution of multi-dimensional variables [6], and it applies optimization algorithms to find the transformation parameters that maximize the sum of the probability density of the transformed point in the coordinate system [15,16]. And because NDT is a probabilistic registration method, it can eliminate the impact of local environmental differences on the results. Also, it can achieve centimeter-level high-precision positioning in a short time by selecting an appropriate grid size [16].

It is inevitable to encounter various obstacles even when intelligent vehicles drive on a prescribed route. In order to improve the active safety of autonomous vehicles, it is necessary to detect and classify the perceived obstacles, then find a reasonable obstacle avoidance strategy according to the results. In this paper, depending on whether the obstacle may affect driving, the detected obstacles are roughly divided into non-road obstacles and road obstacles.

Non-road obstacles are static obstacles which are located on the side of the road and will not hinder basic driving, such as trees, buildings, and municipal constructions. Semantic segmentation or deep learning which is based on visual sensor detection is used for non-road obstacle detection [17,18]. In ref. [19], the authors processed SAR satellite images with the deep fully convolutional neural network algorithm to identify the road and non-road environment. In ref. [20], the authors proposed a detection method which can significantly improve the accuracy of semantic segmentation and detected objects by combining basic and advanced image features. Most of the above detection methods are feature extraction and are not suitable for LIDAR. In recent years, obstacle detections based on LIDAR have made great achievements. In ref. [21], the authors proved that it is reliable to detect buildings based on the concept of high entropy by qualitative and quantitative analysis. In order to overcome the limitation of the LIDAR detection of vegetation, ref. [22] estimated tree size and frequency distribution applying the ITC algorithm. Yi and his colleagues came up with a naive Bayes–RANSAC method, which was used to detect building plans, reduced pseudo-planes, and improved the efficiency of RANSAC [23]. Previous practice has proven that although the detection of non-road obstacles can be achieved, this process requires considerable computing power for segmentation and clustering. These operations would reduce the real-time performance and seriously affect the safety of driving.

Road obstacles are vehicles, pedestrians, roadblocks, scattered objects, and other static or dynamic obstacles around the road which may affect driving safety. Vision-based or vehicle-mounted LIDAR-based detection methods are often used to detect such obstacles [24,25]. In ref. [26], the authors identified static and dynamic objects according to time in the event camera to reduce latency. Considering the reliability of the algorithm, the application of LIDAR is more extensive. In the detection process, processing methods such as point cloud filtering, ground segmentation, and clustering are usually used to obtain the location information of obstacles [27,28], and the detected obstacles are identified and classified. Lee et al. proposed a Geometric Model-Free Approach with a Particle Filter (GMFA-PF), which realized the good recognition and tracking of moving objects in the driving environment [29]. Jin et al. designed a ground segmentation algorithm based on a Gaussian process, which effectively solved the problem of over-segmentation and under-segmentation in the traditional ground segmentation algorithm [30]. Lin et al. proposed an adaptive clustering segmentation method to further segment the preliminary segmented point cloud according to the standard deviation of clustering points and realized the efficient segmentation of scene point cloud data [31]. Additionally, Miller and his partners took screenshots of the point cloud by frame and detected the screenshots with a visual method. However, this method can only obtain obstacle information on a two-dimensional plane [32]. In ref. [33], the authors illustrated the extraction and classification of 3D point cloud features with the help of machine learning algorithms and achieved the purpose of identifying obstacles. However, it is not suitable for changing environments without a rich data set.

In this paper, an obstacle detection process based on high-precision positioning is proposed, including an obstacle detection method based on high-precision positioning and an obstacle map generation method based on filtering non-road factors. Details are explained in Figure 1. The main contributions of this paper are as follows:

(1)

A point cloud data matching and clustering algorithm based on the combination of the KD-tree and NDT matching algorithm is designed, which accelerates the process of location recognition, reduces the deviation between the scanned point cloud and point cloud map, and realizes centimeter-level high-precision positioning.

(2)

A target reduction method based on threshold partitions is proposed. The point cloud is further divided on the basis of down-sampling, which reduces the amount of calculation of the platform, realizes the reduction in communication load, and speeds up the processing speed of external information in the vehicle perception system.

