Dynamic obstacles detection of tram based on laser radar

KUANG Wen-zhen， WU Meng-luo， XU Li

(School of Automation & Electrical Engineering, Lanzhou Jiaotong University, Lanzhou 730070, China)

Abstract： The detection of obstacles in a dynamic environment is a hot and difficult problem. A method of autonomously detecting obstacles based on laser radar is proposed as a safety auxiliary structure of tram. The nearest neighbor method is used for spatial obstacles clustering from laser radar data. By analyzing the characteristics of obstacles, the types of obstacles are determined by time correlation. Experiments were carried out on the developed unmanned aerial vehicle(UAV), and the experimental results verify the effectiveness of the proposed method.

Key words： laser radar; tram; dynamic obstacle detection; spatial obstacle clustering; time correlation; nearest neigbor method

0 Introduction

The modern tram is a rail transit system between the bus and the light rail, and it can adapt to various roads. Different from the subway and the train, the tram often runs the road that overlaps or crosses the motor vehicle road. Because the road conditions are rather complex, there are some factors that threaten the safety of tram running[1-4], the three major factors of which include the status of the driver, the stability of the equipment and the obstacles in front of the tram. The detection of obstacles such as vehicles and pedestrians at the intersection is related to the safe operation of the tram. Usually, tram drivers observe the road ahead through the naked eye, which requires drivers not only to operate trams, but also to observe dangerous situations ahead of time and take corresponding operations in time. Therefore, in the long run of tram, the driver is very easy to produce visual fatigue, affecting the safety of the tram.

In practical situations, stationary object is easy to be detected. But when the tram and the object move simutaneously, the detection will be in a non-static state. In this paper, laser radar is used for the detection and recognition of moving objects. This tram may bump or swing in the process of running, and radar measurement signals may be temporarily lost. These problems need to be solved.

Generally, dynamic obstacles detection can be divided into three stages: target selection, target information detection and target consistency check. The target selection is to primarily select effective targets from all the objects. Among them, the effective targets include any static or moving object nearest to the tram. The target information detection is to predict the motion states of the effective targets at the future time, so as to prepare for the consistency check of the targets. The target consistency check is determined and evaluated by the similarity between the primary target information and the target prediction value.

1 Obstacle detection using laser radar

Most of the existing detection methods are based on static environment, and the detection precision of static objects is high. But real environment is always dynamic. When the tram is running, the ground obstacles will also move, which makes the obstacle detection be in a non-static background. The detection of the motion obstacle in the dynamic environment is a difficult point of environment modeling. Currently, there are cameras, laser radar and other obstacle sensing sensors for it. Owing to its high precision, good performance, and so on, laser radar has been widely used[5-8]. The representative research is the dynamic occupying grid algorithm proposed by Bisw[5]. However, there are still some problems in the algorithm. For example, assuming that the shapes of objects are unchanged, though objects are easily detected from the background, fast moving objects can not be detected. Wang, et al. put forward a dynamic background detection method[6], which is multi-hypothesis tracking data association. But this method can not detect temporary static obstacles. On the basis of the study of the time-varying potential method, Yang, et al. proposed a multi-resolution potential field method[8]. This method is applicable to path planning and obstacle avoidance in mobile robot navigation. But how to use laser radar to distinguish static and dynamic obstacles is not involved. In this paper, a method of detecting independent dynamic obstacles based on laser radar is proposed for tram obstacles detection by clustering the nearest objects scanned with laser radar, analyzing the relevance of the obstacles, and detecting and making decisions for the targets effectiveness.

2 Dynamic obstacle detection

2.1 Obstacle detection model

Laser radar has high accuracy and good robustness out of doors. A schematic diagram of obstacle detection is shown in Fig.1.

Fig.1 Diagram of obstacle detection

γ=ξcosσ+ηsinσ.

(1)

In addition,γon the scanning plane is always less thanδon the horizontal plane through a series of radar measurements (ξi,ηi), namely

(2)

whereδrepresents the distance of the imaging point from the center on the optical detector. In vertical irradiation mode, it is equal to the planar displacement of the object and can be measured directly.

Generally speaking, the scanning radius of the radar is far greater than the distance between two rails. Therefore, the curve of the road between the two rails can be approximated as a straight line. The changes in distance and angle between adjacent scanning points are relatively limited and generally meet the following conditions as

(3)

(4)

whereρdrepresents the distance between two points;Kdshows the angle between two points.

2.2 Clustering of dynamic obstacles

The laser radar is installed in the middle of the tram, away from the ground 2 m. The scanned area by the laser radar is within 0°-180°. By evenly dividing the scanned area at the angle of 1°, there are 181 scanned data points corresponding to 181 obstacle distance data points in the range of 0°-180°. And these obstacle distance data points can be represented by polar coordinates asPi(ρi,λi)(i=0,1,2,…,180), whereρiis the distance of the facula central points of the laser radar from scanned data point, andλiis the angle value of the scanned data point. Based on the clustering principle of adjacent points, if the distance between adjacent points is less than a predetermined thresholdT, these scanned data points are regarded as the same kind of obstacles.

