reference
Real-time trajectory planning for autonomous urban driving: framework, algorithms and verifications
It uses a hierarchical motion planning framework. First, use the command behavior from senior planner to extract a rough path from the digital reference map. Conjugate gradient and nonlinear optimization algorithm for sequentially cubic Bspline smooth curves and interpolation reference path. In order to follow the reference path as well as improved handling of static and moving objects, the longitudinal trajectory planning task is solved in planning curve coordinate frame. Generating a rich set of feasible candidate path kinematic, dynamic flow for processing the reaction and in consideration. At the same time, speed of execution configuration generation to improve driving safety and comfort. After that, through careful evaluation of the objective function to generate the trajectory of the objective function to adjust the traffic to a combination of behavioral decision. Selecting the best collision-free, smooth, and possible dynamic trajectory, and by the lower lateral and longitudinal control command performs conversion. Our test autonomous vehicles conducted a field test on the reality of inner-city roads. Experimental results demonstrate the capacity and effectiveness of the framework and trajectory planning algorithm is proposed to deal with a variety of typical driving scenarios safely, static and moving objects such as avoidance, lane keeping and vehicles to follow, and obey the traffic rules.
Correlation
FIG pixel search method and random sampling algorithm.
Structural road path generation
research priorities
stratified framework. Senior planners dynamic behavior of complex traffic conditions responsible for reasoning and decision-making to develop discrete intentional maneuver, such as lane follow, lane change, vehicle tracking, overtaking slow travel and so on. The use of behavioral decision making, trajectory generation algorithm to generate dynamic feasible and assume responsibility for collision-free trajectories, these lower trajectory tracking controller can be easily tracked. This article focuses on the real-time trajectory generation algorithm.
LiDAR based positioning method by using a precise location information extracting lane level coarse reference path from the digital map, which is similar to [38] and the method of [39] proposed. In order to improve driving comfort and reduces control effort, using the nonlinear optimization conjugate gradient method and the cubic B-spline smoothing algorithm and sequentially inserted into the reference path.
Curve coordinate frame based on the trajectory planning task is decoupled space velocity planning and path planning subtasks. The trajectory optimization scheme is not used to generate a unique optimal trajectory, but can generate a rich set of candidate sub-optimal, which can overcome the perceived noise in the system and positioning. In addition, it ensures that the vehicle can safely stop in case of imminent. In order to ensure driving safety and improve driving comfort, we designed a series of cost objective function has a unique physical meaning of the terms in order to choose the best execution trace.
system framework
Figure provides an overview of the software architecture of the AGV. The system consists of various modules, such as sensors, digital maps, task documents, perception and positioning systems, mission planning, a route planner, behavior planner, trajectory planning and tracking controller, a lower actuator control and human-machine interface (HMI). Each module based Inter Process Communication Toolkit subscribe communication protocol for communication with other modules by specific publisher /.
Such as laser, radar, cameras and other sensors to provide real-time information about the surroundings sensor. Further, other sensors, such as GPS with an inertial measurement unit (IMU) and an encoder wheel in combination, for coarse positioning information acquisition vehicle. Sensing information is mainly used for two purposes: one for the sensing system, such as detection of lane markers, traffic lights, and static and dynamic objects (such as pedestrians, cyclists and other vehicles); the other is to achieve accurate and robust local of. To achieve Lanes stage accurate and reliable positioning, many researchers use GPS, IMU maps and positioning technology, combined laser scanner online 3D point data [38], [40], [41] or a camera visual data [42], [43], [44].
In this article, online from HDL64E Velodyne LiDAR sensing 3D point data combined with high-resolution 3D map data by manually driving record, it has been applied to real-time estimation of the vehicle position and posture. Based on this positioning method, you can use the previous information-rich digital map. In practice, we use detailed digital maps manually constructed, which provides a wealth of information transmission, information (such as location, number and topology) and traffic rules (eg speed limits). Localization and mapping method details are not the main focus of this article.
Results sensing and positioning systems is shown in FIG. 2, FIG master points using LiDAR data to create a 3D. With a magenta cell line is an obstacle detected, including static and dynamic obstacles. Rectangular block represents a moving object, the moving direction indicated by an arrow. White lines indicate lane based on the digital map. In practice, we found that the positioning method to ensure our autonomous vehicles play a key role in a dynamic urban environment in terms of safe driving during the day and night, sunshine or rain.
The use of digital maps and accurate localization information, global route planners can calculate the fastest mission planning based on the results of the reference line. The goal is to rely on the behavior of planners presence awareness information and reference channel plans to make informed decisions. This paper focuses on the trajectory planning module, which plays a key role in ensuring road safety and comfort.