(3)

A test platform with the combination of a cloud platform and mobile terminal is designed, and the real situation of the data communication processing of intelligent vehicle remote control is simulated by the Mulran data set and ROS, which proves the effectiveness of the obstacle detection algorithm based on a high-precision map in practical application.

The rest of the paper is structured as follows: first, the proposal is introduced in Section 2, then the experiment is carried out, the results are analyzed in Section 3, and finally the conclusions and prospects are explained in detail in Section 4.

2. Proposed Obstacle Detection Methods

In order to improve the computational efficiency of obstacle detection and obtain better detection results, an obstacle detection method based on the high-precision location of the occlusion region is proposed. Firstly, taking the maximum registration score of the likelihood function as the optimization objective, the high-precision point cloud matching based on NDT method is realized; then combining threshold judgment and the binary tree search algorithm, the non-road obstacles are filtered; and finally, by using KD-tree to accelerate the segmented clustering progress, the identification of road obstacles affecting vehicle driving is realized. Figure 1 describes the structure of the proposed obstacle detection method.

2.1. Registration Positioning

The purpose of this section is to realize the registration and location based on the point cloud map. The registration score function based on the likelihood function is designed. The system first grids the point cloud map and calculates the probability density function of each grid and then takes the maximum registration score as the optimization goal to obtain NDT registration parameters which match the point cloud and the target point cloud. Finally, according to optimal registration parameters, the registration is realized, and the alignment point cloud is obtained [15]. After the above operations, high-precision map registration is completed. The specific positioning process is shown in Algorithm 1.

Algorithm 1 Register scan Pr to reference Pm using NDT

1: Inputs: Registration Point-clouds: Pr and Allocate cell structure: C

2: Output: Aligned Point-clouds: Pa

3: for scan k = 1 to $N\in C$ do

4: $p(x)\leftarrow density(\mu ,\sum ,y_k)$

5: end for

6: $\psi \leftarrow {\mathrm{\Pi }}_{k=1}^{n}p(T(\overrightarrow{p},{\overrightarrow{x}}_k))$

7: $score(\tilde{p}({\overrightarrow{x}}_k),{\overrightarrow{x}}_k)\leftarrow G(\psi )$

8: while not converged do

9: $scor{e}_{\max }\leftarrow Optimize(score(\tilde{p}({\overrightarrow{x}}_k),{\overrightarrow{x}}_k))$

10: end while

11: $T(\overrightarrow{p},{\overrightarrow{x}}_k)\leftarrow scor{e}_{\max }(\tilde{p}({\overrightarrow{x}}_k),{\overrightarrow{x}}_k)$

12: $P_a\leftarrow Trans(T,P_r)$

2.1.1. Grid Processing

The size of voxels is marked as ‘C’ meters in this paper, as shown in Figure 2. The normal distribution is adopted to normalize the voxels. Then, the information of the point cloud can be represented by the probability density function of the grid. In order to obtain the probability density function of each grid, it is necessary to calculate the mean and covariance matrix of the corresponding grid, and in the calculation, all scanning points in a grid are represented by y_k (k = 1, 2, 3…).

Then, the mean of the normal distribution of a grid is as follows:

$$\mu =\frac{1}{\mathrm{m}}{\displaystyle \sum _{k=1}^m{y}_k}$$

(1)

The expression of the covariance matrix is as follows:

$$\Sigma =\frac{1}{m}{\displaystyle \sum _{k=1}^m(y_k-\mu )}{(y_k-\mu )}^T$$

(2)

Finally, the probability density function of a grid is as follows:

$$p(x)=\frac{1}{{(2\pi )}^{3/2}\sqrt{\left|\Sigma \right|}}{e}^{-\frac{{(x-\mu )}^T{\Sigma }^{-1}(x-\mu )}{2}}$$

(3)

2.1.2. NDT Registration

In order to obtain the best NDT matching results, we use the likelihood function to record the position conversion process of all the points in the space, and the expression is shown as follows:

$$\psi ={\mathrm{\Pi }}_{k=1}^{n}p(T(\overrightarrow{p},{\overrightarrow{x}}_k))$$

(4)

where $T(\overrightarrow{p},{\overrightarrow{x}}_k)$ is the spatial transformation function, $p(T(\overrightarrow{p},{\overrightarrow{x}}_k))$ represents the conversion process, and $\psi$ represents the likelihood function.