Because the density of the scanned data points in the vicinity from radar sensor is higher than that in the distance, the nearest neighbor clustering method with dynamic threshold is adopted. Threshold selection is related to two factors: one is the distance between the facula central points of the laser radar, the other is the degree of attention to the obstacles. The former increases linearly with the extension of the beam; the latter is characterized by different resolutions corresponding to different obstacles such as the objects in the track and the trees along the track. For the area with a radius of 10 m ahead of the tram, the threshold value of 0.3 m is ideal, while for the area with a radius larger than 10 m, the threshold value of 0.03ρiis ideal.

The clustering algorithm is as follows:

1) Let the data pointPibelong to classCj(i=0,j=1)；

2) Calculate the distance fromPi+1toPiby

there isPi+1∈Cj, otherwisej=j+1,Pi+1∈Cj;

4) Ifi<179,i=i+1, turn to step 2, otherwise the algorithm ends.

2.3 Data association of obstacles

By analysis of the clustering of the obstacle distance data points, in order to further obtain the type of each obstacles subset and more detailed movement information, it is necessary to analyze the time correlation of the previous clustering obstacle data, which is to confirm the data correlation of obstacles subset.

Data association problem arises from sensor measurement process and multi-object uncertainty. In practical work, there are some unavoidable factors such as measurement error, noise interference of the sensor, etc. It is impossible to correctly judge whether the measurement data come from the real targets or other false targets, which will lead to a fairly vague relationship between the measured data and the corresponding target.

In recent decades, the research on data association methods has achieved fruitful results. The existing data association algorithms include the nearest neighbor algorithm[9-10], probabilistic data association filtering algorithm[11], multiple false attempts[12], and so on. In this paper, a relatively simple nearest neighbor algorithm is adopted.

The basic principle of the nearest neighbor correlation method is to construct a difference function to judge the difference of obstacles, which considers all kinds of characteristics of obstacles and weights them according to the actual situation. Only the obstacle that is the smallest difference from the target obstacle is considered to be related to it.

LetZAandZBbe the two obstacles that need to be verified in laser radar coordinates. The center coordinate ofZAis (xAm,yAm), the starting coordinate is (xAs,yAs), and the terminal coordinate is (xAe,yAe); the center coordinate ofZBis (xBm,yBm), the starting coordinate is (xBs,yBs), and the terminal coordinate is (xBe,yBe). The difference function is

H(ZA,ZB)=W1ΔD+W2Δα+W3ΔL,

(5)

where ΔDis the difference between the center coordinates distance ofZAandZB, there is

(6)

Δαis the difference between the directions ofZAandZB, there is

(7)

ΔLis the difference between the widths ofZAandZB, there is

(8)

W1,W2andW3are the weighted values of each feature, respectively, whose values are determined according to Ref.[9]. Generally, there are

W1=0.7,W2=0.05,W3=0.25.

2.4 Target effectiveness detection and decision-making method of obstacles

The target effectiveness detection and decision-making process of obstacles are shown in Fig.2.

First, the three-order Kalman filter is used to predict the effective target information in the whole life cycle. The state infromation in then-th detection cycle is expressed asXn=[dnvnan], wheredn,vnandanare the relative distance, velocity and acceleration between the effective target and the tram inydirection. The predictive values of the target state in the next cycle are expressed by

(9)

wheret0is the radar scanning period, which is 50 ms;d(n+1)|n,v(n+1)|nanda(n+1)|nare the state information of the effective target in (n+1)th cycle predicted by the data in then-th cycle.

Fig.2 Target effectiveness detection and decision making process

Comparing the measured values of the effective target with the predicted values in the (n+1)th cycle can determine whether they are consistent, namely

(10)

wheredn+1,vn+1andan+1are the state information of the effective target measured by radar;d0,a0andv0are the permitted errors between the measured values and calculated values, which can be obtained by experiments, there is

(11)

For the primary target obtained in then+1 cycle, if the measured values satisfy Eq.(11), it is considered that the primary target is consistent with the effective target obtained in thenth cycle, and then the target information is updated; otherwise it is necessary to deal with the target inconsistency. The effective target life cycle is introduced to represent the formation, continuity, tracking and extinction processes of the effective target.

3 Experimental results and analysis

The experimental platform is a self-developed four-axis unmanned aerial vehicle(UAV). As shown in Fig.3. UAV can display the fight direction, height and position by using 8 channel remote hand boats. It uses laser radar as an external sensor to realize the perception of the environment. The system processing software is Visual C++. In the course of moving, UAV uses laser radar to scan the environment ahead of operation to judge moving area or obstacle area. The speed of the moving area is 12 m/s, and the speed of the obstacle avoidance is 4 m/s.

Fig.3 Dynamic obstacles detection effect map

In Fig.3, the “*” point is the original data of the environment detected by the radar, the dark square represents the dynamic obstacle, the light square represents the static obstacle, and the triangle represents the location of the robot in the map.

Compared with the actual environment, Z14, Z23, Z28 and Z33 are moving cars, Z36 and Z47 are moving pedestrians, Z13, Z42, Z43 and Z53 are stationary cars, and the rest are stationary or low-speed pedestrians.

The motion obstacle Z23 is randomly selected as the estimation object, and the position and velocity of the motion in the global coordinate system are tracked in 140 s. The results show that the detection algorithm can accurately detect the dynamic obstacles.

4 Conclusion

For the moving obstacle detection in an unknown environment, a method based on 2D laser radar autonomous obstacles detection and data association is proposed. It determines the distribution of front obstacles by means of laser radar scanning, classifies the obstacles by nearest neighbor method, and analyzes the data association of the obstacles by multiple successive scanning processes.