Flowchart shown in FIG. 3 trajectory planning algorithm is first derived reference path. We nonlinear optimization method to smooth the center line, and then using a cubic B-spline curve interpolating them, instead of using the reference channel centerline. In this manner, it is possible to obtain a smooth reference path (or baseline). As described [13], in order to mimic human-like driving behavior of an urban environment, the vehicle trajectory may be isolated as the natural lateral movement and longitudinal movement. In view of this, the curve based on the coordinates instead of Cartesian space trajectory planning path planning task into subtasks and speed plan. Generating a rich set of candidates based on the reference trajectory path. After that, a crash test candidate trajectories to trim the collision with the obstacle. Then, a context-aware application assessment track to drive an optimal trajectory, the trajectory tracking controller to convert it to a lower actuator command.
Reference line smoothing / interpolation
in a structured environment, can directly calculate the desired reference path right border and left lane. In practice, we found that the center line of curvature computing generally neither continuous nor smooth. Low-level tracking controller can not easily track it. Discontinuous track centerline easily lead to control overshoot and ringing. In addition, when the vehicle is turning through the curved driveway or negotiations, following the center line of the need for more control and lead to greater lateral acceleration. To overcome these problems and to improve driving comfort, the conjugate gradient optimization method of the reference path smooth continuous space. It refers to the method similar to [20] and [30].
The first vector to minimize the variation between adjacent vectors, minimizing the accumulation of curvature along the second path. They are performed in order to improve the smoothness of the reference path. The third warranty cost optimized reference path satisfies the constraint condition imposed road boundary. Boundary conditions can be obtained from a digital road map. Right cost items and re-iteration of an effect on the smoothness of the path. Like other nonlinear optimization methods, conjugate gradient algorithm returns only local optimal solution. Indeed, omega] 1 is set to 10, ω2 set to 1, which is empirically obtained. In addition, the maximum number of iterations is limited to 400, in order to meet real-time requirements. In the future, we will use machine learning techniques to adaptively adjust these parameters.
Before and after the smoothing treatment, the number of vertices is set to be the same. In fact, it found that the distance between the apex sometimes too much. In order to obtain dense waypoint, we use cubic B-spline interpolation carried out. Optimization of discrete points of the reference path can be used as a control point. Using discrete control points Pi (i = 0,1, ..., n), a cubic B-spline curve segments
Generating a spline curve path based on coordinate
in order to avoid the obstacle while following the reference frame, [11] proposed a model predictive motion planner, which first group of terminals to a state of lateral movement from the reference frame is sampled. The route generation then expressed as two point boundary value problem [45]. To reduce the solution space while observing movement of a vehicle restraint using a parametric polynomial curvature steering control input space. Although this method takes into account the constraints of the state of the terminal, but it is close to the path during measurement. Thus, it can not generate a long-term trajectory is aligned with the bend in the road, for example, a curved right angle turns. As we mentioned before, generation and road geometry alignment of the track environment is essential. Solve the problem, we refer to [25], [26] path [46] mentioned in the generation strategy. 5, employed to generate a set path candidates based on the motion planning strategy samples having different offset laterally displaceable base frame. In this manner, the vehicle can avoid the obstacle and can be aligned with the static geometry of the base frame. Details are described below.
Based on the curve coordinate frame, the sampling using the terminal state may be represented by two parameters, along a longitudinal distance sf base frame, the base frame from the other lateral deviation lf. Sf longitudinal distance is adjusted to the speed of the vehicle is determined and aligned with the base frame. In practice, it may be adaptively adjusted according to the vehicle speed, without violating the maximum lateral acceleration. L0 lateral offset is the vertical distance between the current position along the path corresponding to the reference point (last point) of the vehicle.
According to [47] mentioned difference smooth control theory, and polynomial splines l0 smooth transition to a sampling offset laterally offset lf from the current terminal. In order to achieve a smooth transition shift offset setting terminal offset from the cubic polynomial wetting initial lateral advantage, which is similar to the algorithm [46] is mentioned.
Parameter 4 four states
Rate plan
rate plan has a great influence on driving safety and comfort, especially for a vehicle driving in real urban traffic participants other dynamic scenes. Therefore, careful distribution speed profile generated at each point along the path generated. In addition, the need to explicitly consider the constraints, such as longitudinal and lateral acceleration limit, a speed limit traffic restrictions. Given these constraints can greatly reduce the speed of the solution space planning and allows speed planner concentrated in the most space possible optimal solution. Thus, after the space velocity generating path generates an execution.