And in order to speed up the process of solving the optimal registration parameters, this paper conducts a negative logarithmic transformation on the likelihood function and approximately simplifies it by using the Gaussian function, and the following formula can be obtained:

$$score(\tilde{p}({\overrightarrow{x}}_k),{\overrightarrow{x}}_k)=-{\displaystyle {\sum }_{k=1}^n\tilde{p}(T(\overrightarrow{p},{\overrightarrow{x}}_k))}$$

(5)

Then, the problem of solving the maximum value of the likelihood function is transformed into solving the minimum value of the negative logarithmic function.

At the same time, the Newton optimization algorithm is used to calculate the maximum value of the objective function. When the optimization result meets the convergence condition, the optimization process stops, and the optimal registration score is obtained. After the optimization, the spatial transformation function will be calculated according to the best registration score, and taking the point cloud to be registered P_r and the spatial transformation function as input, the point cloud P_a which is aligned with the target point cloud P_m after registration will be obtained.

2.2. Non-Road Factor Obstacle Detection

Vision-based image processing methods [17] or airborne LIDAR detection methods [22] are often used to detect buildings or trees. However, these methods require a large amount of calculation, and it is difficult to meet the randomness requirements with them. There are tiny differences for the point cloud of non-road obstacles in different situations; thus, P_a and P_m after positioning have a high degree of registration. But they have a low matching degree of road obstacles because of the randomness and uncertainty. Therefore, in order to reduce the amount of calculation of the computing platform, improve the computing speed of the computing platform, and speed up the recognition process, this paper designs a target reduction method to further deal with the point cloud after ground segmentation. The specific process is shown in Algorithm 2. Firstly, P_a is regarded as the objects to carry on ground segmentation to reduce the amount of data involved in the calculation. Secondly, according to the distribution regulation of the point cloud map, the distance threshold required for the target reduction method is calculated. Finally, according to the target reduction method, P_a is classified.

Algorithm 2 Non-road factor obstacle detection

1: Inputs: Aligned Point Clouds: Pa and Map Point Clouds: Pm, D is the dimension of the ray.

2: Output: Non-Road Point Clouds: PNroad and Road Point Clouds: PYroad

3: for scan j = 1 to N do

4:  $P_j\leftarrow Segment(D,P_a)$

5:  for scan $l=1$ to $M\in P_m$ do

6:   $(P_{Nroad},P_{Yroad})\leftarrow DropT\arg et(P_j,P_m)$

7:  end for

8: end for

2.2.1. Ground Segmentation

The basis to identify the ground point and non-ground point is whether the slope of the two adjoining points on the same ray meets the set slope threshold [34]. However, the slope partition method requires a large number of multiplication and division operations, which requires a high load capacity of the computing platform. In order to reduce the number of multiplication and division operations of the computing platform in the process of ground segmentation, improve the calculation speed, and ensure the real-time performance of the vehicle perception system, a ground segmentation method based on height division is designed in this paper. The system first determines the position of the point cloud according to the ray sequence and the radius of the laser point. Then, the system sorts the points on the same ray and marks them as P_ord. Ground recognition will be realized by comparing the altitude difference between the two adjoining points and the set altitude threshold (h_min, h_max): If the altitude of current point h_c is not within this range and the previous point is a ground point, the current point is also a ground point. If h_c is within this range and the distance s_i between two points is greater than the distance threshold s between two adjacent points, the current point is a ground point. If the above conditions are not met, the point is not a ground point, as shown in Figure 3.

2.2.2. Calculation of Distance Threshold

In order to obtain more reliable obstacle classification results, it is necessary to calculate a reasonable distance threshold to divide the spatial range: Figure 4 shows that the distance between two measurement points on the X-axis obtains the maximum value at 0°. So, the distance at 0° is chosen as the X-axis distance threshold ∆x; the change in the Y-axis is the same as that in the X-axis, so the distance threshold of the Y-axis ∆y equals ∆x; the distance threshold of the Z-axis ∆z is related to the density of point cloud contours.

After considering the distribution characteristics of the point cloud, the distance threshold required for point cloud registration is calculated as follows:

$$\Delta y=\Delta x=\frac{r^2{\cos }^2(\omega )\pi }{2000}$$

(6)

$$\Delta z=K\frac{r\sin (\omega )}{r\sin (\omega )-1}$$

(7)

where r is the measured distance, ω is the vertical angle of the laser, and K is the gain coefficient.

2.2.3. Target Reduction Method

After determining the distance threshold, in order to obtain the target point cloud corresponding to the point cloud to be processed, it is necessary to determine the range of the target point cloud P_m involved in reduction processing. This paper divides the space from the direction of the three axes of the vehicle coordinate system. Taking the X-axis as an example, the range is (P_m._xmin, P_m._xmax). If the range calculated based on a point P_j in the non-ground point cloud P_Nground contains some points which are in P_m, then P_j will continue to be included in the Y-axis judgment process. The formulas of P_m._xmin and P_m._xmax are as follows:

$$P_{m.x\mathit{min}}=P_{j.x}-\Delta x$$

(8)

$$P_{m.x\mathit{max}}=P_{j.x}+\Delta x$$

(9)

If there are some points in the target point cloud that fall within the range of the Y-axis (P_m.ymin, P_m.ymax) and the Z-axis (P_m.zmin, P_m.zmax) calculated based on P_j, and the number of these points exceeds the target threshold p, then P_j is considered as a non-road factor obstacle point; otherwise, the point is considered as a road factor obstacle point. When all points of P_Nroad are traversed, P_a will be divided into non-road obstacle points P_Nroad and road obstacle points P_Yroad. Considering that the data volume of P_m is too large, in order to ensure real time and accuracy, we choose the data of P_m with the current vehicle location as the center and the radius less than 30 m.

2.3. Road Obstacles Detection

Due to the varying degrees of impact that different obstacles have on driving safety, further obstacle detection is required for the road factor obstacle point cloud P_Yroad after the target reduction processing. In this paper, road factor obstacle detection is realized by piecewise clustering. The detection process is shown in Algorithm 3, which includes point cloud clipping and filtering, segmented clustering, and obstacle identification. Firstly, road obstacle point P_Yroad is clipped and filtered to reduce the amount of calculation, then the point cloud is divided into regions according to the radius, and the corresponding clustering radius is used for different regions. Finally, the clustering results are analyzed to complete the obstacle detection.

Algorithm 3 Road factor obstacle detection

1: Inputs: Road Point Clouds: PYroad

2: Output: Obstacle: PO

3: for scan a = 1 to N do

4:  $P_a\leftarrow Cut(h,P_{Yroad})$

5:  $P_a\leftarrow Filter(P_a,d)$

6:  $P_C\leftarrow Cluster(tree(Down(P_a,leaf)),r_e,s_e,q)$

7:  $P_O\leftarrow Frame(P_C)$

8: end for

2.3.1. Tailoring and Filtering

Since the detection range of LIDAR is 1~120 m, there are problems such as too much irrelevant point cloud information caused by the excessive detection range or radar reflection caused by the vehicle body. So, to reduce the unnecessary calculation of the computing platform, it is necessary to tailor and filter P_Yroad. Since obstacles with excessively high elevation do not affect the safety of ground vehicles, this paper filters out the point cloud data above a certain elevation h. At the same time, in order to prevent errors caused by the reflection of LIDAR rays from the vehicle body, we filter the point cloud around the vehicle body according to the size of the information collection vehicle. Here, we filter the point clouds whose distance r is less than d. The point cloud after clipping and filtering is marked as P_f.

2.3.2. Segmented Clustering

Cluster processing is suitable for environment perception in low-speed and simple scenarios [35]. This method mainly includes steps such as down-sampling processing, nearest point searching, and clustering.