1. The maximum road speed
2. Maximum lateral acceleration limit
3. The longitudinal acceleration / deceleration limit value
For comfort purposes, we further smooth trapezoidal speed, to ensure the continuity of acceleration. Inspired by the speed profile generation parametric process [30], the speed curve using a cubic polynomial parameterized. Compared to [30] and method [26] proposed, we trapezoidal velocity profile, and a smooth rise and descent respectively.
Terminal velocity vf decision to determine usage behavior, which determines the traffic situation. For example, when the vehicle following behavioral decision is, the terminal velocity will be our current lane nearest the front of the vehicle speed (if the speed does not exceed the maximum allowed speed). In other words, the speed planner to use a similar ACC (adaptive cruise control) speed control strategies. When the behavior decision to stop at the stop line, the terminal velocity will be zero.
As shown in FIG 7, we are solving the continuous acceleration ramp-up speed of the S-shaped curve (red solid lines) and a ramp-down speed profile (blue dotted line). Spatial path corresponding to the velocity distribution in combination, produce the track. Note cubic polynomial velocity acceleration profile the acceleration exceeding the linear speed at certain points in the curve. In practice, up and ramp-curve decline curve of acceleration and deceleration values are set to a very conservative value to mitigate the negative impact. According to [49] above, these parameters may be adjusted by the demonstration of human study, in order to achieve a human-like driving.
Evaluation and optimal trajectory
After generating the candidate locus, crash tests performed to reduce or cut off the static and the moving locus of collision with the obstacle. To do this, use the online sensing data will be represented as a partial track plans occupancy grid map. 8, the rectangular shape of the vehicle is approximated by a circular set [33]. Using this method, we can significantly reduce the computational complexity of the crash test. To meet safety requirements, the distance between the obstacle and the center must be greater than the radius of the circle. Taking into account the positioning, sensing noise in the system can be enhanced to increase the safety margin between the radius of the vehicle and the obstacle. To evaluate the safety of the generated track, and the need to consider the static obstacle moving object. In order to avoid the occurrence of a collision with a moving object, it is necessary to predict the movement of other moving objects. Due to imperfect perception and prediction information, impossible to predict the movement of other objects within a large range forecast accurately. In practice, the vehicle speed may be utilized to adjust the driving conditions and the surrounding prediction range. Further, it may also be considered to interact with other road users, and since the driving road geometry due to limited manipulation, as described [34] and [35].
Cost values were normalized to [0,1]. Jp cost item reflects the long trajectory of preferences, which can prevent overreaction due to the too short-sighted behavior of the track caused.
In order to enhance the comfort of lateral movement, smoothness criterion is considered cost term Js, Js is the curvature term cost along the trajectory candidate points.
In order to ensure that the vehicle can follow the guide base frame, Jd penalty cost term trajectory departing from the base frame.
Jc cost item punishment path between continuous playback inconsistent. Inconsistencies between adjacent reschedule easily lead to oscillation control system, overshoot or even instability. In order to ensure consistency between the reproduction, we minimize the sampling terminal lateral offset between the selected path continuous reproduction.
Considering the above conditions to evaluate the cost of generating a four path will inevitably have the problem of the trade-off between security and smoothness, accuracy and consistency. In practice, the result of the decision based on a priori information behavior and digital maps, four right experience to adjust cost item weighting factor. For our future research will use machine learning techniques to adaptively adjust the terms of these costs.
9, using the grid map occupied to represent the sensed environment information from the sensing system is based on. Red grid is static obstacles. Purple curve represents the reference path (base frame). Possible to generate dynamic trajectory, and the reference path and alignment are. As shown in the lower right of the color bar, which represents the color of the trace cost value. The green curve is the best trajectory through the above objective function evaluation. It will track by the track controller switches to execute commands.
Verification Test
Figure 11 depicts the execution time planning algorithm obtained from the experimental part of our record data. The maximum number of candidates generated trajectory is set to 500. The total program time includes a reference path and trajectory planning time optimization time. As can be seen from Figure 11, the trajectory plan in each planning cycle in about 20ms, and the time of the reference path optimization process takes approximately 50ms. Road traffic speed test is 25km / h. Our test environment width of the road is very narrow (single lane width less than 3 meters). Where there is oncoming traffic. Cyclists also share the road with cars. In addition, there is a pedestrian walking along the lane. Sometimes there stroller. For security reasons, the maximum allowable speed is set to 22km / h.