The KD-tree algorithm is a method to retrieve the closest point data to the query point in the KD tree to achieve the search purpose. Here, the KD-tree template in the PCL library is selected for accelerating processing. By setting the size of the leaf, we achieve the purpose of reducing point cloud density and meeting the real-time requirements of clustering. The Euclidean clustering algorithm is a mature point cloud segmentation algorithm [36,37]. The main parameters that affect the algorithm are the cluster radius threshold r_e, minimum number of cluster points threshold q_min, and maximum number of cluster points threshold q_max. This paper divides the detection range from near to far according to the distance radius R_i and selects the cluster radius threshold r_e from small to large. It should be noted that the smaller r_e is, the longer the clustering time is, the larger the minimum clustering point threshold q is, and the poorer the clustering effect is. We record the clustered point cloud cluster as P_C for obstacle modeling.

2.3.3. Establishing Obstacle Models

In order to accurately estimate the motion state of the obstacle, the frame model is used to represent the obstacle. The frame model of the obstacle can be calculated by using the point cloud cluster P_C obtained by the clustering processing. Firstly, the centroid of the cluster point cloud O_i is calculated and used as the center of the obstacle. Then, the length, width, and higher parameters of the point cloud cluster are calculated according to the furthest distance from the center point in each direction of P_C. Finally, a border is added to obtain the road factor obstacle. And, in order to obtain the obstacle model with the greatest impact on driving safety and the shortest modeling time, we set the side length of the obstacle in the same direction or perpendicular to the driving direction of the vehicle.

3. Experiment and Analysis

In the experiment, to simulate the real situation of the remote control of an intelligent vehicle, an experimental platform combining a cloud platform and mobile terminal is designed in this paper. And the KAIST of the Mulran data set and ROS PCL library are used to test the proposed method. The KAIST sequence data come from an OS1-64 3D LIDAR (Ouster, San Francisco, CA, USA) and a radar. The radar is located behind the LIDAR with an angle of about 70° that is lost in the horizontal direction [38]. In this paper, the function of registration positioning and obstacle detection are evaluated using qualitative evaluation and quantitative evaluation. At the same time, the computational cost in obstacle detection is recorded to evaluate the efficiency of this strategy.

3.1. Experimental Platform Design and Experimental Parameters

The cloud platform used in the experiment is Huawei Cloud Platform, which provides multiple computing nodes for this experiment. For ease of comparative analysis, each computing node is equipped with X86 architecture, Intel Cascade Lake 2.6 GHz CPU, and 4 GB of memory. The baseline bandwidth and maximum bandwidth are both 1 G/s. The image running on this computing node is Ubuntu 16.04 server 64-bit (40 GB). In order to ensure the reliability of the functionality, the experiment also purchased load balancing and elastic scaling services. The VPC configuration required for the elastic scaling service is consistent with the VPC of the aforementioned computing nodes. This experiment is also equipped with an industrial control computer based on X86 architecture and the Ubuntu Linux operating system, which is responsible for processing sensor data, communicating with the cloud, and sending instructions to the core control module as the onboard computing center. To reduce operational complexity, the experiment chose the same Ubuntu 16.04 operating system as the cloud end.

The parameters that affect the results of obstacle detection based on high-precision positioning are shown in

Table 1. Parameters for obstacle detection.

Symbol/Unit	a/cm	s/cm	p/-	h/m	d/m	leaf/-	qmin/-	qmax/-
Value	2	0.2	5	1.98	2.4	0.1	5	50

. The integrated navigation system adopts the NAV712/SU7 GNSS/INSS system (Beijing Nuogeng SCI-TECH Development co., LTD; Beijing, China), and the LIDAR uses C16 LIDAR(LSLIDAR, Shenzhen, China). The parameters of the integrated navigation system and LIDAR are shown in

Table 2. Parameters for LIDAR.

LIDAR Parameter	Value/Unit	LIDAR Parameter	Value/Unit
Laser band	905/nm	Measuring range	70~200/m
Laser channel	16/-	Ranging accuracy	±3/cm
Vertical field angle	±15/°	Horizontal field angle	360/°

and

Table 3. Parameters for IMU.

IMU Parameter	Value/Unit	IMU Parameter	Value/Unit
Gyro type	MEMS/-	The zero bias stability of the gyroscope	2/(°/h)
Gyro range	±300/(°/s)	The range of the accelerometer	±6/g

.

3.2. Functional Test of Registration and Positioning

3.2.1. Quantitative Evaluation

The evaluation content of the registration positioning results includes the target point error and quantitative index. The evaluation method adopts the visual analysis method with a high utilization rate, that is, by analyzing the results of different detection sequences and quantitatively evaluating the integrity of non-road factor obstacle identification and the accuracy of road factor obstacle identification. In this paper, the registration conversion rate P, the number of registration iterations NI, and the types of obstacles involved in matching are used to evaluate the results of high-precision positioning. The registration conversion rate P is a measure used to evaluate the registration effect of a certain frame, and the formula is as follows:

$$P=\frac{score}{size}$$

(10)

where score refers to the best registration score score_max, and size is the number of grids participating in registration.

Figure 5 and

Table 4. Results of NDT registration.

Seq. Name	Vehicle Situation	Time/(ms)	P	NI	Involved Obstacles
					Trees	Building
1	Stationary	152.616	6.73321	2	√	√
2	Moving	419.857	6.71391	7	√	√
3	Moving	771.78	6.93716	13	√	√
4	Moving	538.09	6.75882	11	√	/

show the matching effect of the scanned point cloud before registration and the alignment point cloud and map after registration. We calculate statistics on the registration conversion rate and the number of registration iterations of the selected 500 frame point clouds. The results show that the average registration conversion rate is 6.5526, and the average registration iteration number is 4.98. This paper not only uses the registration conversion rate and the number of registration iterations as the result analysis index but also adds three analysis indexes: vehicle driving state, registration time, and obstacle matching. In sequence 1, the situation of the vehicle is stationary. There is a short distance between the scanned point cloud and the aligned point cloud; it takes less time to perform iterations and registration. At the same time, there is a lot of vegetation and buildings involved in the matching, which provides conditions for registration positioning. In sequences 2–4, the vehicle is moving. The scanning point cloud is far away from the alignment point cloud, so it takes a longer time to perform iterations and registration, which makes the registration more difficult. Although there are more obstacles in sequence 2 than in sequence 3, there are more ground points in sequence 3 which results in more iterations. The positioning effect of sequence 4 performs well even when there are large amounts of obstacle data, because the point cloud in the figure can be well adapted to the outline of trees. Generally speaking, by observing the point cloud matching of different sequences, it can be concluded that the point cloud matching method designed in this paper can achieve high-precision positioning.

3.2.2. Qualitative Evaluation

In this paper, the average road offset ratio $\overline{D}$ is selected here to express the overall offset result of the measured point cloud, and the formula is as follows:

$$\overline{D}=\frac{\tan {\alpha }_1+\tan {\alpha }_2}{2}$$

(11)

where α₁ is the lane left boundary offset rate, and α₂ is the lane right boundary offset rate.

In Figure 6, sequences 5–6 are the first section and the middle section of the road, respectively. The calculation results show that the $\overline{D}$ of P_r in sequence 5 is −2.619%, and the $\overline{D}$ of P_a is −0.087%. In sequence 6, the $\overline{D}$ of P_r is 1.828%, and the $\overline{D}$ of P_a is 0.262%. The road offset range before positioning is 0~66.7 cm and after positioning is 0~10.5 cm. The range of positioning offset is reduced by 84%. The positioning error of the automatic driving system above L4 must be controlled within 20 cm. The lane positioning error of this paper is 0~10.5 cm, so the registration and positioning method designed in this paper can meet the needs of high-precision positioning.

3.3. Function Test of Obstacle Detection

3.3.1. Quantitative Evaluation

We select multiple different scenarios in the data set to study the algorithm for distinguishing road/non-road obstacles and the detection algorithm for road obstacles. The results are shown in Figure 7. In the experiment, the obstacles detected on the road include vegetation, buildings, pedestrians, vehicles, road construction, and other objects.

1.

The distinction of road/non-road obstacles.

In the results, there are 325 road obstacles and 843 non-road obstacles. The number of the missed inspections is 24. Therefore, 27.83% of obstacles may affect driving safety.

Table 5. The results of road/non-road obstacle distinction.

Seq. Name	Vehicle Situation	Obstacle Type		Missed Obstacles	False Obstacles
		Road Factor	Non-Road
7	Turning	18	27	1	2
9	Turning	10	17	0	1
11	Straight	14	17	1	1
13	Straight	22	21	0	3

illustrates the test results. Table 5 contains the number of correct, missed, and false road/non-road obstacles of sequences 7, 9, 11, and 13 in Figure 7. The results show that most of the missed obstacles are vehicles parked closely on the roadside (sequence 7), and there are three types of obstacles which are detected incorrectly: trees with low height and sparse point clouds (sequence 13); vehicles that are closer to trees (sequences 9 and 11); and vehicles classified as the same obstacle (sequence 7). By comparing the angle between the vehicle and the obstacle, the accuracy of the recognition result is related to the angle between the vehicle and obstacle direction: if the angle is acute or right, the detection performed with high accuracy (sequence 11); if the relative angle is obtuse, there are the results with low accuracy (sequence 7).

2.

The detection of road obstacles

In the test results, there were 479 obstacles in total, of which 67 were pedestrians, 290 were vehicles, and 16 were missed. As a result/therefore, 74.53% of obstacles may affect driving safety; this shows a significant improvement in the effectiveness of obstacle detection.

Table 6. The result of road obstacle detection.

Seq. Name	Vehicle Situation	Obstacle Type		Missed Obstacles	False Obstacles
		Pedestrians	Vehicles
8	Turning	0	17	1	3
10	Turning	2	8	2	3
12	Straight	2	12	1	2
14	Straight	4	18	1	2

reports the number of correct, missed, and false obstacles of sequences 8, 10, 12, and 14 in Figure 7. The detection results show that most of the undetected obstacles are vehicles (sequences 8 and 10), and the wrong obstacles are low-height and luxuriant vegetation (sequences 10 and 14) and obstacles treated as the same obstacle (sequences 8 and 12). After analysis, the missed vehicles are those parked on the side of the road. The reason is that some vehicles overlap the point clouds of nearby trees so that they are all clustered as vegetation (sequence 10), and this is also the reason why the vehicle is blocked by other obstacles during driving. But in general, the factors that may affect driving safety are mainly located in front of the vehicle, and the coincidence of point clouds at different angles is different. Therefore, even if they are regarded as the same obstacles, they can be identified by the strategy designed in this paper. This means that the strategy in this paper can meet the requirements of obstacle identification and ensure safe driving.

The results of the distinction of road/non-road obstacles and the detection of road obstacles show that the efficiency of obstacle detection after using the target reduction method is significantly improved, and the effective rate of the detected obstacles is increased by 46.7%. Road obstacles that affect driving safety are completely retained, while non-road obstacles that do not affect driving safety are successfully filtered except for those who have a small structure or special location. Among them, important detection results such as pedestrians or vehicles are also ideal.

3.3.2. Qualitative Evaluation

We compared and analyzed the detection results of all sequences in Figure 7 to evaluate the proposed method qualitatively according to the attributes of the detection object. All sequences involved in the analysis have achieved high-precision positioning.

Sequences 7 and 8 are the detection results of the steering vehicle. The objects in sequence 7 are classified as obstacles, including pedestrians, trees, vehicles, and buildings. Because the steering angle of the vehicle forms a certain angle with the road, and the clustering did not perform well due to the lush trees, some road obstacles are identified as non-road obstacles. By reducing the target processing, most of the non-road obstacles in the sequence are no longer involved in the detection except for small trees, and vehicles and pedestrians near the vehicle are successfully detected. Sequences 9 and 10 are also the detection results of the steering vehicle. Obstacles detected in sequence 9 also include buildings, pedestrians, vehicles, and trees. Since the vehicle forms a certain angle with itself when it is parked sideways, obstacles 10–13 are identified as the same obstacle. But vehicles or pedestrians can be nested in the frame of buildings or trees. By reducing the target processing, sequence 10 filters out most of the trees and buildings, while the original detected vehicles and missed vehicles are retained. Sequences 11 and 12 are the results of the straight vehicle. The three pedestrians and vehicles along the way in sequence 11 are completely detected, but the frame of vegetation may cover the vehicle. After the target reduction process, the pedestrians and vehicles close to the vehicle are completely detected in sequence 12 except for the distant vehicles surrounded by vegetation. However, small trees on the side of the road are also identified as obstacles of the road. Sequences 13 and 14 are also the results of straight-through detection. The vehicles close to their own vehicle in sequence 13 are all detected completely, and there is no overlap. Parallel pedestrians are clustered as the same obstacle. By reducing the target processing, sequence 14 completely retains all of the detected vehicles and pedestrians. Distant vehicles are not recognized, but they do not have much of an influence on safe driving. The test result of straight vehicles is slightly better than that of steering vehicles. But generally speaking, road obstacles that affect the safety of vehicle driving can be effectively identified.

3.4. Load Assessment

The computational cost comes from the target point cloud gridding, down-sampling processing, and prediction of the current frame pose during NDT registration. Additionally, it comes from target reduction processing and road obstacle detection. This section mainly analyzes the difference in obstacle detection efficiency caused by different methods after positioning. Experimental sequences 7~14 are used to analyze the computational cost. Figure 8 reports the difference in calculation efficiency between the proposed obstacle processing method and the original processing method; the calculation unit is ms/frame. The results show that the proposed method can improve the calculation efficiency. It can be seen from Figure 9 that the load required for obstacle detection in different driving states is reduced more or less by reducing target processing. Although the load change in sequence 11 (12) with a larger proportion of road factors and obstacles is small, only 1.4%, other sequences have achieved the goal of significantly reducing the load percentage.

In general, there is a contradiction between calculation efficiency and the obstacle detection effect, but the method proposed in this paper effectively solves the above problems. By reducing the calculation of road obstacle detection, the strategy in this paper not only improves the positioning effect of the self-vehicle but also realizes the accurate positioning of obstacles and improves the accuracy of target reduction processing and the detection performance of non-road obstacles and then improves the work efficiency of the algorithm.

4. Conclusions and Prospects

To improve the capability of obstacle detection, which can make intelligent driving significantly safer, this paper proposes an obstacle detection method based on high-precision positioning for blocked zones. To meet the positioning requirements, we use the NDT registration algorithm to match the registered point cloud with the point cloud map. In order to shorten the calculation time and improve efficiency in the obstacle detection process, we propose a target reduction method without reducing the precision of positioning. Finally, the following conclusions are obtained through a combination of theoretical analysis and experimental verification:

(1)

The obstacle detection technology based on registration and location designed in this paper not only shortens the obstacle detection time but also improves the detection accuracy, which provides conditions for the real-time control of the vehicle.

(2)

The positioning error of the registration and positioning technology in this paper is controlled within 10.5 cm. It can meet the needs of high-precision positioning in the process of the automatic driving of intelligent vehicles.

(3)

The target reduction method in this paper can realize the classification of road factor obstacles and non-road factor obstacles. This method reduces the number of irrelevant obstacles and the detection complexity.

(4)

Using KD-tree to accelerate clustering processing and obstacle models establishing processing can meet the requirements of obstacle detection. The processing method of this paper can quickly and accurately obtain obstacle position and shape information.

Although the method proposed in this paper can improve the performance of obstacle detection in blocked zones, there is still some work to conduct to improve the performance further. In this paper, the information involved in obstacle detection comes from a single vehicle, but the information obtained by a single vehicle is limited. In view of this deficiency, the environmental information of obstacles perceived by multi-vehicles will be combined to update the fixed obstacles of structured roads in real time (such as roadblocks, etc.). In this idea, the vehicle control system can quickly obtain the static obstacles near the current driving position by subscribing to the global obstacle map service, shorten the obstacle detection time, improve the detection efficiency, and further improve the reliability of the system.